AUTOMATION HISTORY

Nvidia's new Cosmos model is another sign that devices and machines are getting better at understanding their environments. Meta's Ray-Ban glasses are getting popular. Tech gadgets -- whether they be phones, robots or autonomous vehicles -- are getting better at understanding the world around us, thanks to AI. That message rang loud and clear throughout 2024 and became even louder at CES 2025, where chipmaker Nvidia unveiled a new AI model (a CES award winner) for understanding the physical world and a family of large language models for powering future AI agents. Nvidia CEO Jensen Huang is positioning these world-foundational models as being ideal for robots and autonomous vehicles. There's another class of devices that could benefit from better real-world understanding: smart glasses. Tech-enabled eyewear like Meta's Ray-Bans is quickly becoming the hot new AI gadget, with shipments of Meta's spectacles crossing the 1 million mark in November according to Counterpoint Research. Such devices seem like the ideal vessel for AI agents, or AI helpers that can understand the world around you using cameras and processing speech and visual input to help you get things done rather than just answering questions. Huang didn't say whether Nvidia-powered smart glasses are on the horizon. He did explain how the company's new model would power future smart glasses if partners were to adopt the technology for that purpose. "The use of AI as it gets connected to wearables and virtual presence technology like glasses, all of that is super exciting," Huang said in response to a question about whether its models would work on smart glasses during a press Q&A at CES. Read more: Smart Glasses Are Going to Work This Time, Google's Android President Tells CNET Huang pointed to cloud processing as an option, which would mean queries that use Nvidia's Cosmos model could be handled in the cloud rather than on the device itself. Compact devices like smartphones use this method often to lighten the processing load when running demanding AI models. If a device maker wanted to create glasses that could leverage Nvidia's AI models on-device rather than relying on the cloud, Huang said Cosmos would distill its knowledge into a smaller model that's less generalized and optimized for specific tasks. Nvidia's new Cosmos model is being touted as a platform to gather data about the physical world to train models for robots and self-driving cars -- similar to the way a large language model would learn how to generate text responses after being trained on written media. "The ChatGPT moment for robotics is coming," Huang said in a press release. Nvidia also announced a set of new AI models built with Meta's Llama technology called Llama Nemotron, which is designed to accelerate the development of AI agents. It's interesting to think about how these AI tools and models could potentially be applied to smart glasses too. A recent Nvidia patent filing fueled speculation about upcoming smart glasses, although the chipmaker hasn't made any announcements about future products in that space. Nvidia's new models and Huang's comments come as Google, Samsung and Qualcomm announced last month that they're building a new mixed-reality platform for smart glasses and headsets called Android XR, hinting that smart glasses could soon become more prominent. Several new types of smart glasses were also on display at CES 2025, such as the RayNeo X3 Pro and Halliday smart glasses. The International Data Corporation also predicted in September that shipments of smart glasses would grow by 73.1% in 2024. Nvidia's moves are another space to watch for too. For more from CES, check out the best TVs, the best laptops and the most impressive concepts we've seen.

A "neural codec language model", VALL-E uses discrete codes derived from an off-the-shelf neural audio codec model to synthesize high-quality personalised speech with only a 3-second recording of an unseen speaker. The AI is trained with 60,000 hours of English speech with over 7,000 unique speakers. All this data is taken from Libri-Light, the Meta-owned audio library that collects spoken English audio. It can also imitate the speaker's emotional tone and acoustic environment. "Experiment results show that VALL-E significantly outperforms the state-of-the-art zero-shot TTS system in terms of speech naturalness and speaker similarity," say Microsoft researchers in their paper. The three-second voice input needs to match with some other samples in the training data provided to have a better result. This is why VALL-E should be more diverse in the future. The training data will be scaled up to improve the performances of prosody, speaking style, and speaker similarity perspectives, Microsoft says. READ MORE: AI offers solutions, but needs to be more transparent: experts How can we benefit from VALL-E? For now VALL-E can only convert text into speech in the chosen voice. It can’t create new content. Its creators are hopeful that VALL-E can provide various benefits in terms of speech editing and audio content creation. The example of Stephen Hawking using a text-to-speech generator to continue his studies while suffering from classical motor neuron disease (ALS) has shown the world one of the highest benefits one could get from this technology. VALL-E can be used for simultaneous translations, or to create the voice of our loved ones who had passed away. Creating audiobooks would be a lot easier and faster with VALL-E. One can create a voice for any written peace or text message in a short time. For all these uses and more, we need to wait for Microsoft to open VALL-E to public use. Microsoft has not said yet when the new AI will be available for public consumption. READ MORE:Can an invention enabled by artificial intelligence be patented? VALL-E might bring risks While the question of how to use AI technologies safely and ethically is being asked more often than ever these days, people express their ethical concerns over newly launched systems like Chat GPT, Lensa AI, or VALL-E. Chat GPT, a new chatbot AIthat can process natural language tasks such as text generation and language translation started debates on students committing plagiarism by using this AI for their homework. At the same time, Lensa AI, an app that uses algorithms to generate ordinary photos into artistic renderings led to ethical questions on artistic production made by using other artists' works. Many argue that it cannot replace human artists who make digital art. VALL-E, likewise, has potential risks of misuse that can criminalise the users, such as spoofing voice identification or impersonating a specific speaker. Impersonating people's voices without their consent might fuel mischief and deception which could then lead to social harm. Similar to Lensa's risk of replacing real artists, VALL-E also leads to concerns over ethics of art. In a case when music production companies make the AI sing new songs without the consent of the voice-owner singer, it might not be that much fun to use it. READ MORE: Robots can be racist and sexist, new study warns Microsoft's response to concerns Microsoft says it is aware of these concerns and possible risks that robots might bring on. It had apologised in the past for its chatbot Tay's offensive tweets. The researchers who created VALL-E stated in their paper that they are likely to build a measuring mechanism that can prevent such risks. "Since VALL-E could synthesize speech that maintains speaker identity, it may carry potential risks in misuse of the model, such as spoofing voice identification or impersonating a specific speaker," they said in the paper. "To mitigate such risks, it is possible to build a detection model to discriminate whether an audio clip was synthesized by VALL-E. We will also put Microsoft AI Principles into practice when further developing the models." they added. The company will see if it could build such a detection model and if so, how much of these risks will be mitigated by it, only when the project is open to public use, they said. READ MORE:AI offers solutions, but needs to be more transparent: experts 00:00 00:00

September 27, 2022 12:15 ET | Source: Market Research Future Market Research Future New York, US, Sept. 27, 2022 (GLOBE NEWSWIRE) -- According to a comprehensive research report by Market Research Future (MRFR), “Speech Recognition Market, By Type, Technology, Verticals - Forecast Till 2030”, to garner USD 48.8 billion by 2030, registering approximately 21.30% CAGR during the review period (2022-2030). Speech Recognition Market Overview The speech recognition market witness’s rapid revenue growth due to the increasing use in the education sectors globally. Besides, the rising demand for accurate and easy-to-use speech recognition APIs suitable for multiple fields and languages fosters market growth. The speech-to-text industry that perceives brisk digitization impact market shares positively. Also, other industries moving toward digital transformation boost the speech recognition market size. Players active in the speech recognition market are, Get Free Sample PDF Brochure: https://www.marketresearchfuture.com/sample_request/1815 Users of speech recognition technologies are demanding much more than ever. New and changing industrial safety regulations are forcing industries to change workflows and adapt to the changes. Resultantly, the requirements of speech recognition solutions have expanded beyond simple text transcription, adding advanced capabilities like punctuation, speaker notes, global language packs, organizational reforms, and new vocabulary with frequent builds. With the growing adoption of new voice-enabled learning tools that can address equity and bias issues, the speech recognition market is likely to gain significant traction in the next few years. Over the past several years, the introduction of breakthrough technologies and digital transformation in various industries has been growing exponentially. This, alongside the resultant scale-up in the speech-to-text industry, is promoting the need to meet these new technological requirements. Also, new developments of self-supervised learning solutions that enable speech recognition engines to learn from unstructured data across the web are increasingly coming forward. Self-supervised learning allows access to a large range of unlabeled data, social media, etc., offers a wider variety of voices, and eliminates the need for human supervision. Industry Trends There are a number of researches to drive advances in speech recognition technology to equally diversify and understand accents & dialects clearly. Besides, the rising use of new voice-enabled learning tools demonstrating the capability to address equity and bias creates vast opportunities. The growing adoption of speech recognition technology in education drives the speech recognition market growth. Additionally, significant investments in developing high-performance and efficient speech recognition boost the market size. Large upgrades in major offerings have enabled educational applications that are accelerating demand for more joyous and frictionless voice experiences, accelerating the speech recognition market value. These advanced educational tools have boosted the interest of developers, educators, and researchers, further widening the access to educational resources and minimizing the risks of implicit biases. Report Scope: Browse In-depth Market Research Report (100 Pages) on Speech Recognition Market: https://www.marketresearchfuture.com/reports/speech-recognition-market-1815 Speech Recognition Market Regional Analysis North America leads the global market for speech recognition technology. Growing implementations of speech recognition applications in cell phones escalate the market value. Besides, the expanding utilization of this technology for obtaining customer consent/ acknowledgment over speech/voice in mobile banking and buyers & IoT gadgets boost the speech recognition market size. Rapid developments and rising applications of this technology substantiate market revenues. Moreover, the large presence of major technology providers and high spending on automotive infotainment systems increase the market sales of speech recognition. Europe is another profitable market for speech recognition, headed by the rising interest in discourse and voice acknowledgment. Additionally, voice recognition tech observing significant advances and expanded applications in buyer gadgets and retail areas drives the speech recognition market demand. The vast automotive sector in the region creates substantial market demand, increasingly using patterns of associated devices in robots. APAC is an emerging market for speech recognition systems. The growing number of developments and use of voice-empowered gadgets in the auto and medical businesses push the market growth. Furthermore, augmenting demand for standalone products of language recognition, speech-to-text, machine translation, and transliteration provides significant market opportunities. China, Japan, and Singapore are major speech recognition markets supporting the growth of the regional market. Speech Recognition Market Segments This speech recognition market is segmented into vehicle types, technologies, verticals, and regions. The type segment is bifurcated into speaker-dependent and speaker-independent. The technology segment is bifurcated into AI-based and non-AI-based. The vertical segment is bifurcated into military, automotive, healthcare, and others. The region segment is bifurcated into the Americas, Asia-Pacific, Middle East & Africa, Europe, and rest-of-the-world. Ask for Discount: https://www.marketresearchfuture.com/sample_request/1815 Speech Recognition Market Competitive Analysis The highly competitive speech recognition market appears fragmented due to the presence of several well-established players. To gain a competitive advantage, industry players adopt strategic approaches such as collaborations, mergers & acquisitions, expansions, and product/ technology launches. Speech recognition technology enables learning and plays experiences. Therefore, speech recognition market players are integrating this technology with AI to develop efficient literacy solutions to quickly and accurately assess reading and language skills. This gives teaching professionals a seamless, scalable, and more accurate approach to assessing the language reading & learning progress of students. Further, market players make substantial investments in driving their R&D activities and expansion plans. For instance, on Sep.21, 2022, Nvidia Corporation, an American multinational technology company, unveiled new languages and other upgrades to its Riva speech AI platform for enterprise services. Last year, Nvidia launched the Riva Custom Voice neural-based text-to-speech service that can create lifelike human voices. Talk to Expert: https://www.marketresearchfuture.com/check-discount/1815 Nvidia Riva now features seven language models and tools for distinguishing speakers in conversations and more customization options. Nvidia's Riva speech AI software development kit (SDK) features speech recognition and text-to-speech functions required in enterprises to augment the existing automatic speech recognition and domain-specific customization. Related Reports: Global Speech Analytics Market, by Type, by Deployment Type, by End-User, by Organization Size - Forecast 2030 Far-Field Speech and Voice Recognition Market Research Report: By Component, Microphone Solution, Application - Forecast till 2030 Voice Recognition System Market for Automotive Information Report by Vehicle Type, Technology, Application and by Regions - Global Forecast To 2030 VoLTE Technology Market Research Report, By Device, Technology - Forecast till 2030 About Market Research Future: Market Research Future (MRFR) is a global market research company that takes pride in its services, offering a complete and accurate analysis regarding diverse markets and consumers worldwide. Market Research Future has the distinguished objective of providing the optimal quality research and granular research to clients. Our market research studies by products, services, technologies, applications, end users, and market players for global, regional, and country level market segments, enable our clients to see more, know more, and do more, which help answer your most important questions. Follow Us: LinkedIn | Twitter

Author of the article: You can save this article by registering for free here. Or sign-in if you have an account. KYIV, Ukraine (AP) — Drone advances in Ukraine have accelerated a long-anticipated technology trend that could soon bring the world’s first fully autonomous fighting robots to the battlefield, inaugurating a new age of warfare. Enjoy the latest local, national and international news. Enjoy the latest local, national and international news. Create an account or sign in to continue with your reading experience. Create an account or sign in to continue with your reading experience. The longer the war lasts, the more likely it becomes that drones will be used to identify, select and attack targets without help from humans, according to military analysts, combatants and artificial intelligence researchers. That would mark a revolution in military technology as profound as the introduction of the machine gun. Ukraine already has semi-autonomous attack drones and counter-drone weapons endowed with AI. Russia also claims to possess AI weaponry, though the claims are unproven. But there are no confirmed instances of a nation putting into combat robots that have killed entirely on their own. Experts say it may be only a matter of time before either Russia or Ukraine, or both, deploy them. “Many states are developing this technology,” said Zachary Kallenborn, a George Mason University weapons innovation analyst. “Clearly, it’s not all that difficult.” The sense of inevitability extends to activists, who have tried for years to ban killer drones but now believe they must settle for trying to restrict the weapons’ offensive use. Ukraine’s digital transformation minister, Mykhailo Fedorov, agrees that fully autonomous killer drones are “a logical and inevitable next step” in weapons development. He said Ukraine has been doing “a lot of R&D in this direction.” “I think that the potential for this is great in the next six months,” Fedorov told The Associated Press in a recent interview. Ukrainian Lt. Col. Yaroslav Honchar, co-founder of the combat drone innovation nonprofit Aerorozvidka, said in a recent interview near the front that human war fighters simply cannot process information and make decisions as quickly as machines. Ukrainian military leaders currently prohibit the use of fully independent lethal weapons, although that could change, he said. “We have not crossed this line yet — and I say ‘yet’ because I don’t know what will happen in the future.” said Honchar, whose group has spearheaded drone innovation in Ukraine, converting cheap commercial drones into lethal weapons. Russia could obtain autonomous AI from Iran or elsewhere. The long-range Shahed-136 exploding drones supplied by Iran have crippled Ukrainian power plants and terrorized civilians but are not especially smart. Iran has other drones in its evolving arsenal that it says feature AI. Without a great deal of trouble, Ukraine could make its semi-autonomous weaponized drones fully independent in order to better survive battlefield jamming, their Western manufacturers say. Those drones include the U.S.-made Switchblade 600 and the Polish Warmate, which both currently require a human to choose targets over a live video feed. AI finishes the job. The drones, technically known as “loitering munitions,” can hover for minutes over a target, awaiting a clean shot. “The technology to achieve a fully autonomous mission with Switchblade pretty much exists today,” said Wahid Nawabi, CEO of AeroVironment, its maker. That will require a policy change — to remove the human from the decision-making loop — that he estimates is three years away. Drones can already recognize targets such as armored vehicles using cataloged images. But there is disagreement over whether the technology is reliable enough to ensure that the machines don’t err and take the lives of noncombatants. The AP asked the defense ministries of Ukraine and Russia if they have used autonomous weapons offensively — and whether they would agree not to use them if the other side similarly agreed. Neither responded. If either side were to go on the attack with full AI, it might not even be a first. An inconclusive U.N. report suggested that killer robots debuted in Libya’s internecine conflict in 2020, when Turkish-made Kargu-2 drones in full-automatic mode killed an unspecified number of combatants. A spokesman for STM, the manufacturer, said the report was based on “speculative, unverified” information and “should not be taken seriously.” He told the AP the Kargu-2 cannot attack a target until the operator tells it to do so. Fully autonomous AI is already helping to defend Ukraine. Utah-based Fortem Technologies has supplied the Ukrainian military with drone-hunting systems that combine small radars and unmanned aerial vehicles, both powered by AI. The radars are designed to identify enemy drones, which the UAVs then disable by firing nets at them — all without human assistance. The number of AI-endowed drones keeps growing. Israel has been exporting them for decades. Its radar-killing Harpy can hover over anti-aircraft radar for up to nine hours waiting for them to power up. Other examples include Beijing’s Blowfish-3 unmanned weaponized helicopter. Russia has been working on a nuclear-tipped underwater AI drone called the Poseidon. The Dutch are currently testing a ground robot with a .50-caliber machine gun. Honchar believes Russia, whose attacks on Ukrainian civilians have shown little regard for international law, would have used killer autonomous drones by now if the Kremlin had them. “I don’t think they’d have any scruples,” agreed Adam Bartosiewicz, vice president of WB Group, which makes the Warmate. AI is a priority for Russia. President Vladimir Putin said in 2017 that whoever dominates that technology will rule the world. In a Dec. 21 speech, he expressed confidence in the Russian arms industry’s ability to embed AI in war machines, stressing that “the most effective weapons systems are those that operate quickly and practically in an automatic mode.” Russian officials already claim their Lancet drone can operate with full autonomy. “It’s not going to be easy to know if and when Russia crosses that line,” said Gregory C. Allen, former director of strategy and policy at the Pentagon’s Joint Artificial Intelligence Center. Switching a drone from remote piloting to full autonomy might not be perceptible. To date, drones able to work in both modes have performed better when piloted by a human, Allen said. The technology is not especially complicated, said University of California-Berkeley professor Stuart Russell, a top AI researcher. In the mid-2010s, colleagues he polled agreed that graduate students could, in a single term, produce an autonomous drone “capable of finding and killing an individual, let’s say, inside a building,” he said. An effort to lay international ground rules for military drones has so far been fruitless. Nine years of informal United Nations talks in Geneva made little headway, with major powers including the United States and Russia opposing a ban. The last session, in December, ended with no new round scheduled. Washington policymakers say they won’t agree to a ban because rivals developing drones cannot be trusted to use them ethically. Toby Walsh, an Australian academic who, like Russell, campaigns against killer robots, hopes to achieve a consensus on some limits, including a ban on systems that use facial recognition and other data to identify or attack individuals or categories of people. “If we are not careful, they are going to proliferate much more easily than nuclear weapons,” said Walsh, author of “Machines Behaving Badly.” “If you can get a robot to kill one person, you can get it to kill a thousand.” Scientists also worry about AI weapons being repurposed by terrorists. In one feared scenario, the U.S. military spends hundreds of millions writing code to power killer drones. Then it gets stolen and copied, effectively giving terrorists the same weapon. To date, the Pentagon has neither clearly defined “an AI-enabled autonomous weapon” nor authorized a single such weapon for use by U.S. troops, said Allen, the former Defense Department official. Any proposed system must be approved by the chairman of the Joint Chiefs of Staff and two undersecretaries. That’s not stopping the weapons from being developed across the U.S. Projects are underway at the Defense Advanced Research Projects Agency, military labs, academic institutions and in the private sector. The Pentagon has emphasized using AI to augment human warriors. The Air Force is studying ways to pair pilots with drone wingmen. A booster of the idea, former Deputy Defense Secretary Robert O. Work, said in a report last month that it “would be crazy not to go to an autonomous system” once AI-enabled systems outperform humans — a threshold that he said was crossed in 2015, when computer vision eclipsed that of humans. Humans have already been pushed out in some defensive systems. Israel’s Iron Dome missile shield is authorized to open fire automatically, although it is said to be monitored by a person who can intervene if the system goes after the wrong target. Multiple countries, and every branch of the U.S. military, are developing drones that can attack in deadly synchronized swarms, according to Kallenborn, the George Mason researcher. So will future wars become a fight to the last drone? That’s what Putin predicted in a 2017 televised chat with engineering students: “When one party’s drones are destroyed by drones of another, it will have no other choice but to surrender.” —— Frank Bajak reported from Boston. Associated Press journalists Tara Copp in Washington, Garance Burke in San Francisco and Suzan Fraser in Turkey contributed to this report. —— Follow the AP’s coverage of the war at https://apnews.com/hub/russia-ukraine —— This story has been updated to correct when the U.N. report was issued. It came out in 2021, not last year. Postmedia is committed to maintaining a lively but civil forum for discussion. Please keep comments relevant and respectful. Comments may take up to an hour to appear on the site. You will receive an email if there is a reply to your comment, an update to a thread you follow or if a user you follow comments. Visit our Community Guidelines for more information. SHA Mexico pairs beach bliss with boot-camp rules for a serious reality check Maido is just one of many eateries that make Peru’s capital a culinary mecca The 3 best beauty products we tried this week from Dior, SVR and L'Oréal Paris Plus, our favourite shoe styles for spring Plus best products for dry, oily and flaky scalps 365 Bloor Street East, Toronto, Ontario, M4W 3L4 © 2026 National Post, a division of Postmedia Network Inc. All rights reserved. Unauthorized distribution, transmission or republication strictly prohibited. This website uses cookies to personalize your content (including ads), and allows us to analyze our traffic. Read more about cookies here. By continuing to use our site, you agree to our Terms of Use and Privacy Policy. You can manage saved articles in your account. and save up to 100 articles! You can manage your saved articles in your account and clicking the X located at the bottom right of the article.

Tim Sweezy Stay updated with the latest news and updates. Subscribe to our newsletter! Home Reviews News Blogs Full Site Sitemap PC Components Systems Mobile IT Infrastructure Leisure Videos About Advertise News Tips Contact Privacy And Terms HotTech Accessibility Shop Twitter Facebook YouTube RSS Or sign in manually:

Copyright 2026 The Associated Press. All Rights Reserved. Copyright 2026 The Associated Press. All Rights Reserved. FILE - A Switchblade 600 loitering missile drone manufactured by AeroVironment is displayed at the Eurosatory arms show in Villepinte, north of Paris, on June 14, 2022. Drone advances in Ukraine have accelerated a long-anticipated technology trend that could soon bring the world’s first fully autonomous fighting robots to the battlefield, inaugurating a new age of warfare. (AP Photo/Michel Euler, File) FILE - Ukraine’s Minister of Digital Transformation Mykhailo Fedorov gestures during a news conference at the Web Summit technology conference in Lisbon, Portugal, Thursday, Nov. 3, 2022. Fedorov, agrees that fully autonomous killer drones are “a logical and inevitable next step” in weapons development. (AP Photo/Armando Franca, File) FILE - Mykhailo Fedorov, Ukrainian minister of digital transformation, poses for a photo illuminated by a lamp in the shape of a Ukrainian trident, a national symbol, in his office in Kyiv, Ukraine, Wednesday, Dec. 21, 2022. Fedorov, agrees that fully autonomous killer drones are “a logical and inevitable next step” in weapons development. (AP Photo/Felipe Dana) FILE - This undated photograph released by the Ukrainian military’s Strategic Communications Directorate shows the wreckage of what Kyiv has described as an Iranian Shahed drone downed near Kupiansk, Ukraine. (Ukrainian military’s Strategic Communications Directorate via AP, File) FILE - A drone is seen in the sky seconds before it fired on buildings in Kyiv, Ukraine, on Oct. 17, 2022. (AP Photo/Efrem Lukatsky, File) FILE - Ukrainian soldiers shoot a drone that appears in the sky seconds before it fired on buildings in Kyiv, Ukraine, on Oct. 17, 2022. (AP Photo/Vadym Sarakhan, File) FILE - Ukrainian soldiers launch a drone at Russian positions near Bakhmut, Donetsk region, Ukraine, Thursday, Dec. 15, 2022. (AP Photo/LIBKOS, File) FILE - Firefighters work after a drone attack on buildings in Kyiv, Ukraine, Monday, Oct. 17, 2022. (AP Photo/Roman Hrytsyna, File) KYIV, Ukraine (AP) — Drone advances in Ukraine have accelerated a long-anticipated technology trend that could soon bring the world’s first fully autonomous fighting robots to the battlefield, inaugurating a new age of warfare. The longer the war lasts, the more likely it becomes that drones will be used to identify, select and attack targets without help from humans, according to military analysts, combatants and artificial intelligence researchers. That would mark a revolution in military technology as profound as the introduction of the machine gun. Ukraine already has semi-autonomous attack drones and counter-drone weapons endowed with AI. Russia also claims to possess AI weaponry, though the claims are unproven. But there are no confirmed instances of a nation putting into combat robots that have killed entirely on their own. Experts say it may be only a matter of time before either Russia or Ukraine, or both, deploy them. “Many states are developing this technology,” said Zachary Kallenborn, a George Mason University weapons innovation analyst. ”Clearly, it’s not all that difficult.” The sense of inevitability extends to activists, who have tried for years to ban killer drones but now believe they must settle for trying to restrict the weapons’ offensive use. Ukraine’s digital transformation minister, Mykhailo Fedorov, agrees that fully autonomous killer drones are “a logical and inevitable next step” in weapons development. He said Ukraine has been doing “a lot of R&D in this direction.” “I think that the potential for this is great in the next six months,” Fedorov told The Associated Press in a recent interview. Ukrainian Lt. Col. Yaroslav Honchar, co-founder of the combat drone innovation nonprofit Aerorozvidka, said in a recent interview near the front that human war fighters simply cannot process information and make decisions as quickly as machines. Ukrainian military leaders currently prohibit the use of fully independent lethal weapons, although that could change, he said. “We have not crossed this line yet – and I say ‘yet’ because I don’t know what will happen in the future.” said Honchar, whose group has spearheaded drone innovation in Ukraine, converting cheap commercial drones into lethal weapons. Russia could obtain autonomous AI from Iran or elsewhere. The long-range Shahed-136 exploding drones supplied by Iran have crippled Ukrainian power plants and terrorized civilians but are not especially smart. Iran has other drones in its evolving arsenal that it says feature AI. Without a great deal of trouble, Ukraine could make its semi-autonomous weaponized drones fully independent in order to better survive battlefield jamming, their Western manufacturers say. Those drones include the U.S.-made Switchblade 600 and the Polish Warmate, which both currently require a human to choose targets over a live video feed. AI finishes the job. The drones, technically known as “loitering munitions,” can hover for minutes over a target, awaiting a clean shot. “The technology to achieve a fully autonomous mission with Switchblade pretty much exists today,” said Wahid Nawabi, CEO of AeroVironment, its maker. That will require a policy change — to remove the human from the decision-making loop — that he estimates is three years away. Drones can already recognize targets such as armored vehicles using cataloged images. But there is disagreement over whether the technology is reliable enough to ensure that the machines don’t err and take the lives of noncombatants. The AP asked the defense ministries of Ukraine and Russia if they have used autonomous weapons offensively – and whether they would agree not to use them if the other side similarly agreed. Neither responded. If either side were to go on the attack with full AI, it might not even be a first. An inconclusive U.N. report suggested that killer robots debuted in Libya’s internecine conflict in 2020, when Turkish-made Kargu-2 drones in full-automatic mode killed an unspecified number of combatants. A spokesman for STM, the manufacturer, said the report was based on “speculative, unverified” information and “should not be taken seriously.” He told the AP the Kargu-2 cannot attack a target until the operator tells it to do so. Fully autonomous AI is already helping to defend Ukraine. Utah-based Fortem Technologies has supplied the Ukrainian military with drone-hunting systems that combine small radars and unmanned aerial vehicles, both powered by AI. The radars are designed to identify enemy drones, which the UAVs then disable by firing nets at them — all without human assistance. The number of AI-endowed drones keeps growing. Israel has been exporting them for decades. Its radar-killing Harpy can hover over anti-aircraft radar for up to nine hours waiting for them to power up. Other examples include Beijing’s Blowfish-3 unmanned weaponized helicopter. Russia has been working on a nuclear-tipped underwater AI drone called the Poseidon. The Dutch are currently testing a ground robot with a .50-caliber machine gun. Honchar believes Russia, whose attacks on Ukrainian civilians have shown little regard for international law, would have used killer autonomous drones by now if the Kremlin had them. “I don’t think they’d have any scruples,” agreed Adam Bartosiewicz, vice president of WB Group, which makes the Warmate. AI is a priority for Russia. President Vladimir Putin said in 2017 that whoever dominates that technology will rule the world. In a Dec. 21 speech, he expressed confidence in the Russian arms industry’s ability to embed AI in war machines, stressing that “the most effective weapons systems are those that operate quickly and practically in an automatic mode.” Russian officials already claim their Lancet drone can operate with full autonomy. “It’s not going to be easy to know if and when Russia crosses that line,” said Gregory C. Allen, former director of strategy and policy at the Pentagon’s Joint Artificial Intelligence Center. Switching a drone from remote piloting to full autonomy might not be perceptible. To date, drones able to work in both modes have performed better when piloted by a human, Allen said. The technology is not especially complicated, said University of California-Berkeley professor Stuart Russell, a top AI researcher. In the mid-2010s, colleagues he polled agreed that graduate students could, in a single term, produce an autonomous drone “capable of finding and killing an individual, let’s say, inside a building,” he said. An effort to lay international ground rules for military drones has so far been fruitless. Nine years of informal United Nations talks in Geneva made little headway, with major powers including the United States and Russia opposing a ban. The last session, in December, ended with no new round scheduled. Washington policymakers say they won’t agree to a ban because rivals developing drones cannot be trusted to use them ethically. Toby Walsh, an Australian academic who, like Russell, campaigns against killer robots, hopes to achieve a consensus on some limits, including a ban on systems that use facial recognition and other data to identify or attack individuals or categories of people. “If we are not careful, they are going to proliferate much more easily than nuclear weapons,” said Walsh, author of “Machines Behaving Badly.” “If you can get a robot to kill one person, you can get it to kill a thousand.” Scientists also worry about AI weapons being repurposed by terrorists. In one feared scenario, the U.S. military spends hundreds of millions writing code to power killer drones. Then it gets stolen and copied, effectively giving terrorists the same weapon. To date, the Pentagon has neither clearly defined “an AI-enabled autonomous weapon” nor authorized a single such weapon for use by U.S. troops, said Allen, the former Defense Department official. Any proposed system must be approved by the chairman of the Joint Chiefs of Staff and two undersecretaries. That’s not stopping the weapons from being developed across the U.S. Projects are underway at the Defense Advanced Research Projects Agency, military labs, academic institutions and in the private sector. The Pentagon has emphasized using AI to augment human warriors. The Air Force is studying ways to pair pilots with drone wingmen. A booster of the idea, former Deputy Defense Secretary Robert O. Work, said in a report last month that it “would be crazy not to go to an autonomous system” once AI-enabled systems outperform humans — a threshold that he said was crossed in 2015, when computer vision eclipsed that of humans. Humans have already been pushed out in some defensive systems. Israel’s Iron Dome missile shield is authorized to open fire automatically, although it is said to be monitored by a person who can intervene if the system goes after the wrong target. Multiple countries, and every branch of the U.S. military, are developing drones that can attack in deadly synchronized swarms, according to Kallenborn, the George Mason researcher. So will future wars become a fight to the last drone? That’s what Putin predicted in a 2017 televised chat with engineering students: “When one party’s drones are destroyed by drones of another, it will have no other choice but to surrender.” ___ Frank Bajak reported from Boston. Associated Press journalists Tara Copp in Washington, Garance Burke in San Francisco and Suzan Fraser in Turkey contributed to this report. ___ Follow the AP’s coverage of the war at https://apnews.com/hub/russia-ukraine ___ This story has been updated to correct when the U.N. report was issued. It came out in 2021, not last year. Copyright 2026 The Associated Press. All Rights Reserved.

Generative AI is reshaping industries and with accelerated computing, "we're pushing the boundaries of what's possible and driving the next wave of technological advancement," Nvidia CEO Jensen Huang (黃仁勳) said in a speech in Taipei Sunday. (Full text of the story is now in CNA English news archive. To view the full story, you will need to be a subscribed member of the CNA archive. To subscribe, please read here.)

Personalise the news and stay in the know Emergency Backstory Newsletters 中文新闻 BERITA BAHASA INDONESIA TOK PISIN Find any issues using dark mode? Please let us know Topic:Unrest, Conflict and War Ukrainian soldiers launch a drone at Russian positions near Bakhmut in Ukraine's Donetsk region in December 2022. (AP: LIBKOS/ File) Drone advances in Ukraine have accelerated a long-anticipated technology trend that could soon bring the world's first fully autonomous fighting robots to the battlefield, inaugurating a new age of warfare. The longer the war lasts, the more likely it becomes that drones will be used to identify, select and attack targets without help from humans, according to military analysts, combatants and artificial intelligence (AI) researchers. That would mark a revolution in military technology as profound as the introduction of the machine gun. Ukraine already has semi-autonomous attack drones and counter-drone weapons endowed with AI. Russia also says it possesses AI weaponry, though the claims are unproven. But there are no confirmed instances of a nation putting into combat robots that have killed entirely on their own. Experts say it may be only a matter of time before either Russia or Ukraine, or both, deploy them. "Many states are developing this technology," George Mason University weapons innovation analyst Zachary Kallenborn said. "Clearly, it's not all that difficult." The sense of inevitability extends to activists, who have tried for years to ban killer drones, but now believe they must settle for trying to restrict the weapons' offensive use. Ukraine's digital transformation minister Mykhailo Fedorov agreed that fully autonomous killer drones were "a logical and inevitable next step" in weapons development. He said Ukraine has been doing a lot of [research and development] in that direction. "I think that the potential for this is great in the next six months," Mr Fedorov said. Ukraine's Minister of Digital Transformation Mykhailo Fedorov believes that fully autonomous killer drones are "a logical and inevitable next step" in weapons development. (AP: Armando Franca/ File) Ukrainian Lieutenant Colonel Yaroslav Honchar, co-founder of the combat drone innovation nonprofit Aerorozvidka, said that human war fighters simply cannot process information and make decisions as quickly as machines. Ukrainian military leaders currently prohibit the use of fully independent lethal weapons, although that could change, according to Lieutenant Colonel Honchar, whose group has spearheaded drone innovation in Ukraine, converting cheap commercial drones into lethal weapons. "We have not crossed this line yet – and I say 'yet' because I don't know what will happen in the future," he said. Russia could obtain autonomous AI from Iran or elsewhere. The long-range Shahed-136 exploding drones supplied by Iran have crippled Ukrainian power plants and terrorised civilians, but are not especially smart. Iran has other drones in its evolving arsenal that it says feature AI. Without a great deal of trouble, Ukraine could make its semi-autonomous weaponised drones fully independent in order to better survive battlefield jamming, their Western manufacturers say. Those drones include the US-made Switchblade 600 and the Polish Warmate, which both currently require a human to choose targets over a live video feed. AI finishes the job. The drones, technically known as "loitering munitions", can hover for minutes over a target, awaiting a clean shot. That will require a policy change — to remove the human from the decision-making loop — that he estimates is three years away. From a shabby office in Hong Kong to a house in suburban Florida, these are the businesses allowing Russia to build drones that are killing Ukrainians. Drones can already recognise targets such as armoured vehicles using catalogued images. But there is disagreement over whether the technology is reliable enough to ensure that the machines don't err and take the lives of non-combatants. The Associated Press (AP) asked the defence ministries of Ukraine and Russia if they have used autonomous weapons offensively – and whether they would agree not to use them if the other side similarly agreed. Neither responded. If either side were to go on the attack with full AI, it might not even be a first. An inconclusive United Nations report suggested that killer robots debuted in Libya's internecine conflict in 2020, when Turkish-made Kargu-2 drones in full-automatic mode killed an unspecified number of combatants. A spokesman for the manufacturer STM said the report was based on "speculative, unverified" information and should not be taken seriously. He told the AP that the Kargu-2 cannot attack a target until the operator tells it to do so. Fully autonomous AI is already helping to defend Ukraine. Utah-based Fortem Technologies has supplied the Ukrainian military with drone-hunting systems that combine small radars and unmanned aerial vehicles, both powered by AI. The radars are designed to identify enemy drones, which the UAVs then disable by firing nets at them — all without human assistance. The number of AI-endowed drones keeps growing. Israel has been exporting them for decades. Its radar-killing Harpy can hover over anti-aircraft radar for up to nine hours waiting for them to power up. Other examples include Beijing's Blowfish-3 unmanned weaponised helicopter. Russia has been working on a nuclear-tipped underwater AI drone called the Poseidon. The Dutch are currently testing a ground robot with a .50-calibre machine gun. Lieutenant Colonel Honchar believes Russia, whose attacks on Ukrainian civilians have shown little regard for international law, would have used killer autonomous drones by now if the Kremlin had them. "I don't think they'd have any scruples," agreed Adam Bartosiewicz, vice-president of WB Group, which makes the Warmate. AI is a priority for Russia. President Vladimir Putin said in 2017 that whoever dominates that technology will rule the world. In a speech, Mr Putin expressed confidence in the Russian arms industry's ability to embed AI in war machines. (Reuters/ File) He stressed that "the most effective weapons systems are those that operate quickly and practically in an automatic mode". Russian officials already claim their Lancet drone can operate with full autonomy. "It's not going to be easy to know if and when Russia crosses that line," former director of strategy and policy at the Pentagon's Joint Artificial Intelligence Center Gregory Allen said. Switching a drone from remote piloting to full autonomy might not be perceptible. To date, drones able to work in both modes have performed better when piloted by a human, Mr Allen said. The technology is not especially complicated, top AI researcher from University of California-Berkeley professor Stuart Russell said. In the mid-2010s, colleagues he polled agreed that graduate students could, in a single term, produce an autonomous drone "capable of finding and killing an individual, let's say, inside a building", he said. An effort to lay international ground rules for military drones has so far been fruitless. Nine years of informal United Nations talks in Geneva made little headway, with major powers including the United States and Russia opposing a ban. The last session, in December, ended with no new round scheduled. Washington policymakers say they won't agree to a ban because rivals developing drones cannot be trusted to use them ethically. Toby Walsh, an Australian academic who, like Mr Russell, campaigns against killer robots, hopes to achieve a consensus on some limits, including a ban on systems that use facial recognition and other data to identify or attack individuals or categories of people. "If you can get a robot to kill one person, you can get it to kill a thousand." Scientists also worry about AI weapons being repurposed by terrorists. In one feared scenario, the US military spends hundreds of millions writing code to power killer drones. Then it gets stolen and copied, effectively giving terrorists the same weapon. To date, the Pentagon has neither clearly defined "an AI-enabled autonomous weapon" nor authorised a single such weapon for use by US troops, Mr Allen said. Drones are getting more stealthy, and detection technology is struggling to keep up. Now researchers are reporting a marked improvement in detection ranges, thanks to an unlikely source. Any proposed system must be approved by the chairman of the Joint Chiefs of Staff and two undersecretaries. That's not stopping the weapons from being developed across the US. Projects are underway at the Defense Advanced Research Projects Agency, military labs, academic institutions and in the private sector. The Pentagon has emphasised using AI to augment human warriors. The US Air Force is studying ways to pair pilots with drone wingmen. Former Deputy Defense Secretary Robert Work said in a report last month that it "would be crazy not to go to an autonomous system" once AI-enabled systems outperform humans — a threshold that he said was crossed in 2015, when computer vision eclipsed that of humans. Humans have already been pushed out in some defensive systems. Israel's Iron Dome missile shield is authorised to open fire automatically, although it is said to be monitored by a person who can intervene if the system goes after the wrong target. Multiple countries, and every branch of the US military, are developing drones that can attack in deadly synchronised swarms, according to Mr Kallenborn, the George Mason researcher. So will future wars become a fight to the last drone? That's what Vladimir Putin predicted in a 2017 televised chat with engineering students: "When one party's drones are destroyed by drones of another, it will have no other choice but to surrender". AP Topic:Accidents and Emergency Incidents Topic:Property Prices Topic:Government and Politics Topic:Drownings Topic:Explainer Topic:Unrest, Conflict and War Topic:Government and Politics Ukraine Unrest, Conflict and War Topic:Accidents and Emergency Incidents Topic:Property Prices Topic:Government and Politics Topic:Drownings Topic:Explainer Topic:Property Prices Topic:Unrest, Conflict and War Topic:Government and Politics Topic:Missing Person Your home of Australian stories, conversations and events that shape our nation. This service may include material from Agence France-Presse (AFP), APTN, Reuters, AAP, CNN and the BBC World Service which is copyright and cannot be reproduced. We acknowledge Aboriginal and Torres Strait Islander peoples as the First Australians and Traditional Custodians of the lands where we live, learn, and work. Sign up to get the latest on your favourite topics from the ABC

Technology Karan Kamble Oct 16, 2024 | Updated Nov 13, 2024, 02:26 PM GMT+5:30 Save & read from anywhere! Bookmark stories for easy access on any device or the Swarajya app. One fond childhood memory for kids growing up in the last century was watching the animated sitcom The Jetsons, where robotic hands for household chores were always abundant. It may have even sparked the fantasy in more than a few to have robots helping around the home one day. While robots — even those that look like humans, called ‘humanoid robots’ — were certainly around back then, they tended to be more of an experimental nature, where scientists and engineers were more concerned with tweaking a part here, a part there in order to simply identify and extend fundamental robotic capabilities. As time passed, robots grew more and more advanced, yet largely remained limited to an academic, scientific, or defence setting. However, key advancements in artificial intelligence (AI) over the last decade have promised to propel robots from experiments to helping hands, whether at a factory, a service or a retail outlet, or, as that childhood fantasy goes, in homes. Companies like Boston Dynamics, Tesla, Agility Robotics, Softbank Robotics, and Hanson Robotics are swiftly helping materialise and hopefully soon commercialise miraculously useful humanoid robots. Tesla’s suite of humanoid robot “friends,” showcased on 11 October 2024, provides a taste of things to come. Here in India, in the thriving tech city of Bengaluru, Karnataka, a two-member startup operating out of the co-founder’s apartment is working hard to move the robotics dial forward by designing and developing humanoid robots in India. While the global robotics giants boast of ever-improving, advanced robots that look surprisingly human and are capable of handling a variety of mundane tasks, the key differentiator with the Bengaluru initiative is set to be the price point. Whereas the market rate for humanoid robots presently is anywhere between $10,000 for the basic machines and around $100,000 for the top-of-the-line ones, the Bengaluru company that goes by the name of ‘Cosine Robots’ is targeting a $4,000 price point for their humanoid robot. The company has started piecing together the individual systems, starting with the actuators, to quickly commercialise them and use the resulting profits to build the rest of the parts until they all eventually add up to a full-fledged humanoid robot. For Cosine Robots co-founder Kiran Mathew Koshy, whose home doubles up as the office, building affordable humanoid robots is not just a personal business goal. He believes India must get a move on and develop robots that the Indian public can afford, lest the prices elsewhere make these next-generation robots largely inaccessible to the Indian market. “Even if we are not successful, someone will eventually succeed in building these robots at an affordable price. However, that price may not be affordable in India, even though it might be in the United States, Europe, or elsewhere,” he explains, adding, “It is crucial that we avoid this situation as a country by developing robots that are affordable for Indian industries, made in India, and deployable here.” The Come Up Koshy has done his professional time, accumulating close to a decade of experience working at multinational giants Amazon and Google. His work in these companies was standard software engineering fare, but he harboured the idea of starting up someday. “I've always wanted to do a startup, right from my college days, but I come from a very middle-class family, so that was not something I could jump into right away. Getting a job and getting a couple of promotions in the job was a natural progression,” he says. The time was right in mid-2023. The Covid-19 pandemic had passed. Enough professional experience had been gained. It was time to dust off those old startup dreams. Koshy had maintained a longstanding interest in robotics, but he was smart to notice that recent advancements in AI had changed the game. While building humanoid robots was no longer the dominant challenge, getting them to do stuff without having to program their every move remained a problem. Luckily, AI opened a pathway. In the pre-AI-infused robotics world, “you would use robotics only in a place where you have extreme levels of scale, where you can write the code once and then deploy it to, let's say, 10,000 robots at least,” Koshy says. “The only realistic use cases of robots were in things like the Amazon warehouse. And those are also not robotic arms or anything. Those are more like wheeled robots, which carry entire shelves so that the picking and packing is more convenient.” However, with advancements in AI, the need to hard-code the robot’s every move is fulfilled by AI’s ingenuity. So, AI can operate the robot and get useful work out of it. The AI advances that have propelled robotics development include the transformer-based architectures underlying large language models (LLMs) like ChatGPT, vision transformers, and the technology driving self-driving cars. They have enabled robots to be given instructions via text or speech, through which they scope out the area and the objects contained therein, and then proceed to execute tasks, like moving a glass of water from one place to another. Buoyed by these advances, the focus has shifted to making robots more accessible—specifically, by constructing them simply, swiftly, and at scale to enhance affordability. This is the mission Koshy and his co-founder Delip Thomas, a year senior to Koshy during their Bachelor of Technology studies at the Indian Institute of Technology (IIT), Patna, have undertaken. Building From Scratch Building a humanoid robot, broadly speaking, requires several systems, chief among which are the actuators, sensors, and AI and software. “We’re taking a ground-up approach where we build the hardware first, then add a vision stack that can recognise objects. After that, we will have an AI stack to execute actions based on instructions,” Koshy says. He and his colleague are starting off with the design and development of actuators. Actuators are essential because they enable robots to move. They convert energy into physical motion, enabling robots to walk, pick up objects, and make gestures. “You can think of an actuator as a single joint of the robot. We will soon start selling these actuators, and use the profits to finance building more robots,” Koshy says. Building actuators first is a good strategy because those currently on the market are expensive. A robot maker needs as many actuators as joints designed for the robot. This bumps up the overall cost of the robot. Cosine is, therefore, planning to make cheaper actuators, which will help them lower the cost of their eventual robot offerings. “A single joint actuator on the market today would be around $1,000 plus. I'm talking about precise ones that you would use in industrial arms. We are building ours to be priced between $300 and $400,” Koshy says. The joints of a robot must be interconnected, which brings us to the static components that connect the actuators. Again, with affordability in mind, Koshy and his colleague plan to employ mass-manufacturing techniques like casting and sheet metal bending and cutting to make the static components, borrowing from the ways of automobile manufacturing, where affordable and scale-friendly methodologies are used. “In a car, you have several components which have a very high level of precision, like the engine, crankshaft, or, in an electric car, the motor. Everything else, like the body and chassis, just has to be strong; they don't have to be very precise. It's the precision that raises the cost. The components that need to be precise need to be manufactured using machining or milling, whereas sheet metal operations and stamping, for example, are a lot cheaper and faster,” he says. Similarly, while building robots, the components demanding precision can be machined, whereas the other parts, like the static components, can be made relatively quickly and affordably, like in car manufacturing. Koshy has gone about designing their actuators in such a way that they could be made on the standard automotive original equipment manufacturer (OEM) platforms. “Say, someone who builds pistons for a Maruti Suzuki engine would be able to build actuators for this, except the controller,” he says. Actuators aside, the Bengaluru-based robot makers will also need sensors, along with AI and software. Koshy plans to use the standard Hall effect sensors procured off the shelf. A Hall effect sensor measures the strength and direction of a magnetic field. In the automotive industry, it is used in wheel speed sensors (ABS systems), camshaft or crankshaft position sensors, and throttle position sensors. A more familiar application would be using the sensor to detect the opening and closing of flip cases on our smartphones or laptop lids by sensing embedded magnets. As for the AI stack, Cosine Robots plans to opt for existing off-the-shelf or open-source LLMs. They recognise the need to build AI models of their own at some point, but that’s for later. For now, the focus is on getting their self-made actuators off the ground. Among other things, the ‘actuator first’ strategy will enable the bootstrapped initiative to generate initial revenue, allowing them to pursue funding to gradually enhance their robot offerings. The startup’s earlier attempts to woo investors came to nought. “Deep tech investors have had their hands burnt by a lot of robotics startups. Also, there aren’t really that many deep tech investors in India to begin with. There are some, but they have a policy of not funding until there is some level of revenue, which is hard to get in deep tech before you build a product — where significant risk capital is needed first. This creates a gap in the VC (venture capital or capitalists) market in India.” Therefore, Koshy plans to generate revenue by selling actuators first. This will allow him to deploy capital the way he sees fit. “At that point, if we get investors, that will be even better. But we are not entering a path where we need to have VCs deploy capital before we can do the work,” he says. The good news is that there are takers for Cosine’s actuators already. “We have a couple of pre-orders from people looking to use our actuators. We will start selling them as quickly as we can, probably in the next month.” Building Robots That People Can Buy Cosine is not out to make the most advanced robots. Rather, they are in the business of making robots that most people can buy. Explaining the current state of robotics, Koshy states, “We are not in a phase where the bottleneck is in creating something that doesn’t exist. Instead, we are at a stage where many of these technologies have been developed by various people historically, some even 20 or 30 years ago. However, they are produced at such a low scale and volume today that the per-unit costs are astronomically high. Everyone is racing to reduce this price as much as possible while remaining profitable.” The pursuit of affordable humanoid robots ties into Koshy’s dream of a world, particularly India, where luxuries typically affordable only to the well-off, such as hiring a housekeeper, cook, or driver to make life easier, become accessible to nearly anyone. “It doesn't matter how rich the entire world becomes, but, in our current state of technology, someone in the bottom 50 per cent is never going to have a cook or a maid. Whereas in a world with, let's say, 10 billion robots, this is very trivial. Everyone can then have a robotic cook or a robotic maid to clean up the apartment, and maybe robotic security so that women are safe.” According to Koshy, a society with affordable robots will be as different from one without them as society was before and after the Industrial Revolution. “Imagine a world where you can do construction with robots. You can easily provide five-bedroom apartments for everyone. A lot of the things that are wrong in society are primarily because we don’t have infrastructure. We don't have good roads; we don't have good houses for everyone. These are very simple labour problems which immediately go away once you have affordable robots.” Cosine Robots have begun this journey. They have built a prototype humanoid robot — think of it as Version 1 — that was put together using off-the-shelf motors. The more capable, the more indigenous Version 2 will incorporate its own actuators. Once the robot is ready for the market, the company will initially target industrial settings. This environment will provide controlled real-world experience before the robots are deployed in homes. “In the industry, the person using the robot has full control of the setting and where it is used, so they can easily have a walled-off area where the robot operates, and no one in their right mind will walk into that area. However, in a home environment, you can't really do that if you have pets or your kids running around,” Koshy explains. The home version of the robot will simply be the industry version with additional safety features built into it. “Safety controls would be things like contact sensors, which can detect skin touch on the robot the same way a capacitor screen on your phone can detect a skin touch. Similarly, you can have similar sensors around the robot’s body so that it stops whenever a human touches it. Everything required for home safety is industrial plus some safety stuff,” he says. Sociopolitical Implications Any talk of introducing robots in greater numbers in society typically accompanies some nervousness. The resulting job losses are perhaps the most prominent concern. According to Koshy, though some jobs would be lost, especially those that “involve pressing some buttons,” new opportunities will open up in a society that welcomes humanoid robots. “When we started introducing computers, there were a lot of people who started protesting because they feared the loss of jobs. But the reality is, in India, computers enabled us to do a lot of work and generate a lot of income as a country, primarily by providing service work to other nations,” he says. Koshy argues that a similar scenario could unfold in India's robotics sector. “If you have a robot in the United States (US) and you have someone in India who can control that robot, there is quite a lot of work which you can do remotely, even without AI, even without any automation, by simply doing manual control.” So, through remote robot control, someone in India could help run an American factory at night. Moreover, robotics and especially AI will shake things up in ways we can’t quite grasp just yet. With such an unwritten future ahead of us, where chapters are waiting to be written, India can’t afford to hang back and wait for other authors to do the writing. “I would give an example of what happened to the Arab nations after the printing press was introduced. The Arab nations banned the printing press because they wanted to save the jobs of the royal scribes. What happened as a result of the printing press was a massive increase in living standards in the West, whereas the living standards in most of Asia pretty much remained the same,” Koshy says. “In the long term, things are 100 per cent going to be better. In the short term, will we have turmoil as a result? Yes. Do we need to adjust as a society? Yes,” he adds. As is often the case, governments will have to play a key role. “The best way to tackle it is to ease the pain of people whose jobs are getting displaced without entering full-fledged socialism. So, you still need capitalism, but you can reduce capitalism from all segments to some segments.” Drawing a distinction between the paths taken by Europe and Russia on one hand and the US on the other after the Industrial Revolution, Koshy says, “There was also a very high increase in productivity compared to what was there earlier. In Russia, they decided, ‘Hey, we have reached peak productivity. Let’s just distribute the gains and enter into socialism.’ The problem with that approach is that it effectively caps your growth, resulting in further productivity at a very low rate once you enter socialism.” Koshy’s view is that a mixed economy where socialist elements are added on top of a capitalist system would be the way to go. It will retain the incentive to pursue further increases in productivity while providing a safety net for those in need of protection. The introduction of the Social Security Act in the US in the 1930s serves as an example of this approach, in contrast to the paths taken by Europe and Russia after the Industrial Revolution. Koshy believes that in the age of robotics and AI, governments can harness technology to ensure that even if someone is without a job, they are not without food to eat or a place to live. Widespread use of robotics can make that happen. Karan Kamble writes on science and technology. He occasionally wears the hat of a video anchor for Swarajya's online video programmes. Join our WhatsApp channel - no spam, only sharp analysis Comments ↓ About Swarajya Shaping the modern Indian's worldview, speaking on behalf of those invested in the cultural and economic prosperity of India. Published since 1956. Swarajya is a publication by Kovai Media Private Limited. Useful Links Useful Links Participate Stay Connected

Business Naver to debut new OS for robots at Saudi tech show Published : March 5, 2024 - 17:07:26 Link copied! Korean internet giant Naver said Tuesday that it is debuting the world’s first web platform-enabled operating system for robots, called “Arc mind,” at this week’s LEAP 2024, a tech trade show held in Saudi Arabia. The company said the new platform combines the web platform technology of Naver Cloud’s Whale and the robotics software technology of Naver Labs. "Team Naver has continuously expanded its tech portfolio from web platform to robotics software by developing and operating Whale browser, Whale OS, and the multi-robot intelligence system Arc (AI-Robot-Cloud). Such technical capabilities became the basis for Arc mind, which converges advanced technologies of the future," Naver Cloud General Leader Kim Hyo said at a press briefing in Seoul on Tuesday. "Arc mind will create the optimal environment to develop and provide a wider variety of robots and robotic services to future cities. This will open up a new field of robotics to web developers around the world and speed up the mass adoption of robotics in our everyday lives." Peck Jong-yoon, Naver Labs executive officer, also explained how its new OS aims to compete with Robot Operating System, the widely used open-source robotics OS. "We do want to build a new ecosystem similar to Android, but it's not the same. The difference is that it is not creating an entirely new OS, but using the existing web platform. Web platforms are universal and it is excellent in compatibility and productivity,” Peck said. To introduce a new service like a payment system for a delivery robot, the developers would have to start from scratch on most occasions. But with Arc mind, they can use existing payment systems such as Naver Pay to speed up the overall development process. Naver said it plans to implement Arc mind on its own robots first, with the ultimate plan of expanding it to a fully open ecosystem. At LEAP 2024 running from Monday to Thursday, Naver Labs CEO Seok Sang-ok plans to deliver a keynote speech, titled “Tech Convergence for Future Cities,” and the company has set up an exhibition booth to showcase its latest technologies such as digital twin, AI, cloud and Arc mind. The firm also plans to debut a new robotics platform, called Robotics Edge Computing Platform, a collaboration with Samsung Electronics. The new platform combines Samsung’s chip solutions and Naver’s software in a bid to accelerate the wider adoption of cooperative robots, the company added. Strong exports, weak currency: Why the won can’t catch a break The Korean won is expected to remain under pressure throughout 2026 as it confronts a new set of forces reshaping global financial hegemony — most notably a US-led artificial intelligence boom that is New year brings host of new rules More than 80% of South Koreans watch streaming Yearlong exercise halves depression risk: study [Graphic News] Year of the Fire Horse Russia, North Korea exchange New Year's gifts Diplomatic Circuit North Korean POWs’ transfer to S. Korea still pending Diplomatic Circuit K-pop 101 The world of K-pop explained, for both fans and newcomers Inside Gen Z Giving Korea's new generation a voice Oddities From the funny to the strange and downright unbelievable Herald Interview A series of in-depth interviews. Coupang’s W1.7tr payout plan fails to quell public anger Global banks see 1,400 won as new baseline What Georgia raid changed for journalists heading to CES Samsung develops first inhouse mobile GPU for Exynos 2600 Coupang founder apologizes for data breach amid mounting scrutiny Korea wants foreign students, but does it want the graduates? AI is no longer optional in K-pop — it’s becoming the new normal [Graphic News] Year of the Fire Horse Veteran actor Ahn Sung-ki in critical condition Fans mark BTS’ V’s 30th birthday with stone bench at Versailles Address : Huam-ro 4-gil 10, Yongsan-gu,Seoul, Korea Tel : +82-2-727-0114 Online newspaper registration No : Seoul 아03711 Date of registration : 2015.04.28 Publisher. Editor : Choi Jin-Young Juvenile Protection Manager : Choi He-suk The Korea Herald by Herald Corporation. Copyright Herald Corporation. All Rights Reserved.

Want to read in a language you're more familiar with? ByteDance releases GR-2 robot AI large model, with an average task completion rate of 97.7%, simulating human learning to handle complex tasks. The research team of ByteDance has recently launched the second-generation large-scale model GR-2 (Generative Robot 2.0). Its highlight lies in innovatively constructing the "robot infancy" learning stage, imitating human growth and learning complex tasks, possessing outstanding generalization ability and multitask versatility. Like many other AI models, the GR-2 model undergoes two processes: pre-training and fine-tuning. During the pre-training phase, GR-2 'watched' up to 38 million internet videos from various public datasets and 500 billion tokens, covering a wide range of daily scenes such as home, outdoor, and office environments. This enables GR-2 to have generalization capabilities across a wide range of robot tasks and environments in subsequent reinforcement learning. In the fine-tuning stage, the team used robot trajectory fine-tuning for video generation and action prediction, demonstrating outstanding multitasking capabilities, achieving an average success rate of 97.7% in over 100 tasks. In addition, GR-2 demonstrates excellent generalization ability in novel and previously unseen scenarios, including new backgrounds, environments, objects, and tasks. SEE ALSO: ByteDance’s Doubao Video Generation Large Model Released Related posts coming soon... Pandaily is a tech media based in Beijing. Our mission is to deliver premium content and contextual insights on China's technology scene to the worldwide tech community. © 2017 - 2026 Pandaily. All rights reserved.

A humanoid robot has stunned the tech world by nailing complex drum performances with over 90 percent rhythmic accuracy, even executing humanlike techniques such as stick switching and cross arm hits. Developed by researchers at SUPSI, IDSIA, and Politecnico di Milano, this “Robot Drummer” system relies on reinforcement learning rather than preprogrammed sequences, marking a milestone in machine creativity and physical coordination. Instead of feeding the robot exact instructions, its creators transformed drum scores into “Rhythmic Contact Chains” precise sequences of timed contact events and trained the model on over 30 rock, metal, and jazz tracks. In simulation, the robot mastered rhythmic timing while developing expressive behaviors such as planning strike patterns and adapting stick use spontaneously. The idea originated from a casual chat between researchers who recognized that humanoid robots rarely venture into the expressive realm of music. Drumming became a natural test case for combining rhythm, coordination, and creative expression. The team believes the robot could eventually perform live alongside human musicians. The team’s next ambition is transferring these simulated drumming skills to actual hardware, empowering the robot to improvise live based on musical cues essentially allowing it to respond to its environment just like a human drummer. Figures humanoid robots now march naturally using reinforcement learning trained in simulation and Boston Dynamics is enhancing agility through similar AI driven methods. This reflects a growing movement away from manual programming toward autonomous learning and adaptability in robotics. Drumming demands split second precision, full body coordination, and dynamic timing traits essential to musical artistry. That a robot can emulate these human qualities opens new doors for both entertainment and educational applications. Whether guiding music students or performing on global stages, robotic musicians may no longer be sci fi. Abdul Wasay explores emerging trends across AI, cybersecurity, startups and social media platforms in a way anyone can easily follow. The federal government and NADRA have officially amended the National Identity Card (NIC) Rules to widen the legal definition of biometrics. As of Wednesday, facial. Russia has introduced a new quantum computer prototype, signaling tangible progress in its long running effort to build advanced computing systems without relying on foreign. Taiwan Semiconductor Manufacturing Company has quietly begun volume production of its first 2nm class chips, marking a major manufacturing milestone as the world’s largest contract. Researchers have developed a new physical artificial intelligence system that enables electric vehicles to detect loss of control in real time by combining traditional physics-based. Premier Pakistan technology news website with special focus on startups, entrepreneurship and consumer products. © 2025 TechJuice.PK – All rights reserved.

From daily news and career tips to monthly insights on AI, sustainability, software, and more—pick what matters and get it in your inbox. Explore The Most Powerful Tech Event in the World with Interesting Engineering. Stick with us as we share the highlights of CES week! Access expert insights, exclusive content, and a deeper dive into engineering and innovation. Engineering-inspired textiles, mugs, hats, and thoughtful gifts We connect top engineering talent with the world's most innovative companies. We empower professionals with advanced engineering and tech education to grow careers. We recognize outstanding achievements in engineering, innovation, and technology. All Rights Reserved, IE Media, Inc. Follow Us On Access expert insights, exclusive content, and a deeper dive into engineering and innovation. Engineering-inspired textiles, mugs, hats, and thoughtful gifts We connect top engineering talent with the world's most innovative companies We empower professionals with advanced engineering and tech education to grow careers. We recognize outstanding achievements in engineering, innovation, and technology. All Rights Reserved, IE Media, Inc. The robot uses a combination of machine learning and a flexible “back” to chimney climb with ease. A team of researchers from Japan’s Jouhou System Kougaka Laboratory at the University of Tokyo (JSK) has developed a new robot that can both walk on uneven terrain and climb vertical walls. Called KLEIYN, the robot can scale vertical surfaces using a technique called chimney climbing (pressing its legs against two opposing walls for support). To support their paper on the subject, published in the journal arXiv, the team has also released footage of KLEIYN in action. The robot itself is a four-legged (quadruped) robot with an active waist joint (it can bend in the middle). According to its developers, the robot uses quasi-direct-drive motors for precise and powerful movement. All in all, the robot weighs approximately 40 pounds (18 kg), has 13 joints, and measures around 2.5 feet (76 cm) in length. Furthermore, the robot utilizes a form of machine learning called Reinforcement Learning (RL) in simulation to learn how to move and climb. The team also integrated a special training method called Contact-Guided Curriculum Learning (CGCL), which gradually teaches it how to transition from flat ground to vertical surfaces. Other technical innovations include Asymmetric Actor-Critic RL, an efficient training setup that enables the robot to learn from rich simulation data while utilizing only basic sensors in real-life applications. What makes KLEIYN different from the many other climbing robots out there is the design of its “feet.” Most climbing robots use grippers (like claws) to grab onto things, which limits their walking ability. Instead, KLEIYN uses chimney climbing, which works by pressing its feet against two walls, eliminating the need for grippers. The waist joint lets it adapt to different wall widths, especially narrow ones. This design appears to have paid off remarkably well. During test climbs of walls spaced 31.5 inches (80 cm) to 39.4 inches (1 m) apart, the robot was able to climb at a rate of 6 inches (15 cm) to 6.7 inches (17 cm) per second. That is roughly 50x faster than the previous best (SiLVIA). The robot was also able to walk on rough terrain and climb steps outdoors successfully. The robot can learn how to recover from slipping, making it more robust. But not all is going good for KLEIYN. According to the team, it struggles with any wall gaps wider than 77 inches (1.05 m) due to torque limits. It also moved sideways unintentionally during climbs, meaning it needs better environmental sensing (e.g., LiDAR input). The robot’s motors also tend to overheat during long climbs, suggesting a need for better load balancing. That said, this research pushes the boundaries of robot mobility, enabling one robot to handle both flat ground and vertical obstacles. That could make it valuable for tasks such as search and rescue in collapsed buildings, exploration of complex environments (like caves or disaster zones), and transportation tasks in uneven terrain. You can view the study for yourself in the journal arXiv. You can also check out the design team’s GitHub page for more information on the project. Christopher graduated from Cardiff University in 2004 with a Masters Degree in Geology. Since then, he has worked exclusively within the Built Environment, Occupational Health and Safety and Environmental Consultancy industries. He is a qualified and accredited Energy Consultant, Green Deal Assessor and Practitioner member of IEMA. Chris’s main interests range from Science and Engineering, Military and Ancient History to Politics and Philosophy. Premium Follow

0 By clicking the button, I accept the Terms of Use of the service and its Privacy Policy, as well as consent to the processing of personal data. Don’t have an account? Signup Karnataka teacher Akshay Mashelkar wanted to make education more interactive for children in his village, and so he built Shiksha, a humanoid robot that imparts education to children under Class 4. Karnataka teacher Akshay Mashelkar wanted to make education more interactive for children in his village, and so he built Shiksha, a humanoid robot that imparts education to children under Class 4. google-news Follow Us Dressed in a blue shirt and tunic, with neatly parted hair styled into two pleats, the humanoid robot named ‘Shiksha’ bears a striking resemblance to the rest of the students of Sirsa village. As she begins delivering the day’s lessons — from rhymers to the days of the weeks, names of different shapes, and more — there’s a sense of wonder in the eyes of each student as they take in this remarkable teaching experience. Shiksha is the brainchild of 30-year-old Akshay Mashelkar, and aims to make learning fun and interactive “Growing up in a village I was very aware of the limitations of schools in rural areas. We still use printed charts and blocks as a means of learning. There are no scientific methods available. I want to change that,” Akshay tells The Better India. Born and raised in the village of Sirsi in Karnataka’s Uttara Kannada district, Akshay grew up in a teaching household. “My mother was a teacher and from a very young age, I knew I wanted to become an educator too. While studying, I realised that I wanted to work towards improving the education system,” he says. Following in the footsteps of his mother, Akshay became a professor at a college in Sirsi after completing his degree in Physics. “While I enjoyed my job as a professor, I had many ideas to implement in the education system. With the work, there was no time for me to start working on it though,” he says. When the COVID pandemic hit and the education sector moved online, Akshay found himself relatively free. “I found the perfect opportunity to work on my ideas. One of the most important things that I have seen in the education sector, especially in Tier-2 and -3 cities and rural areas, is the lack of modern and scientific methods of teaching. On one of my several visits to schools in the village, I saw that teachers were still using charts and blocks to teach,” he says. “Those techniques were used when I was in school. It is sad that the world has advanced so much with smart boards and whatnot, but schools in rural areas are still stuck with handmade charts. This pushed me further to give all my attention to bringing an easier and cheaper solution,” he adds. It took Akshay a good one and half years to do the research. In 2022, ‘Shikha’ — a humanoid robot capable of teaching in regional languages up to Class 4 — was ready. In India, the education sector has been incorporating technology for teaching purposes for several years. Nevertheless, its implementation has primarily been observed in urban regions and expensive schools. On the contrary, rural schools continue to rely on conventional tools like charts and drawings to facilitate learning. Moreover, teachers of government schools are overburdened with students. A recent Quint report states, “The number of teachers in government schools in Karnataka has dropped from 2.08 lakh to 1.99 lakh, forcing 6,529 schools in the state to have only one teacher. The student-teacher ratio is now 23:1 when compared to 21:1 in 2020-21.” The inclusion of such a device could help fix this problem. The robot took nearly Rs 2 lakh to build, which he took out of his own savings. “A lot of money was involved in the research and development. On average, making only a robotic arm costs nearly Rs 50,000. ‘Shiksha’ is an entire robot with several features. The reason why I was able to cut costs was I used jugaad. For instance, I did not use a mould for the body of the robot, instead for the arms I used plastic cricket stumps that you find in toy shops,” he says. Siksha can teach various subjects including rhymes in Kannada and English rhymes; the days of the week; names of shapes; English alphabets, and maths topics such as multiplication, addition and tables. Explaining how the robot works, Akshay says, “The robot has two main cards — the master card that unlocks it, and the normal card to start the desired programme. The teacher has to put the master card on Shiksha’s hand to start it and then they can use the programme cards to start different programmes. She moves her arms to take the card and returns it once scanned. She asks questions, recites poems, and even has trivia options,” he says. The robot has visited over 25 schools in the Uttar Kannada district, including KHB School and Urdu School in Sirsi. So far, Shiksha can teach up to Class 4 and has syllabus accommodation across boards. Sunaina Hegde, who teaches Science and Maths at Model Higher Primary School in Sirsi, says, “Akshay came with Shiksha to our school in April. The children were so happy to see her and they took a greater interest in the class. For them, Shiksha was not a robot, but more like a friend as it was dressed like them too.” “While it is great for students to learn, it is also a great tool to be included in schools for teachers. It reduces our burden, as there are fewer teachers in government schools. Something so interactive helps children to gain more interest in science and technology,” she adds. Akshay notes, “The importance of involving village children in technology is because they are also the future of the country. An average child living in an urban setting, from a very young age, knows how to operate laptops and computers. Sadly, this is not true for kids in rural areas. When the kids saw Shiksha for the first time, I could see the sparkle in their eyes. They were intrigued, amazed and excited.” “My motive behind making Siksha was not only to introduce technology in the classroom but also to encourage children to make their robots,” he says. Taking the thought forward, Akshay also opened a research centre where young robotics enthusiasts can come and learn for free. “In order to keep the cost of the operations of the centre low, we keep our centre mobile. Whenever we get low-cost places to rent in Sirsi, we move to that place. Over 200 children have visited the centre and many are regulars now. They have the space to learn from me and use the tool available in the research centre,” he says. Although the first Shiksha cost him lakhs, Akshay says that he can reduce the cost even more. “There were a lot of errors and a lot of investment in R&D initially, but now there won’t be. With the help of grants and support from the government and NGOs, I can possibly reduce the cost to Rs 35,000. This way it will be cheaper to afford rural schools. My only wish is to take Shiksha to every rural school in Karnataka and make learning fun,” he adds. If you wish to know more about his research centre and be a part of his initiative, you can reach him at 74832 76508. (Edited by Divya Sethu) Subscribe to our Newsletter! Quick Links

The TOI Tech Desk is a dedicated team of journalists committed to delivering the latest and most relevant news from the world of technology to readers of The Times of India. TOI Tech Desk’s news coverage spans a wide spectrum across gadget launches, gadget reviews, trends, in-depth analysis, exclusive reports and breaking stories that impact technology and the digital universe. Be it how-tos or the latest happenings in AI, cybersecurity, personal gadgets, platforms like WhatsApp, Instagram, Facebook and more; TOI Tech Desk brings the news with accuracy and authenticity.Read More The rainbow of the wild: 10 colourful and rare insects across the globe From Neeru Bajwa to Shehnaaz Gill: Leading ladies shaping Punjabi cinema’s bright future In pics: Vijay Sethupathi’s fierce avatar in Bigg Boss Tamil 9 promo 10 small dog breeds perfect for apartment living: Family-friendly and easy to care for Disha Parmar approved top 10 stylish looks Navratri 2025: Shraddha Kapoor, Alia Bhatt, and other divas' inspired pink ethnic looks for day 9 From Chic to Classy: Priyanka Mohan’s style evolution Mamitha Baiju mesmerizes; nature meets grace Monalisa stuns in red saree for Durga Puja Best family games that will make your kids forget all about screens

Digital Replica Edition Sign up for Newsletters and Alerts Sign up for Newsletters and Alerts Digital Replica Edition Trending: LEGO has been making childhood toys for more than 50 years and now even offer kits that allow your child to build a robot. These sets encourage young kids to explore and get excited about STEM. LEGO robot kits are great gifts that combine fun with learning to build, code and program. They come in variations designed for every age and skill level. If you or your child wants to build a fully functional intelligent robot that walks, talks, plays games and completes many different tasks, try the LEGO Mindstorm EV3 31313 Robot Kit. Start by choosing one of two main types of robot kits: model robots or programmable robots. The greater the detail involved, the more time and effort required to build the robot. Programmable robot kits will always be more detailed than model robot kits. Look for two key indicators: the number of pieces included and the suggested age range. More detailed and complicated kits will have at least 500 separate pieces. Robot kits with high levels of detail will require tools to assemble and will also be more expensive. Small programmable robot kits have one motor while larger and more detailed robot kits will have two or even three motors. High-tech add-ons like transmitters and sensors have electrical components that need more than one motor to operate properly. You will find simple model kits for $15-$25 and more detailed ones for $100 or more. Programmable model kits start at around $150 and increase in cost as you add pieces, motors, sensors and possibilities. A. Yes, on two different levels. LEGO Boost kits teach younger kids the basic rules of programming through the use of drag-and-drop icons. Those who choose LEGO Mindstorm kits learn how to code by writing their own programs using more complex processes. A. Anyone can have fun with a LEGO robot kit. If you enjoy technology, try using a more complex model. If you’re new to coding, a simpler kit is a good start. LEGO Mindstorm EV3 31313 Robot Kit What you need to know: This fully functional intelligent robot kit allows you to create five different robots that walk, talk and play games. What you’ll love: Builders of all ages will enjoy creating and commanding their own 16- by 15- by 14-inch robot with this 601-piece kit. It includes the intelligent EV3 Brick, three motors and color, touch and infrared sensors. What you should consider: This is a pricey robot that requires some technical and programming skills. LEGO Star Wars VIII BB-8 75187 Building Kit What you need to know: This Star Wars droid has more than 1,100 parts and is designed for kids 10 to 16. What you’ll love: The detailing is authentic. Turn one wheel at the side to rotate the head and another to open the access hatch and extend the welding torch. It also comes with a display stand, decorative fact plaque and mini-figure. What you should consider: This kit is expensive for a model that can’t be programmed. Prices listed reflect time and date of publication and are subject to change. Check out our Daily Deals for the best products at the best prices and sign up here to receive the BestReviews weekly newsletter full of shopping inspo and sales. BestReviews spends thousands of hours researching, analyzing and testing products to recommend the best picks for most consumers. BestReviews and its newspaper partners may earn a commission if you purchase a product through one of our links. Copyright © 2026 MediaNews Group

e-Edition Get the latest news delivered daily! Get the latest news delivered daily! e-Edition Trending: LEGO has been making childhood toys for more than 50 years and now even offer kits that allow your child to build a robot. These sets encourage young kids to explore and get excited about STEM. LEGO robot kits are great gifts that combine fun with learning to build, code and program. They come in variations designed for every age and skill level. If you or your child wants to build a fully functional intelligent robot that walks, talks, plays games and completes many different tasks, try the LEGO Mindstorm EV3 31313 Robot Kit. Start by choosing one of two main types of robot kits: model robots or programmable robots. The greater the detail involved, the more time and effort required to build the robot. Programmable robot kits will always be more detailed than model robot kits. Look for two key indicators: the number of pieces included and the suggested age range. More detailed and complicated kits will have at least 500 separate pieces. Robot kits with high levels of detail will require tools to assemble and will also be more expensive. Small programmable robot kits have one motor while larger and more detailed robot kits will have two or even three motors. High-tech add-ons like transmitters and sensors have electrical components that need more than one motor to operate properly. You will find simple model kits for $15-$25 and more detailed ones for $100 or more. Programmable model kits start at around $150 and increase in cost as you add pieces, motors, sensors and possibilities. A. Yes, on two different levels. LEGO Boost kits teach younger kids the basic rules of programming through the use of drag-and-drop icons. Those who choose LEGO Mindstorm kits learn how to code by writing their own programs using more complex processes. A. Anyone can have fun with a LEGO robot kit. If you enjoy technology, try using a more complex model. If you’re new to coding, a simpler kit is a good start. LEGO Mindstorm EV3 31313 Robot Kit What you need to know: This fully functional intelligent robot kit allows you to create five different robots that walk, talk and play games. What you’ll love: Builders of all ages will enjoy creating and commanding their own 16- by 15- by 14-inch robot with this 601-piece kit. It includes the intelligent EV3 Brick, three motors and color, touch and infrared sensors. What you should consider: This is a pricey robot that requires some technical and programming skills. LEGO Star Wars VIII BB-8 75187 Building Kit What you need to know: This Star Wars droid has more than 1,100 parts and is designed for kids 10 to 16. What you’ll love: The detailing is authentic. Turn one wheel at the side to rotate the head and another to open the access hatch and extend the welding torch. It also comes with a display stand, decorative fact plaque and mini-figure. What you should consider: This kit is expensive for a model that can’t be programmed. Prices listed reflect time and date of publication and are subject to change. Check out our Daily Deals for the best products at the best prices and sign up here to receive the BestReviews weekly newsletter full of shopping inspo and sales. BestReviews spends thousands of hours researching, analyzing and testing products to recommend the best picks for most consumers. BestReviews and its newspaper partners may earn a commission if you purchase a product through one of our links. Copyright © 2026 MediaNews Group

Fasciné par le fonctionnement complexe du cerveau animal, Thomas Deneux, chercheur au CNRS, décide de mettre au point une IA permettant de visualiser des réseaux de neurones. Il fonde alors Learning Robots, startup qui propose un logiciel capable d'entraîner un robot pour des activités simples et de visualiser le fonctionnement de ses neurones, et s’associe avec Axel Haentjens. L’objectif ? Sensibiliser le grand public à l’IA en la démystifiant grâce à une vision claire sur ses mécanismes et son apprentissage. « Avec la généralisation des chatbot et des IA génératives comme Chat GPT, l’IA est partout, sauf que peu de gens comprennent comment cela fonctionne ! », rappelle Axel Haentjens, Directeur Général de Learning Robots. « Nous voulons changer la vision du public sur l'IA, la démystifier et la rendre accessible à tous, et pas seulement à une élite. Or, c'est en manipulant et en expérimentant que l'on apprend le mieux.» Grâce à un logiciel, il est possible de visualiser en temps réel les décisions prises par un petit robot, l’AlphAI, et d’expliquer les processus d'algorithmes sous-jacents qui permettent l’apprentissage de la machine. Cette approche rend l'IA tangible et compréhensible, sans concepts abstraits souvent difficiles à saisir, grâce à la mise en pratique. N’importe quel public peut comprendre le fonctionnement de l’IA, comment ses neurones fonctionnent et même observer comment le robot « voit » et prend ses décisions. Les évolutions du robot, au fur et à mesure de son apprentissage, permettent de mieux comprendre l’importance de la qualité des données dans son évolution. « L’IA se trompe beaucoup au début. Mais plus elle apprend, mieux elle va fonctionner. Cela montre aux enfants qu’il est normal de faire des erreurs, pour mieux comprendre et apprendre », précise le Directeur Général. « En entraînant ce robot, tout le monde peut comprendre ce qu’est un algorithme. On peut commencer à l’utiliser dès la classe de sixième, pour sensibiliser à l’IA, puis en seconde, on peut rentrer dans le détail des calculs des algorithmes, et enfin dans le supérieur, commencer à coder ces algorithmes. » La jeune pousse vend des kits robotiques et des licences logicielles pour le suivi et l'entraînement des robots aux établissements. Plusieurs centaines d’établissements, comme l’Université Paris-Saclay, utilisent déjà cette technologie. Néanmoins, si Learning Robots s’intègre parfaitement dans le système éducatif, elle a su trouver sa place également dans le milieu professionnel : des entreprises, notamment du CAC 40, utilisent déjà cette technologie pour former leurs cadres et les aider à comprendre l’IA et réduire leurs appréhensions. Après avoir remporté le Prix International Next Innov by Banque Populaire Val de France, la startup souhaite désormais s’étendre à l’international pour développer de nouvelles fonctionnalités, notamment la connexion de son logiciel à des modèles de langage génératif et ainsi faciliter les interactions homme-machine. Des discussions sont en cours pour établir des partenariats à Hong Kong, en Suisse, au Brésil, en Allemagne et au Royaume-Uni. « Lorsque nous avons entendu parler de Next Innov, j’ai directement flashé sur le nom : Next, c’est préparer le coup d’après. Et innover, c’est notre ADN », conclut Axel Haentjens. « Au-delà des aspects financiers, ce prix va nous permettre d’obtenir une certaine notoriété, la Banque Populaire Val de France étant très connue, et de disposer d’un rayonnement médiatique certain, d’un accompagnement et d’un réseau : ce sont des rencontres qui entraînent des rencontres, et cela est passionnant ».

In a striking demonstration of robotic agility, Tesla’s Optimus humanoid robot has been captured on video executing a series of kung fu moves with remarkable fluidity, mirroring the motions of a human trainer. Shared by Elon Musk on the social platform X, the 36-second clip shows Optimus blocking strikes, dodging attacks, and performing precise hand gestures, all while maintaining balance on a mat. This isn’t just a parlor trick; it highlights significant advancements in Tesla’s AI-driven robotics, where the machine learns complex physical tasks through imitation and self-correction, without direct human intervention during the performance. The video, which has garnered millions of views, underscores Optimus’s evolution from earlier prototypes that struggled with basic locomotion to a system capable of nuanced, dynamic interactions. Musk emphasized that the robot’s actions are powered entirely by artificial intelligence, with no teleoperation or pre-programmed sequences involved. This leap forward comes amid Tesla’s broader push into humanoid robotics, aiming to deploy these machines for factory work, household chores, and beyond. Advancements in AI Training and Hardware Engineers at Tesla have reportedly trained Optimus using end-to-end neural networks, similar to those in the company’s autonomous driving systems, allowing the robot to process visual data and translate it into physical responses in real time. According to a report from Digital Trends, the display is impressive for an adult-sized robot, showcasing speed, balance, and accuracy that rival human capabilities. The publication notes that while earlier videos featured Optimus dancing or folding laundry, this kung fu routine demonstrates improved joint flexibility and coordination, potentially paving the way for applications in security or entertainment. Comparisons to competitors like Boston Dynamics’ Atlas or Figure’s humanoid bots reveal Tesla’s unique edge: scalability. Musk has projected low-volume production for internal use by next year, with high-volume output targeted for 2026. Insights from Interesting Engineering highlight how Optimus’s kung fu training refines its motor skills, drawing from simulation-based learning where virtual models endure millions of iterations before real-world deployment. Implications for Industry and Ethical Considerations Industry insiders see this as a milestone in humanoid robotics, potentially disrupting labor markets in manufacturing and service sectors. Tesla’s approach leverages its vast data from electric vehicles to accelerate robot development, with Optimus priced under $30,000 per unit for mass appeal. A piece in The Times of India describes the robot flawlessly replicating kung fu alongside a trainer, emphasizing AI’s role in achieving human-like balance without engineer oversight. However, such capabilities raise questions about safety and job displacement. If robots can master martial arts, their potential in hazardous environments—like disaster response—becomes evident, but so do concerns over misuse. Musk has addressed this by stressing ethical AI frameworks, though skeptics on platforms like X argue for more transparency in training data. Future Trajectories and Competitive Pressures Looking ahead, Tesla plans to integrate Optimus with its Grok AI for conversational abilities, expanding beyond physical feats. Posts on X from robotics enthusiasts, including those analyzing the video’s frame-by-frame mechanics, suggest this kung fu demo is a precursor to more autonomous behaviors, like adaptive learning in unstructured settings. Rivals are not idle; companies like Unitree have showcased similar agile robots, but Tesla’s ecosystem integration gives it an advantage. As reported by TechEBlog, Optimus V3’s moves start with a fist touch, evolving into sparring that blurs lines between machine and human. For industry leaders, this signals a new era where robots don’t just assist but emulate human prowess, promising efficiencies while challenging us to redefine work in an automated world. In wrapping up, Tesla’s kung fu-capable Optimus isn’t merely entertainment—it’s a harbinger of robotics’ mainstream integration, driven by relentless innovation and AI prowess. Subscribe for Updates Help us improve our content by reporting any issues you find. Get the free daily newsletter read by decision makers Get our media kit Deliver your marketing message directly to decision makers.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Advertisement Scientific Reports volume 15, Article number: 32506 (2025) Cite this article 4434 Accesses Metrics details In modern industrial and laboratory environments, robotic arms often operate in complex, cluttered spaces. Ensuring reliable obstacle avoidance and efficient motion planning is therefore essential for safe performance. Motivated by the shortcomings of traditional path planning methods and the growing demand for intelligent automation, we propose a novel reinforcement learning framework that combines a modified artificial potential field (APF) method with the Deep Deterministic Policy Gradient algorithm. Our model is formulated in a continuous environment, which more accurately reflects real-world conditions compared to discrete models. This approach directly addresses the common local optimum issues of conventional APF, enabling the robot arm to navigate complex three-dimensional spaces, optimize its end-effector trajectory, and ensure full-body collision avoidance. Our main contributions include the integration of reinforcement learning factors into the APF framework and the design of a tailored reward mechanism with a compensation term to correct for suboptimal motion directions. This design not only mitigates the inherent limitations of APF in environments with closely spaced obstacles, but also improves performance in both simple and complex scenarios. Extensive experiments show that our method achieves safe and efficient obstacle avoidance with fewer steps and lower energy consumption compared to baseline models, including a TD3-based variant. These results clearly demonstrate the significant potential of our approach to advance robot motion planning in practical applications. Multidegree-of-freedom robotic arms are an integral part of today’s industrial automation, manufacturing, and service applications. As these systems are deployed in increasingly complex and dynamic environments, achieving robust motion planning and reliable obstacle avoidance has become a critical challenge. In many practical scenarios, the robot must navigate through cluttered spaces and avoid collisions while optimizing its trajectory for efficiency and safety. Efficient motion planning is essential for safely performing tasks in environments that are not only cluttered but also subject to dynamic changes. Conventional planning methods often struggle with adapting to unpredictable obstacles or high-dimensional configuration spaces. Traditional methods can be broadly categorized into searching-based, bionic and evolutionary, sampling-based, and gradient-based techniques. Searching-based algorithms, such as A*1, D*2, Dijkstra3, and simulated annealing4, are efficient in static, small-scale environments. However, they become impractical in large-scale, dynamic scenarios. Bionic methods like Genetic Algorithms (GA)5, Particle Swarm Optimization (PSO)6, Ant Colony Optimization (ACO)7, and fuzzy logic8 can handle more complex environments but often suffer from slow convergence and local optima. Sampling-based approaches, such as Rapidly-Exploring Random Trees (RRT)9 and Probabilistic Roadmaps (PRM), can navigate intricate spaces but at the cost of computational complexity. Gradient-based methods, especially the artificial potential field (APF) method10, offer intuitive and computationally efficient solutions by modeling the workspace as a potential field. APF methods attract the end-effector toward the goal and repel it from obstacles, yet they are inherently vulnerable to local minima problems where the robot becomes trapped due to balanced attractive and repulsive forces11,12. To address these limitations, hybrid strategies have been proposed. Integrating methods such as RRT with APF9, ACO with APF13, and Dynamic Window Approach (DWA) with GA14 have been explored. Researchers have also enhanced APF directly by modifying repulsive field functions or introducing auxiliary constructs. Cao15 proposed a velocity potential field (VPF) using robot velocity instead of distance, Flacco et al.16 leveraged repulsive vectors for collision avoidance, and Zhang et al.17 adjusted repulsive force functions to mitigate local minima. In parallel, reinforcement learning (RL) methods, particularly Deep Deterministic Policy Gradient (DDPG), have emerged as robust tools for continuous control problems18. Early RL algorithms, such as Deep Q-Networks (DQNs)19, DDPG18, and SARSA20, enabled agents to learn optimal policies by interacting with their environment. Recent advancements have further expanded these techniques. Chen21 utilized DDPG, TD3, and SAC effectively in robotic tasks, while Imtiaz22 demonstrated the versatility of PPO in fruit-picking applications. Improvements such as the multi-critic mechanism by Sze23 and refined reward functions by Li et al.24 have enhanced policy evaluation and APF optimization, respectively. Hybrid approaches combining RL with traditional methods, like RRT with TD325, DWA with Q-learning26, and PRM with RL27, further illustrate the complementary potential of these strategies. Despite these advancements, critical challenges persist, notably the local minima issue in APF methods and the sensitivity of RL algorithms to reward function design. The APF method’s vulnerability to local minima arises when nearby obstacles counterbalance the attractive goal force, trapping the robot prematurely. Conversely, DDPG and similar RL methods can learn complex continuous action policies but often struggle with limited or misleading rewards in complex environments. Our proposed approach integrates a modified APF mechanism within the DDPG framework, combining APF s structural guidance to alleviate local minima and DDPG s adaptive learning capabilities. Specifically, we enhance the reward signal with APF-derived attractive and repulsive components. Our state space includes both joint angles and end-effector positions, and actions comprise fine-grained joint rotations. This hybrid strategy enables efficient obstacle avoidance and optimized motion planning in a fully continuous state and action space, in contrast to the discrete formulations used in some prior work. While some prior approaches combine APF with RL methods such as Q-learning or PPO28,29, they typically rely on discretized state spaces, task-specific constraints, or lack the fine-grained control needed for articulated robotic arms. Others focus primarily on mobile robots, where the configuration space is of lower dimensionality and local minima are easier to mitigate. In contrast, our approach is explicitly designed for continuous 6-DOF robotic arms operating in 3D space, where both configuration complexity and the risk of local minima are substantially higher. Moreover, our APF-informed reward function is designed to remain informative throughout training, facilitating faster convergence and more stable behavior across environments of varying complexity. The contributions of our work are threefold: First, we present a novel APF-DDPG hybrid framework that combines the advantages of traditional and learning-based methods in the continuous setting. Second, we propose an improved reward mechanism explicitly designed to mitigate local minima. Finally, we validate our framework through extensive experiments across diverse environments, demonstrating significant improvements in convergence, solution quality, and collision avoidance compared to baseline models. Our findings provide robust solutions for real-world robotic arm motion planning. In this work, we employ a standard 6-degree-of-freedom industrial manipulator30 as our experimental platform. The kinematic model of the robot is derived using the standard Denavit–Hartenberg (DH) convention. According to this convention, each joint is assigned a coordinate frame, enabling the definition of spatial transformations between consecutive joints. Using these DH parameters, we calculate the forward kinematics to determine the robot configuration. Each transformation matrix \({}^{i-1}\mathbf{T}{i}\) from joint \(i-1\) to joint i is defined as The complete forward kinematic transformation from the base frame to the end-effector frame is obtained by multiplying these matrices sequentially: From \({}^{0}\mathbf{T}_{6}\), the end-effector position (x, y, z) is extracted and used for collision detection and distance computations in the obstacle avoidance algorithm. By incorporating forward kinematics directly into our methodology, we simplify computational requirements while ensuring accurate spatial data for our reinforcement learning framework. In this work, we propose an obstacle avoidance approach that integrates APF with DDPG reinforcement learning31. This integration addresses critical limitations of each method individually. APF methods often encounter local minima problems, particularly when attractive and repulsive forces oppose each other32. Pure DDPG methods, in contrast, typically require extensive exploration and lack direct spatial information, potentially slowing convergence and complicating navigation. The combined APF-DDPG framework thus leverages structured spatial information from APF and the robust exploratory learning ability of DDPG, facilitating efficient and reliable obstacle avoidance in continuous state-action scenarios. We define the obstacle avoidance problem as a Markov Decision Process (MDP), described by a state space \(\mathscr {S}\), action space \(\mathscr {A}\), and a reward function R(s, a). The continuous state vector \(s \in \mathbb {R}^{9}\) comprises the robot s six joint angles \((\theta _1,\dots ,\theta _6)\) and the end-effector’s Cartesian coordinates (x, y, z) computed from forward kinematics. The action vector \(a \in \mathbb {R}^{6}\) specifies incremental joint angle adjustments within the interval of \([-0.5^\circ , 0.5^\circ ]\) per timestep, allowing smooth and controlled movements. The APF-DDPG framework consists of actor and critic neural networks, their respective target networks, and an experience replay buffer for stable training, as illustrated in Fig. 1. At each timestep, the actor network outputs continuous incremental actions given the current state, which are then executed by the robot. The critic network evaluates state-action pairs by estimating the expected cumulative rewards (Q-values). Integrated APF-DDPG framework for continuous obstacle avoidance. The main networks (actor and critic) use spatial information from APF directions to guide policy updates. Target networks stabilize training. Experiences \((s,a,r,s')\) are sampled from the replay buffer to decorrelate training updates. The networks are trained using mini-batches sampled from the replay buffer, where each stored experience tuple \((s,a,r,s')\) captures interactions with the environment. To ensure stable training and smooth convergence, we follow32 and employ soft target network updates, defined by where \(\theta '\) are target network parameters, \(\theta\) are main network parameters, and \(\tau\) being the smoothing coefficient. We integrate APF into our DDPG framework to explicitly inform the actor network of desirable spatial directions. To do this, we first recall the standard artificial potential field definition10 where \(q\) is the end-effector position, \(q_{\textrm{goal}}\) the goal position, \(\rho (q)\) the distance to the nearest obstacle, and \(\rho _0\) its influence radius. Forces arise as the negative gradients of these potentials: The net direction of the APF is then \(F_\textrm{net} = F_a + F_r\). However, when \(F_a\) and \(F_r\) directly oppose each other, the robot can become trapped in a local minimum (\(\Vert F_\textrm{net}\Vert \approx 0\)). To systematically escape such traps, drawing inspiration from the classical BUG principle33, we add a small orthogonal perturbation. Specifically, we define Because \(F_\perp\) is normal to the plane spanned by \(F_a\) and \(F_r\), this small perturbation guarantees movement off any planar saddle without significantly diverting the primary APF direction. The adjusted force \(F_{\textrm{adjusted}}\) is then provided to the actor network at each time step, biasing policy updates toward collision-free, goal-directed motion. The compensation mechanism is schematically illustrated in Fig. 2. Illustration of perpendicular compensation in APF direction. When repulsive and attractive forces nearly oppose each other, a perpendicular correction is introduced, steering the robot away from local minima. The reward function is carefully designed to balance obstacle avoidance and goal-oriented behavior by combining two components: proximity to the goal and directional alignment with APF guidance. Formally, we write where \(\textrm{distanceGE}\) is the Euclidean distance between the end-effector and the goal in Cartesian space, explicitly encouraging movements toward the goal. The term \(\cos (\sigma )\in [-1,1]\) measures the alignment between the executed displacement vector \(\overrightarrow{\textrm{EE}}\) and the APF-suggested vector \(\overrightarrow{\textrm{APF}}\), computed as We choose \(\mu >1\) so that \(\mu (\cos \sigma -1)\le 0\) with equality only when \(\cos \sigma =1\), ensuring every deviation from perfect alignment incurs a negative reward and thereby shaping the agent to follow efficient, APF-consistent motions. To incorporate safety constraints we distinguish collision scenarios by using separate distance-weighting parameters, defining where \(\lambda _{c}>\lambda _{nc}\) is chosen based on the ratio between the workspace size and the robot arm length so that the distance penalty scales appropriately with the environment s spatial dimensions relative to the manipulator. We allow collisions during training rather than terminating episodes to promote extensive exploration of the state action space and to enable the agent to learn both avoidance and recovery strategies. Unlike pure APF methods that stall in local minima, our formulation maintains a negative distance term \(R_{\textrm{dist}} = -\lambda \,\textrm{distanceGE}\) until the goal is reached. We apply a per-step penalty to prevent hesitation or oscillation within shallow basins. We allow non-terminating collisions with a larger penalty \(\lambda _{c}\), encouraging risky detours and subsequent recovery to uncover escape pathways. We use the alignment term \(R_{\textrm{align}} = \mu (\cos \sigma - 1)\le 0\) to penalize moves that conflict with APF guidance. As a result, the agent converges on efficient, collision-free trajectories. Due to the absence of specialized datasets for robotic arm path planning and obstacle avoidance, we developed a series of custom environments to simulate diverse real-world scenarios. Each obstacle within these environments is assigned a region characterized by a specific repulsion scaling factor, resulting in localized repulsive fields that directly affect the agent s trajectory planning. All environments share a common workspace dimension of \(10 \times 10 \times 10\) units. Spherical obstacles are placed strategically and vary in radius from \(0.2\) to \(0.5\) units, while the robotic arm itself, when fully extended, approximates a cylinder of length \(10\) units with a radius of \(0.1\) units. We designed two distinct categories of task environments to rigorously evaluate the proposed approach: Simple task environments These environments include fewer than ten spherical obstacles strategically arranged to create deliberate local optima traps, as depicted in Fig. 3a. These scenarios provide baseline conditions to assess fundamental obstacle avoidance capabilities. Left: Example of a simple task environment containing fewer than ten spherical obstacles positioned to create local optima traps. Right: Example of a complex task environment, illustrating increased obstacle density and strategic placement to enhance task difficulty. (These images were made with Unity-2022.3.18f1, https://unity.com/de). Complex task environments These scenarios incorporate a higher number and strategic complexity of obstacles, significantly increasing the likelihood of collisions and entrapment in local optima. Obstacles are placed near the initial position to hinder initial movements and along intended trajectories to elevate the complexity of the planning problem. Figure 3b illustrates typical complex environments. An additional challenging scenario includes environments featuring cuboid wall obstacles (0.5 units thickness, 5 units width, and 10 units height), as shown in Fig. 4. The initial placement (Fig. 4a), along with the front (Fig. 4b) and side (Fig. 4c) views of the final state, emphasize how the dimensions of the wall impose substantial restrictions on the agent’s movement, covering approximately 30–40% of the width of the workspace and 50% height, providing a strict test for obstacle avoidance capabilities. A complex task environment featuring cuboid wall obstacles, significantly restricting movement space compared to spherical obstacles. (These images were made with Unity-2022.3.18f1. The experimental setup comprises a total of 30 simple and 30 complex environments. Additionally, multi-goal trajectory planning tasks were designed, where the final state of one iteration serves as the initial state of the next. Task completion is achieved when the end-effector reaches the goal within a 0.1-unit tolerance, optimizing for computational efficiency and practical accuracy. For intuitive visualization and precise control of the robotic agent, Unity, a real-time 3D development platform, is employed. Control scripts are implemented using C# to directly interact with the Unity simulation environment. Experiments were conducted on a computing system with an 11th Gen Intel(R) Core(TM) i5-11320H processor and NVIDIA GeForce MX450 GPU. The reinforcement learning framework was implemented using PyTorch 2.0.1. Detailed hardware and software configurations are summarized in Table 1. The choice of baseline methods was motivated by their contrasting characteristics, strengths, and limitations relevant to robotic arm path planning. To fully evaluate the effectiveness of the proposed DDPG-APF approach, we compare with the following baselines: A* Search (Non-RL) A classical best-first search algorithm operating in a discretized joint-action space (rotations of − 1 , 0 , + 1 per joint). A* is run without a step limit, relying on an admissible and consistent heuristic to guarantee optimality in discrete settings34. This method serves as a non-learning baseline, highlighting the challenges of discretization and potential local-optima loops in our custom environments. Pure-DDPG A standard Deep Deterministic Policy Gradient model without Artificial Potential Field guidance, using only collision penalties and distance-to-goal in its reward function. This baseline isolates the benefit of APF integration and demonstrates the intrinsic performance limitations and convergence challenges faced by reinforcement learning without heuristic guidance. TD3-APF Twin Delayed DDPG with the same APF-based reward as our DDPG-APF agent. TD3 employs dual critics and delayed actor updates to address DDPG’s overestimation bias35, enabling a direct comparison under identical reward and network settings to assess whether these algorithmic improvements provide tangible benefits in complex, continuous task spaces. All reinforcement learning methods share identical network architectures and training regimens, ensuring a fair evaluation of feasibility, optimality, computational efficiency, and convergence behavior. Together, these methods provide a comprehensive performance context, ranging from classical planning paradigms to modern deep reinforcement learning techniques. We evaluated the performance of four distinct algorithms: DDPG-APF, Pure-DDPG, TD3-APF, and the classical A* search algorithm. This evaluation was conducted across tasks of varying complexity, focusing specifically on path optimality, solution feasibility, computational efficiency, and learning dynamics. All RL algorithms were trained under identical hyperparameters, explicitly detailed in Table 2. Initially, we analyzed training dynamics and convergence characteristics. The progression of best solutions, quantified by the number of steps required to reach the goal, revealed clear distinctions among the evaluated methods. Initially, all algorithms maintained a step count at the maximum threshold of 200 steps, indicating the absence of early feasible solutions. Upon discovering feasible solutions, the DDPG-APF algorithm exhibited notably faster convergence, consistently achieving shorter paths at an earlier stage, as demonstrated in Fig. 5. Feasible solution evaluation illustrating episodes required for finding the first feasible path. DDPG-APF identifies feasible paths sooner and more consistently than other methods, reflecting superior stability and efficiency. While the episode count to first feasible solution quantifies when each agent begins to find successful paths, analyzing the evolution of the average reward provides deeper insights into how effectively each policy learns over time. Specifically, tracking rewards captures both the stability of learning dynamics and the efficiency with which algorithms balance exploration, collision avoidance, and path-length minimization (see Fig. 6). Figure 6a illustrates the overall average reward curves across all task environments. Here, DDPG-APF consistently achieves a higher average reward compared to Pure-DDPG and TD3-APF, signaling more efficient policy improvements and a superior exploration-exploitation trade-off. In contrast, Pure-DDPG demonstrates a smoother but slower progression, reflecting a stable yet less efficient learning pattern, whereas TD3-APF displays noticeable fluctuations, likely resulting from increased complexity associated with its dual-critic network structure. Comparison of the presented reward functions: (a) overall model performance across all tasks; (b) sensitivity to reward-function design. We further assessed the sensitivity of the RL algorithms to variations in reward function design. Two distinct reward functions were tested: the original APF method without modifications and our proposed compensated APF reward function. Reward curves averaged across all tasks, displayed in Fig. 6b, demonstrated that DDPG-APF maintained stable performance regardless of reward function variations. Conversely, TD3-APF exhibited greater sensitivity, likely due to additional complexity in its dual-network architecture aimed at mitigating value function overestimation. In comparing RL methods to classical non-learning approaches, we evaluated the A* search algorithm, which operates using discrete action spaces and heuristic search. A comprehensive summary of key performance metrics across all 60 scenarios is provided in Table 3. Despite having unlimited search steps per episode, A* exhibited significantly lower success rates compared to the RL-based methods. Specifically, DDPG-APF demonstrated superior performance in average steps, success rate, and episodes required to achieve first success. Moreover, a dispersion analysis shows that DDPG-APF s variability in average steps is roughly half that of Pure-DDPG and one-third that of TD3-APF, and its variability in episodes to first success is about 45% lower than Pure-DDPG and over 65% lower than TD3-APF clear evidence of its consistency across diverse scenarios. The results highlight the advantages of continuous action adjustments offered by DDPG-APF, particularly in scenarios sensitive to local optima. Further evaluation against other RL metrics, such as time per episode and stability of solution improvement, reinforces the robustness of the DDPG-APF method. DDPG-APF consistently outperformed Pure-DDPG and TD3-APF across all metrics, particularly in computational efficiency (Time/Episode), indicating faster convergence and more effective exploration strategies. Notably, DDPG-APF reduces time-per-episode variability by approximately 25% relative to TD3-APF, underscoring its reliable runtime behavior. TD3-APF, while theoretically robust, showed notably higher variability and longer episodes to achieve first success, highlighting challenges associated with its dual-critic architecture. Computational efficiency was further analyzed by measuring average runtime per episode across tasks with different complexity levels. Table 4 summarizes these results, showing DDPG-APF consistently offering computational advantages over TD3-APF and Pure-DDPG. These runtime trends align well with our task complexity classification. Average episode runtime increases steadily from simple tasks through complex tasks without wall to complex tasks with wall. The additional runtime in scenarios containing walls reflects the greater number of collision checks and potential field updates required to navigate planar obstacles. This monotonic increase in computational load confirms that our categorization of simple and hard tasks is reasonable and that wall environments rightly belong in the hard category. To identify the precise contributors to these runtimes we divide each episode into agent runtime and environment preparation runtime. Agent runtime comprises action selection at every step and network updating every fifty steps and is dominated by reinforcement learning computations. Environment preparation updates the environment state around the agent and is driven primarily by artificial potential field calculations. For the DDPG-APF agent action selection requires around 0.5 ms network updating approximately 240 ms and environment preparation about 2 ms. For the TD3-APF agent action selection requires around 0.7 ms network updating approximately 280 ms and environment preparation about 2 ms. Across all analyzed metrics, DDPG-APF consistently outperforms both RL baselines and the classical A* method, validating the benefit of APF-guided reward shaping in continuous control settings. The APF-DDPG framework illustrates how combining structural guidance with adaptive learning can yield both stability and efficiency in robotic motion planning. Our experiments highlight clear benefits over classical approaches and pure reinforcement learning, but they also expose new compromises that arise from this hybrid design. In the following, we reflect on these strengths and weaknesses, emphasizing the trade-offs between guidance and exploration, efficiency and flexibility, as well as simulation and real-world applicability. Using a continuous state-action formulation provides clear benefits over discretized or inverse kinematics-based methods. Continuous incremental actions allowed the agent to adapt flexibly to dynamic obstacle configurations, enabling responsive trajectory adjustments. At the same time, the use of forward kinematics reduced computational complexity by avoiding costly inverse kinematics calculations, resulting in shorter episode runtimes compared to baseline methods. A drawback is that this formulation inherently enlarges the search domain, requiring careful reward shaping and parameter tuning to maintain stability; without strong guidance, training in such a continuous space may become less robust. The framework demonstrated improved learning efficiency, with faster convergence and more stable reward trajectories. The dense reward function provided continuous and informative feedback, facilitating effective optimization and reducing the number of collision trials. On the other hand, this reliance on a dense reward structure represents a limitation: if such shaping is not available, or must be inferred from noisy real-world signals, efficiency and stability may degrade substantially. Integrating a modified APF within the DDPG framework shaped the reward structure into a smoother potential field that continuously guided the agent toward the goal while repelling it from obstacles. This reduced fluctuations, stabilized learning, and biased exploration toward safer paths, improving sample efficiency and runtime compared to Pure-DDPG. At the same time, this integration introduced important trade-offs. The strong directional bias of APF may suppress the discovery of unconventional but potentially more efficient trajectories, and the perpendicular compensation term, while effective in our experiments, could in principle induce oscillatory motion in narrow passages. Moreover, the continuous computation of potential field forces and collision checks adds a runtime overhead of approximately 2 ms per step. Our experiments indicate that this cost is outweighed by fewer detours, reduced collisions, and faster convergence, yet its impact under noisy sensing or dynamic obstacles remains an open question. Together, these findings highlight both the stabilizing benefits and the potential risks of APF guidance. Our experiments revealed two concrete shortcomings of the current implementation. First, the framework lacks adaptability in adjusting step limits to different task complexities, which led to degraded performance in harder scenarios. Second, the method struggled with smooth transitions in multi-goal trajectory planning, limiting its effectiveness in sequential manipulation tasks. These weaknesses point to the need for adaptive step management and hierarchical or goal-decomposition strategies in future work. The dataset and evaluation design impose further limitations. All environments were manually created to challenge the robot arm, which restricts generalizability and prevents a definitive claim that APF-DDPG consistently outperforms other agents. The absence of standardized benchmarks for manipulator path planning further complicates comparisons across studies. Moreover, although oscillatory behavior was not observed in our fewer than 100 task instances, it cannot be excluded on larger or more uniform benchmarks. Parameter tuning across algorithms is another source of uncertainty: TD3, for example, did not achieve the expected improvements over DDPG, possibly because its double actor-critic structure was less advantageous in our task setup. These findings suggest that notions such as overestimation in reinforcement learning should be reconsidered in context, as optimistic Q-value estimates in DDPG may sometimes aid exploration rather than hinder it. The APF-DDPG framework is particularly suited to industrial scenarios where manipulators must operate safely in highly cluttered but largely static workcells. Examples include dynamic pick-and-place in warehouses or laboratory automation tasks, where dense static obstacles constrain free exploration and safety demands are high. In such settings, the spatial guidance of APF can provide reliable safety margins, while reinforcement learning enables adaptability to variations in object placement. In contrast, long-range navigation or human robot collaboration with highly dynamic obstacles may be less suited, since strong APF bias could hinder the discovery of unconventional solutions. Short- to mid-range manipulation in dense, safety-critical environments therefore appears to be the most promising niche for our approach. Finally, the transferability of our findings from simulation to real-world deployment remains a challenge. The modeling assumptions made in our setup are fragile under realistic conditions. For instance, the UR5 model neglects actuator dynamics, backlash, and latency, which influence fine-grained incremental control. Likewise, APF forces were computed under idealized, noise-free conditions, whereas real-world sensing inevitably introduces uncertainty in obstacle localization. The obstacles themselves were represented by simple geometric primitives, in contrast to the irregular and textured shapes encountered in practice. These simplifications suggest that APF-DDPG may be particularly sensitive to sensor noise and geometric mismatches. Future work should therefore investigate robustness under imperfect perception and more realistic robot models to narrow the sim-to-real gap. In this paper, we presented a reinforcement learning framework integrating a modified APF method with DDPG for robotic arm motion planning in continuous 3D environments. Our approach addresses the local optimum issues inherent in traditional APF methods through a tailored reward mechanism. Experimental results demonstrated that our method can find feasible solutions in fewer episodes and typically achieves more optimal paths compared to baseline models. Although APF introduces extra computation per step, in practice this overhead is outweighed by more efficient trajectories, leading to faster overall execution. The findings suggest that APF-based guidance not only accelerates convergence but also improves computational efficiency, making it suitable for complex obstacle avoidance tasks. However, the method’s current limitations, particularly the inability to automatically adjust step limits for varying complexities and difficulties in smoothly handling multi-goal trajectory planning, highlight important areas for further improvement. Future research should explore adaptive parameter tuning methods to enhance performance consistency across different task complexities. Additionally, developing more efficient algorithms for trajectory planning involving multiple sequential goals and extending the framework to support multi-agent cooperation and dynamic obstacle avoidance are promising directions. Ultimately, addressing these challenges will significantly narrow the gap between theoretical advances and practical robotic applications. The data and code that support the findings of this study are available at the following GitHub repository: https://github.com/JunyanPeng330/RobotPlanning. For further information or access to additional data, please contact the corresponding author. Fu, B. et al. An improved A* algorithm for the industrial robot path planning with high success rate and short length. Robot. Auton. Syst. 106, 26–37. https://doi.org/10.1016/j.robot.2018.04.007 (2018). Article Google Scholar Chen, G. et al. Path planning method with obstacle avoidance for manipulators in dynamic environment. Int. J. Adv. Robot. Syst. 15(6), 1729881418820223 (2018). Article Google Scholar Fusic, S. J., Ramkumar, P., & Hariharan, K. Path planning of robot using modified Dijkstra algorithm. In National Power Engineering Conference (NPEC), 1–5 (IEEE, 2018). Miao, H. & Tian, Y.-C. Dynamic robot path planning using an enhanced simulated annealing approach. Appl. Math. Comput. 222, 420–437 (2013). Google Scholar Xiuli, Yu., Dong, Mingshuai & Yin, Weimin. Time-optimal trajectory planning of manipulator with simultaneously searching the optimal path. Comput. Commun. 181, 446–453 (2022). Article Google Scholar Zhu, A. et al. Trajectory planning of rotor welding manipulator based on an improved particle swarm optimization algorithm. Adv. Comput. Signals Syst. 8(6), 122–129 (2024). Google Scholar Miao, C. et al. Path planning optimization of indoor mobile robot based on adaptive ant colony algorithm. Comput. Ind. Eng. 156, 107230 (2021). Article Google Scholar Hentout, A., Maoudj, A. & Aouache, M. A review of the literature on fuzzy-logic approaches for collision-free path planning of manipulator robots. Artif. Intell. Rev. 56(4), 3369–3444 (2023). Article Google Scholar Liu, W. et al. Path planning for mobile robots based on the improved DAPF-QRRT* strategy. Electronics 13(21), 4233 (2024). Article Google Scholar Khatib, O. Real-time obstacle avoidance for manipulators and mobile robots. In Proceedings. 1985 IEEE International Conference on Robotics and Automation, Vol. 2. 500–505 (1985). https://doi.org/10.1109/ROBOT.1985. Chen, Z., Ma, L., & Shao, Z. Path planning for obstacle avoidance of manipulators based on improved artificial potential field. In 2019 Chinese Automation Congress (CAC), 2991–2996 (2019). https://doi.org/10.1109/CAC48633.2019.8996467. Zhao, H. B. & Ren, Y. Improved robotic path planning based on artificial potential field method. Int. J. Adv. Robot. Syst. 37(02), 360–364 (2020). Google Scholar Wang, H., Wang, S. & Yu, T. Path planning of inspection robot based on improved ant colony algorithm. Appl. Sci. 14(20), 9511 (2024). Article CAS Google Scholar Li, Y. et al. A robot path planning method based on improved genetic algorithm and improved dynamic window approach. Sustainability 15(5), 4656 (2023). Article ADS Google Scholar Gao, S., Xu, Q. & Cao, J. Computing streamfunction and velocity potential in a limited domain of arbitrary shape. Adv. Atmos. Sci. 28(6), 1433–1444 (2011). Article Google Scholar Flacco, F. et al. A depth space approach to human–robot collision avoidance. In 2012 IEEE International Conference on Robotics and Automation, 338–345 (2012). https://doi.org/10.1109/ICRA.2012.6225245. Zhang, Y., Gong, P. & Hu, W. Mobile robots path planning based on improved artificial potential field. In 2022 6th International Conference on Wireless Communications and Applications (ICWCAPP), 41–45 (IEEE. 2022). Lillicrap, T. P. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971 (2015). Mnih, C. et al. Playing Atari with deep reinforcement learning. arXiv:1312.5602. https://api.semanticscholar.org/CorpusID:15238391 (2013). Rummery, G. & Niranjan, M. On-line Q-learning using connectionist systems. In Technical Report CUED/F-INFENG/TR 166 (1994). Chen, Y. et al. Integrated intelligent control of redundant degrees-of-freedom manipulators via the fusion of deep reinforcement learning and forward kinematics models. Machines 12(10), 667 (2024). Article CAS Google Scholar Imtiaz, M. et al. Trajectory planning and control of serially linked robotic arm for fruit picking using reinforcement learning. In 2023 International Conference on IT and Industrial Technologies (ICIT), 1–6 (IEEE, 2023). Sze, T. & Chhabra, R. Obstacle-free trajectory planning of an uncertain space manipulator: Learning from a fixed-based manipulator. In: American Control Conference (ACC), 3618–3624 (IEEE, 2024). Li, H., Gong, D. & Yu, J. An obstacles avoidance method for serial manipulator based on reinforcement learning and artificial potential field. Int. J. Intell. Robot. Appl. 5, 186–202 (2021). Article Google Scholar Andarge, E. W., Ordys, A. & Abebe, Y. M. Mobile robot navigation system using reinforcement learning with path planning algorithm. Acta Physi. Pol. A 30, 452–452 (2024). Article ADS Google Scholar Chang, L. et al. Reinforcement based mobile robot path planning with improved dynamic window approach in unknown environment. Auton. Robots 45, 51–76 (2021). Article Google Scholar Faust, A. et al. Prm-rl: Long-range robotic navigation tasks by combining reinforcement learning and sampling based planning’. In IEEE International Conference on Robotics and Automation (ICRA), 5113–5120 (IEEE, 2018). Li, P. et al. UGV navigation in complex environment: An approach integrating security detection and obstacle avoidance control. IEEE Trans. Intell. Veh. https://doi.org/10.1109/TIV.2024.3492539 (2024). Article Google Scholar Xu, X., Zhang, C. & Zhang, W. APF guided reinforcement learning for ship collision avoidance. In IET Conference Proceedings CP886, Vol. 2024. 12, 823–828 (IET, 2024). Parham, M. et al. Kinematic and dynamic modelling of UR5 manipulator. In IEEE International Conference on Systems, Man, and Cybernetics (SMC), 004229–004234 (IEEE, 2016). Lillicrap, T. P. et al. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971 (2015). Chen, L. et al. Deep reinforcement learning based trajectory planning under uncertain constraints. Front. Neurorobot. 16, 1662–5218. https://doi.org/10.3389/fnbot.2022.883562 (2022). Article Google Scholar Liu, K. A comprehensive review of bug algorithms in path planning. Appl. Comput. Eng. 33, 259–265 (2024). Article Google Scholar Kusuma, M. et al. Humanoid robot path planning and rerouting using A-Star search algorithm. In 2019 IEEE International Conference on Signals and Systems (ICSigSys), 110–115 (IEEE, 2019). Fujimoto, S., Hoof, H., & Meger, D. Addressing function approximation error in actor-critic methods. In International Conference on Machine Learning, 1587–1596 (PMLR, 2018). Download references Open Access funding enabled and organized by Projekt DEAL. Linda-Sophie Schneider and Junyan Peng contributed equally to this work. Pattern Recognition Lab, Friedrich-Alexander-University Erlangen-Nuremberg, Erlangen, Germany Linda-Sophie Schneider, Junyan Peng & Andreas Maier Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar All authors conceived the experiments, L. S. and J. P. conducted the experiments and analysed the results. J. P. designed the dataset. All authors reviewed the manuscript. Correspondence to Linda-Sophie Schneider. The authors declare no competing interests. Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Reprints and permissions Schneider, LS., Peng, J. & Maier, A. Robot movement planning for obstacle avoidance using reinforcement learning. Sci Rep 15, 32506 (2025). https://doi.org/10.1038/s41598-025-17740-5 Download citation Received: 12 May 2025 Accepted: 26 August 2025 Published: 12 September 2025 Version of record: 12 September 2025 DOI: https://doi.org/10.1038/s41598-025-17740-5 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Advertisement Scientific Reports (Sci Rep) ISSN 2045-2322 (online) © 2026 Springer Nature Limited Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

Neuracore, a British robot learning platform to scale and deploy faster, has closed a €2.5 million ($3 million) pre-Seed funding round to fuel team expansion, accelerate product development, and support Neuracore’s growth. The round was led by Earlybird Venture Capital, with participation from Clem Delangue (Co-founder & CEO at Hugging Face) and advisors spanning academia, hardware, and AI. “After years in academic and industrial robotics, I saw every team – from research labs to warehouse automation startups – rebuilding the same infrastructure from scratch,” says Stephen James, founder and CEO. “Our mission is to eliminate that duplication and democratise access to high-performance robot learning tools. With this funding and our free academic programme, we’re enabling both researchers and companies to focus on advancing robotics itself, not on building the pipelines to support it.” Neuracore’s pre-Seed round sits within a steadily active 2025 European robotics and physical-AI landscape, where several startups have secured funding across adjacent segments. In Sweden, Stockholm-based Rerun raised €15.6 million to advance multimodal data infrastructure for robotics and autonomous systems. Switzerland saw two sizeable rounds: Zürich’s Flexion secured €43 million to develop reinforcement-learning “brains” for humanoid robots, while mimic closed €13.8 million to expand its dexterous robotic-hand technology. Germany’s Energy Robotics raised €11.5 million for autonomous inspection software for industrial sites, while Italy’s Adaptronics attracted €3.15 million to scale its adaptive gripper technology. Together, these 2025 rounds represent approximately €86 million in disclosed capital across companies building data infrastructure, robotic intelligence layers, and advanced manipulation hardware. Within this context, Neuracore’s focus on robot-learning infrastructure aligns with the increasing investor interest in foundational robotics software rather than solely application-layer systems. Although no UK-based robotics infrastructure company appears alongside these examples, the cluster of European funding rounds indicates a wider continental shift toward platforms that standardise data handling, training, and deployment for robotics teams. Laura Waldenstrom, Principal at Earlybird Venture Capital, adds: “The robotics industry is at an inflection point, moving from the ROS 1.0 era to a data-first paradigm powered by deep learning. Teams are still wasting months building and maintaining their own infrastructure instead of focusing on deployment. Neuracore provides what AWS did for web applications: a reliable, scalable platform that just works. We’re thrilled to support Stephen’s vision to become the infrastructure layer for the coming wave of intelligent robotics.” Founded in 2024 by Stephen James, Assistant Professor of Robot Learning at Imperial College London, Neuracore is building the core infrastructure powering the next generation of intelligent robots. Its platform enables robotics teams to go from data collection to deployed machine learning models in “days not months“, eliminating the bottlenecks that currently consume up to 80% of engineering time. Neuracore explains that their software stack replaces fragmented ‘Frankenstein’ robotics setups with a unified, cloud-based system that handles asynchronous data collection, visualisation, training, and deployment. By integrating the full robot learning pipeline into one platform, Neuracore helps teams focus on innovation rather than infrastructure. The platform is being used by over 50 organisations across commercial and academic robotics, including partnerships with hardware manufacturers. Alongside the funding, Neuracore is launching a free academic programme. This programme gives universities and research institutions worldwide unrestricted access to the full enterprise platform, or in other words, the same infrastructure used by its commercial customers. “Academic researchers are building the foundation for tomorrow’s robots,” James adds. “They shouldn’t waste months setting up data pipelines – they should be innovating. We want Neuracore to be the backbone that lets them do that.” As per data provided by the company – by 2030, the AI-powered robotics sector is projected to exceed €43 billion ($50 billion), driven by advances in deep learning and real-world robot deployment. Yet, Neuracore says most robotics teams remain limited by infrastructure constraints, often spending months integrating disparate systems for data logging, training, and deployment. Neuracore’s platform aims to solve this by providing purpose-built infrastructure optimised for robotics workflows – from asynchronous data streaming and ML-native storage to one-click deployment and continuous improvement loops. With its new academic programme, Neuracore aims to close the accessibility gap between research and industry. Universities and robotics labs will now have free, unlimited access to the full platform, enabling faster experimentation, collaboration, and reproducibility across institutions. EU-Startups.com is the leading online magazine about startups in Europe. Learn more about us and our advertising options. © Menlo Media S.L. - All rights reserved.

This thesis introduces a learning-based framework to improve the efficiency and scalability of Task and Motion Planning (TAMP) for robotic manipulation tasks in cluttered and complex environments. Traditional TAMP methods face significant challenges due to the high computational cost of repeated geometric feasibility checks and the combinatorial explosion of the search space, especially in multi-robot settings. To address these issues, two deep learning-based models are proposed: AGFPNet, which predicts the feasibility of manipulation actions to reduce reliance on geometric planners, and GRN, which leverages graph neural networks and scene graphs to provide interpretable and generalizable feasibility predictions, including the causes of action failure. These models are further extended to support collaborative multi-robot scenarios and complex, mesh-shaped objects. Building on these models, the thesis proposes a multi-robot, feasibility-informed TAMP algorithm that integrates learned geometric reasoning into a search-based planner. This integration enables faster exploration of the search space by prioritizing promising actions and avoiding infeasible ones. Additionally, a novel informed backtracking mechanism uses predicted infeasibility causes to refine planning constraints and heuristics. Extensive experiments demonstrate that the proposed approach significantly accelerates planning, improves scalability, and maintains generalizability, even in highly cluttered or large-scale environments. Cette thèse propose une approche basée sur l'apprentissage automatique pour améliorer l’efficacité et la scalabilité de la planification de tâches et de mouvements (Task and Motion Planning, TAMP) dans des environnements complexes et encombrés. Les approches TAMP traditionnelles sont confrontées à des défis importants en raison du coût computationnel élevé des vérifications répétées de faisabilité géométrique, ainsi que de l’explosion combinatoire de l’espace de recherche, notamment dans des contextes multi-robots. Pour surmonter ces limitations, deux modèles d’apprentissage profond sont introduits : AGFPNet, un réseau de neurones permettant de prédire la faisabilité des actions de manipulation afin de limiter les appels coûteux aux planificateurs géométriques ; et GRN, un réseau de neurones exploitant des représentations de scènes 3D sous forme de graphes pour fournir des prédictions inter! prétables et généralisables, y compris les causes d’infaisabilité des actions. Ces modèles sont également étendus aux scénarios multi-robots et à la manipulation d’objets de formes complexes (maillés). À partir de ces modèles, la thèse présente un algorithme TAMP multi-robots guidé par les prédictions de faisabilité, qui intègre le raisonnement géométrique par apprentissage au sein d’un planificateur classique. Cette intégration permet d’explorer l’espace de recherche plus efficacement en priorisant les actions prometteuses et en évitant les actions infaisables. Un nouveau mécanisme de backtracking informé est également proposé, exploitant les causes d’infaisabilité pour définir des contraintes de planification et affiner les heuristiques de coût vers l’objectif. Des expérimentations approfondies démontrent que cette approche permet de résoudre des tâches de manipulation complexes avec des horizons de planification longs, dans des environnements fortement encombrés, tout en réduisant considérablement les temps de planification. Connectez-vous pour contacter le contributeur https://laas.hal.science/tel-05330516 Soumis le : lundi 4 août 2025-14:38:22 Dernière modification le : lundi 24 novembre 2025-14:47:05 Contact Ressources Informations Questions juridiques Portails CCSD

Aktuelles Biodiversität langfristig beobachten Frohe Festtage und ein glückliches Neues Jahr „Steht für eure Überzeugungen ein!“ „Alle Forschenden stehen vor philosophischen Fragen“ Forschen mit der Zivilgesellschaft Wichtiger Brennstofftest für FRM II erfolgreich no events today.

Elon Musk has shared an exciting update about Tesla’s humanoid robot, Optimus. A video shared by Tesla shows the Optimus robot learning new tasks by watching humans do them, just like we do when watching tutorial videos. For those wondering why this matters, it’s a big change from the old way of programming robots. It brings Optimus closer to acting like a human trainee. By learning from everyday actions, the robot could become far more useful and learn new things much faster. Musk says Tesla wants Optimus to study real-world videos, like those on YouTube, and then use what it sees to perform similar tasks on its own. Engineering lead at Tesla, Milan Kovac, says this breakthrough has already helped the robot pick up jobs like vacuuming, sorting items, stirring food, or taking out the rubbish. This new learning method means Optimus no longer needs detailed programming for every single task. Instead, it can copy from human examples and be told what to do with simple voice or text commands. Musk believes this could give Optimus “task extensibility,” meaning it could learn almost anything very quickly if shown the right videos. So far, Tesla has shown Optimus doing tasks in short clips — from cleaning and putting things together to even dancing. However, earlier demos faced criticism, as some people believed many of these actions were actually being controlled by humans rather than done by the robot itself.Despite the doubts, Musk believes this way of learning from videos is a key turning point. He has called Optimus “the biggest product of all time” and even thinks it could become more valuable than Tesla’s car business in the future. He expects the robot to be made on a large scale, with possible use in Tesla factories as soon as late 2025, or maybe earlier if progress continues quickly.What could this mean going forward? Instead of coding every single move, developers might simply have Optimus watch a set of instructions — whether it’s a cooking video or a guide to folding clothes — and then simply copy it. In theory, the robot could learn just like a person: by watching, imitating, and improving.Still, many experts point out that Tesla’s current robots are less skilled with fine hand movements and walking compared to leaders like Boston Dynamics. While the idea is promising, critics say there’s still a long way to go before we see a truly general-purpose robot that works well in homes or workplaces.- EndsPublished By: Aman rashidPublished On: Aug 6, 2025Also read | Nostalgia max: Prince of Persia remake is coming and it is a big deal for gamers in IndiaAlso read | Xiaomi 15 Ultra review: A chonky camera kingAlso read | Vivo V50e review: No thrills, but plenty of value Despite the doubts, Musk believes this way of learning from videos is a key turning point. He has called Optimus “the biggest product of all time” and even thinks it could become more valuable than Tesla’s car business in the future. He expects the robot to be made on a large scale, with possible use in Tesla factories as soon as late 2025, or maybe earlier if progress continues quickly. What could this mean going forward? Instead of coding every single move, developers might simply have Optimus watch a set of instructions — whether it’s a cooking video or a guide to folding clothes — and then simply copy it. In theory, the robot could learn just like a person: by watching, imitating, and improving. Still, many experts point out that Tesla’s current robots are less skilled with fine hand movements and walking compared to leaders like Boston Dynamics. While the idea is promising, critics say there’s still a long way to go before we see a truly general-purpose robot that works well in homes or workplaces.- EndsPublished By: Aman rashidPublished On: Aug 6, 2025Also read | Nostalgia max: Prince of Persia remake is coming and it is a big deal for gamers in IndiaAlso read | Xiaomi 15 Ultra review: A chonky camera kingAlso read | Vivo V50e review: No thrills, but plenty of value

Add the following snippet to your HTML:<iframe frameborder='0' height='385' scrolling='no' src='https://www.hackster.io/HiwonderRobot/turbopi-a-rpi-based-ros-2-robot-kit-for-ai-vision-learning-ec86c1/embed' width='350'></iframe> Skip the setup, start building. TurboPi is the ready-to-run Raspberry Pi 5 & ROS 2 robot kit that gets you straight to AI vision & voice pro Read up about this project on Skip the setup, start building. TurboPi is the ready-to-run Raspberry Pi 5 & ROS 2 robot kit that gets you straight to AI vision & voice pro Getting started with robotics and AI can feel daunting. Between complex setup, fragmented software, and the fear of hardware becoming a money pit, many beginners hit a wall before they even write their first line of code. The TurboPi kit is designed to break down these barriers, offering a pre-configured, all-in-one platform that gets you from unboxing to a running AI robot in minutes. Built around the Raspberry Pi 5 and the ROS 2 (Robot Operating System 2) framework, it provides a solid foundation for learning modern robotics development. Here’s how it structures the learning journey. The biggest hurdle is often the initial setup. TurboPi tackles this by providing a complete, pre-configured system image. All necessary drivers, dependencies, and demo programs are installed. Simply flash the image to an SD card, boot up, and within minutes you can have the robot performing face detection, gesture control, or object tracking. This immediate, tangible success provides the crucial first feedback loop to keep motivation high. 💡Get all codes, videos, and diagrams from TurboPi Robot Tutorials. Instead of being a shallow jack-of-all-trades, TurboPi concentrates on AI vision and voice interaction—areas that deliver visually impressive and intuitive results quickly. Its vision stack is ready to use out of the box, integrating tools like: With these, you can immediately build projects like an object-following robot, a simple autonomous driving simulator, or an FPV (First-Person View) exploration rover. Beyond vision, TurboPi includes voice interaction capabilities. Using its onboard audio hardware and integrated multimodal models, it can understand natural language commands like "find a red ball, " moving beyond simple remote control to more natural interaction. A frustrating hardware experience can kill a project. TurboPi is designed for robustness and ease of use: This combination allows the platform to execute complex maneuvers reliably, letting you focus on software and AI rather than hardware debugging. Self-directed learning needs a map. TurboPi is supported by extensive documentation, step-by-step TurboPi tutorials, and open-source code examples. The platform leverages the mainstream Raspberry Pi and ROS 2 ecosystems, ensuring the skills you learn are transferable. You're not just buying a robot; you're gaining a guided entry point into a vast world of open-source robotics development. If you are a student, hobbyist, or developer looking for a gentle yet powerful on-ramp to embedded AI, computer vision, and ROS 2, TurboPi is a compelling choice. It removes the initial friction of system integration, letting you spend your time where it matters: building, coding, and experimenting with intelligent robotic behaviors. Hackster.io, an Avnet Community © 2025

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Advertisement Nature Communications volume 16, Article number: 11038 (2025) Cite this article 1690 Accesses 1 Altmetric Metrics details In contrast to the diverse landing solutions that the animals have in nature, human-made aircraft struggle when it comes to perching on vertical walls. However, traditional dynamic simulations and experiments lack the high efficiency required to analyze the perching and design the robot. This paper develops an efficient machine learning framework to predict vertical-wall perching success for flying robots with spines, overcoming traditional methods’ inefficiency. A validated knowledge-based model computes the robot’s transient dynamics during high-speed perching, identifying key success factors. By training the mixed sample data, a data-driven model has been proposed to predict the success or failure of an arbitrary perching event. Here, we show that this high-precision prediction optimizes robot control and structural parameters, ensuring stable perching while drastically reducing the time and cost of conventional design approaches, advancing the flying robot capabilities. Perching is an innate ability of animals, which means they can land and stay on a branch or another suitable surface. For example, flies can land upside down on a ceiling1, gliding geckos crash into tree trunks with their tails to stabilize landing2, and bats land on the tops of caves to rest3. Meanwhile, creatures have evolved a variety of methods for landing, perching, locomotion, and take-off to adapt to various irregular natural and artificial surfaces4. In contrast to animals, the perching approach of robots (especially aircraft) for surfaces is tougher, a flat ground is necessary. However, having these perching abilities similar to animals can help expand the use of the robots in extreme environments and improve their mobility, robustness, and working time5,6. For instance, in search-and-rescue operations, perching mechanisms would allow drones to land on walls or power lines, enabling persistent environmental monitoring from vantage points while conserving power7. Consequently, achieving reliable perching capability on non-horizontal surfaces has emerged as a significant research frontier in aerial robotics. The optimization challenges pertaining to pitch angle, velocity, and lift force during avian perching8 are equally encountered in developing perching technologies for aerial robots. Zufferey and Tormo-Barbero9 presented a process to autonomously land an ornithopter on a branch. This method explains the combined operation of a pitch-yaw-altitude flapping flight controller, an optical close-range correction system, and a bistable claw appendage design that can grasp a branch within 25 milliseconds and reopen. Guo, Tang et al.10 developed an actuated modular landing gear system achieving stable perching and resting of UAVs on diverse structures. Roderick, Cutkosky, and Lentink11 designed a biomimetic robot achieving dynamic perching on complex surfaces with concurrent grasping of irregular objects, demonstrating that closed-loop balance control critically expands viable perching parameter ranges. Hsiao, Bai et al.12 presented a flying robot achieving perching on diverse substrates – including wet/dry overhangs and walls – via an ultralight attachment module constituting merely 3.3% of total mass. Liu, Tian et al.13 proposed a contact-adaptable robot for perching featuring all-in-one electroadhesive pads, enabling ceiling surface attachment and significant onboard energy conservation. Liu, Dong et al.14 developed a multifunctional aerial manipulation system using a composite cup structure (soft inner cup + rigid outer shell), enabling execution of perching or lateral aerial grasping tasks while reducing reliance on precise control through the soft cup’s adaptability to multicopter-induced angular errors. While the aforementioned studies have made substantial progress in developing and characterizing novel perching mechanisms, no research to date has established predictive frameworks for perching. Nevertheless, perching—which involves rapid maneuvers and is subject to strict velocity constraints—remains one of the most critical challenges for flying robots. This is largely because the inability to predict unsuccessful landings on vertical surfaces is highly likely to result in severe damage to the robot15. Establishing a predictive framework for perching would therefore greatly enhance the success rate of wall-perching maneuvers and reduce the losses associated with landing failures. To predict the wall-perching effect and design the structure configuration of the flying robot with spines, people have recourse to traditional dynamic simulation16 and experimental method14 at present. However, traditional dynamic simulations lack the high efficiency required to handle the substantial computational workload. Furthermore, the limited simulation results are difficult to form the practical control strategies for perching. Similarly, the observation of a limited amount of experimental data from perching experiments is insufficient for the design of flying robots. Fortunately, machine learning, as a part of Artificial Intelligence (AI) technology, offers a promising solution for providing robots with perching experience and improving their capabilities. This potential will be realized through advanced ML implementations, as demonstrated by recent breakthroughs in flight control domains. The integration of machine learning (ML) into flight systems to enhance autonomous capabilities has reached a sophisticated level of maturity such as Kaufmann, Bauersfeld et al.17 presented a champion-level autonomous racing system employing deep reinforcement learning fusing simulated training with physical vehicle telemetry. Ouahouah et al.18 proposed an obstacle avoidance algorithm for UAVs based on probability and Deep Reinforcement Learning (DRL). Recently, ML-driven perching control studies for horizontal surfaces or non-horizontal surfaces are constituting a rapidly evolving frontier. Waldock, Greatwood et al.19 implemented a DRL framework integrating a nonlinear constraint optimizer (IPOPT) with a DQN agent to compute perching trajectories for horizontal surfaces. This approach derives optimal pre-contact pitch/velocity parameters that guarantee structurally stable perching on horizontal surfaces. de Croon, De Wagter et al.20 proposed a machine learning method enabling flying robots to acquire optical-flow-based strategies for complex tasks such as landing, resulting in faster, smoother landings. Ladosz et al.21 employed visual sensing and DRL, incorporating a Deep Regularized Q-learning algorithm and a custom-designed reward scheme to enable autonomous landing of UAV systems on moving platforms. Habas and Cheng15 replicated fly landing behaviors in micro-quadcopters using a generalizable control policy for arbitrary ceiling-approach conditions. Their framework employs reinforcement learning in simulation to optimize discrete sensorimotor pairs across diverse approach velocities/directions, with the first-stage “Flip-Trigger” control implementing a one-class support vector machine (SVM). Despite these advances, no existing ML addresses the challenge of predicting wall-perching outcome while co-designing robot structures − a critical capability for designing flying robotics in complex environments. In this work, to bridge this gap, a machine learning-based framework for the wall-perching prediction and design of the flying robot with spines is developed in this paper. The framework synthesizes the machine learning, computational impact dynamics and experiment. Firstly, an accurate knowledge-based model, validated for accuracy through experimentation, has been developed to calculate the transient responses of the flying robot during perching. The key factors that influence the success rate of perching at relatively low landing speeds are identified. Then, a flying robot prototype is constructed to conduct the perching experiment. In addition, the mixed sample datasets are obtained, comprising experimental data from perching experiments and simulation data computed using the knowledge-based model. By training the mixed sample datasets, a data-driven model is established to predict the success or failure of an arbitrary perching event. The results of this study can aid in designing flying robots and preventing perching failure. To achieve the landing and climbing functions, a spiny flying and wall-climbing robot is designed based on the multimodal robotic design philosophy5. The whole robot is composed of three subsystems, namely the flight system, landing system, and attachment system (Fig. 1a). Here, the flight system mainly consists of a micro quadcopter, which can provide the robot with flying capability. The landing system consists of two flexible tail structures connected by a transverse beam with two wheels (Fig. 1a). During landing, the tail will make initial contact with the vertical wall. The huge impact force is reduced thanks to the large structure compliance of the tail we designed. The attachment system is designed with two carbon fiber rods with spines (Supplementary Figs. 1, 2). The spine morphology exhibits structural similarity to beetle tarsal claws15. Due to the very tiny tip of the spine, it can crawl stably on the wall. a The structure composition of the robot including attachment, flight and landing systems. b Robot flies approaching wall, where v0 and θ0 are the initial horizontal incidence velocity and initial pitch of flying robot, respectively. c The robot contacts the wall by its tail, where αB is the robot’s body acceleration. d The robot executes an upward flip action, where vB is the velocity of the mass center of the robot. e The robot grasps and perches the wall. The robot achieves perching through four designed stages (Fig. 1b-e). In the first stage, the robot approaches the vertical wall using its flight system. Then, the second stage is a contact-impact phase, where the robot uses its flexible tail rods to impact the wall. Subsequently, it enters the third stage (i.e., rotation stage). The robot’s wheel contacts the wall, and then the entire robot body rotates upwards around the wheel. In the fourth stage, the robot uses spines in its attachment system to climb on the wall, and the whole perching process is completed. Although the attachment system (Fig. 1a) can help the robot climb on the wall, this paper focuses on the landing and perching problems of the robot. For detailed descriptions of the spine structure and wall-climbing mechanism, please refer to the Supplementary Note 1. To analyze the landing and perching behavior of the robot, humans resort to a contact-impact dynamic model naturally. This model is a typical knowledge-driven model based on the principles of impact dynamics of a flexible multibody system. In this study, the model is discretized using the finite element method (FEM) (Fig. 2a), and calculations are performed in the LS-DYNA software. We use this model to evaluate the success or failure of perching in different maneuvering states before landing. The numerical results guide the experiments and are used to build the subsequent data-driven model, which will be discussed the next section. a Discrete model based on FEM, which is used to solve the contact-impact problem during the landing event. b Displacement of the spine under three sets of initial landing conditions with lift strengthening (I indicates θ0=−5° and v0=0.4 m/s; II indicates θ0=−5° and v0=0.8 m/s; III indicates θ0=−10° and v0=0.8 m/s). The solid and short dash lines are the displacements in the x and z direction, respectively. The symbol star represents the occurrence of collision between the spine and wall. c The lift curves during the landing event for receding cases. d The lift curves during the landing event for strengthening cases. e The resultant velocity of the spine during the landing process for 4 typical landing behaviors. The success and failure envelop regions (i.e., the green and red areas) that we summarized for the landing events under 114 sets of initial landing conditions are also shown. The three values in legend in this figure are pitch of flying robot θ0,initial horizontal incidence velocity v0 and lift FL, respectively (where LR means lift receding and LS means lift strengthening). f The robot undergoes a fixed-axis rotation where the rotation center is the contact point of the tail during the “perfect landing” process. g The robot body flips upward using the wheels as the rotation center during the “perfect landing” process. h The wheel separates from the wall and then contacts it again during the “normal landing” process. i The maximum von-Mises stress appears at the root of the landing rod when the robot contacts the wall. j The maximum von-Mises stress appears in the connecting piece. If the high velocity causes excessive stress, the structure will be damaged. k Configuration of robot when it is in the success perching state on the wall. The numerical simulations for 114 cases of maneuver states before landing have been performed. The alterable parameters of maneuver states include initial pitch of the flying robot θ0, the initial horizontal incidence velocity v0 and the lift FL. Figure 2b illustrates the displacement of the spine after the tail contacts the wall, which was calculated based on the knowledge-driven model. The peaks of the dashed curves represent the quick release of deformation energy from the tail. In Fig. 2b, when the positive direction of the X-axis rotates counterclockwise to align with the robot’s axis, θ0 is defined as positive. Conversely, when the rotation is in the opposite direction, θ0 is considered negative. The other attitude angles and flying velocities are all equal to zero. It should be noted that the lift FL is an active force, it will be controlled during the contact-impact process. It is set as an increasing or decreasing linear function of time t (Fig. 2c, d) to analyze the effect of FL on the success of landing. The gravity is a constant value during the whole landing process. By our computation for 114 cases based on the knowledge-driven model, it can be found that there are 4 types of typical landing behaviors (Fig. 2e). And regardless of what the initial conditions are, the resultant velocities of spine always experience multiple oscillations since the tail occurs obvious vibration under contact-impact. Figure 2e also shows the envelop regions that are summarized from 114 landing events corresponding to the 114 data points in section “Data-driven model for perching”. The green area means the robot landed successfully while red and purple areas mean robot failed to land (LR means lift receding and LS means lift strengthening). Meanwhile, some simultaneous robot postures and stress states during the landing are also given. From Fig. 2e the 4 types of typical landing behaviors can be summarized as follows: Perfect landing (success). The characteristic feature of such landing behavior is that the robot’s mass center trajectory approximates a semicircular path, with negligible body sway after attachment. The corresponding mass center velocity is indicated by the black curve in Fig. 2e. To get these results, the robot must have the most ideal landing state before contacting the wall. i.e., the robot almost flies horizontally, (−8° < θ0 < −5°) and the sum of velocity of the mass center and the pitch are mid-value (0.6 m/s <θ0 <0.8 m/s). Then, the tail contacts the surface (Fig. 2f) and it occurs bending deformation. Then, the robot performs a fixed-axis rotation where the rotation center is the contact point of the tail. The tail does not keep in contact with the wall until the wheels contact with the wall. Then, the robot body flips upwards (Fig. 2g) using the wheels as the rotation center. Finally, the robot’s spines grasp the surface and attach successfully. Normal landing (success). The characteristic feature of such landing behavior is that the robot’s mass center trajectory approximates a horseshoe-shaped path, with significant body sway after attachment. The corresponding mass center velocity is indicated by the blue curve in Fig. 2e. To achieve this landing process, the robot requires the initial states of −5° < θ0 < − 2° for the angle and 0.8 m/s < v0<1.0 m/s for the velocity before contacting the wall. Subsequently, the robot performs a plane motion to achieve upward flips. However, the wheel undergoes a transition between contact and separation with the wall (Fig. 2h) during the entire upward motion. Finally, the wheel will contact the wall again at the end of the landing process. In other words, the wheel makes contact twice during the landing process. Low-kinetic energy landing (failure). This typical behavior and landing process is shown as the red curve in Fig. 2e. The initial conditions for this landing behavior are v0<0.4 m/s and θ0>−4°. Since the initial kinetic energy is low, the robot cannot perform an upward flip motion, despite the fact that the lift of the quadrotor aircraft is sufficient and its tail does not rebound from the wall. The robot will execute an upward straight motion parallel to the wall surface. Rebound landing (failure). This landing behavior has two forms of rebound landing. The first is when the robot achieves an upward flip motion, but the spine rebounds when it contacts the wall due to the large contact force at the spine. This typical behavior and landing process are shown by the purple curve in Fig. 2e. In this case, the initial conditions are generally v0>1.0 m/s and θ0<−10°. The failure is caused by the spine’s inability to grasp the wall tightly. The second form is when the robot directly rebounds as the tail contacts the wall, preventing it from achieving an upward flip motion. In this case, the initial conditions are generally v0>1.0 m/s and θ0>−2°. The occurrence probability of this form is very low because θ0 is generally less than −2° during the actual flight of the robot under artificial operation. Besides, Fig. 2i–j shows the stress concentration phenomenon in the structure of the landing system. The critical areas where structural failure can occur are primarily located in the root zone of the tail. Overall, the conditions such as initial velocity, initial pitch, and lift of the robot have very complicated effects on landing behavior and the landing process. They sometimes work together to achieve successful landings (Fig. 2k), but other times they conflict with each other and lead to landing failures. It is more intuitive to understand that if the velocity is slower and the absolute value of the pitch angle is smaller, lift assistance is required. On the other hand, if the velocity is faster and the absolute value of the pitch angle is greater, increasing lift may cause the robot to rebound from the wall surface. However, these relationships are not absolute. Therefore, one of the greatest challenges in landing on a vertical wall is predicting success or failure. Through our calculations, it has been determined that the time required for each computation using the knowledge-driven model is substantial. Moreover, it is impractical to exhaustively explore all working conditions and control parameter combinations using this method. On a computer equipped with an Intel CPU 6700 K and DDR4 16GB RAM, the cost of each landing process generally requires around 48 h. Thus, there is an urgent need for an efficient and accurate calculation model to enable numerical evaluation of all landing conditions. Fortunately, the emergence of machine learning methods has made this achievable. A comprehensive discussion on this topic will be presented in the section titled “Data-driven model for perching”. In the experiment of landing on the wall, the robot firstly takes off from a horizontal surface and then is operated randomly to land on the wall, with different initial pitch of flying robot θ0,initial horizontal incidence velocity v0 and lift FL. The video of the experiments were recorded to obtain experimental data (Supplementary Figs. 3, 4). For more details about the process of experiment, refer to Supplementary Note 2. The parameters, including θ0, v0 and FL, were intentionally varied before colliding with the wall to provide a diverse range of experimental results for machine learning in subsequent chapters. This experimental data serves two main purposes: firstly, it can be used to validate the accuracy of the computational results obtained from the knowledge-driven model. Additionally, it can be utilized to identify general patterns in the landing behavior by observing the robot’s actual landing event. Figure 3 shows the experimental images of 4 typical landing behaviors. The video of the continuous movement of the flying robot can be viewed in the Supplementary Movie 1. The movement trajectories for perfect landing (success), normal landing (success), low-kinetic energy landing (failure), and rebound landing (failure) are depicted in Fig. 3a–g, respectively. It has been observed that the detailed characteristics captured by the experiment are consistent with those summarized by the numerical evaluation. Figure 3h–k represents the velocities of the mass center and the pitches of the robot during the perfect landing and rebound landing processes, respectively. During the perfect landing, the pitches change linearly to a constant value, while in the rebound landing, the pitches change non-linearly from negative to positive. a The motion track for the perfect landing (success). b The tail contacts the surface during the perfect landing (success). c The wheels contact with the wall during the perfect landing (success). d The robot’s spines grasp the surface and attach during the perfect landing (success). e The motion track for the normal landing (success). f The motion track for the low-kinetic energy landing (failure). g The motion track for the rebound landing (failure). h The velocities of the robot’s mass center during the perfect landing (success). The shadow area indicates the robot is transitioning from a tail collision to a spine collision. i The robot’s pitches during the perfect landing process. j The velocities of the robot’s mass center during the rebound landing (failure). k The pitches of robot during the rebound landing (failure). The control parameters have a great influence on the landing responses. Due to its high cost and time consumption, merely a limited number of landing experiments can be conducted, and the experimental data under a limited set of control parameters can be recorded. It is hard to conduct experiments covering all working conditions. Nevertheless, the limited experimental results are still valuable, which can be used as sample data for establishing the data-driven model in the next section. Generally, “learning for simulation” is a traditional idea in AI fields. Recently, some scientists have hypothesized that “simulation for learning” may be one trend in computer graphic science22. Here, based on mixed samples including computational data obtained from a knowledge-driven model and experimental data, we have established a data-driven model to predict the landing of a robot. In this work, we propose a technique that combines simulations based on knowledge-driven models and machine learning based on data-driven models, in order to leverage the best aspects of both. The entire training and prediction process is shown in Fig. 4, and includes the following 5 steps: a Simulation for robot landing based on the knowledge-driven model. b Actual experiment for robot landing. c Extreme state when the robot is out of control. d Data points obtained from simulations based on the knowledge-driven model and actual experiments under different lift forces. e Data-driven model of landing behavior is established after training using machine learning. f If the data-driven model predicts a failure in landing, the robot will abort the landing, fly away from the wall, and prepare for another landing attempt. g The robot perform a new landing. When it approaches the landing surface again, predictions are conducted once more through the data-driven model. h The robot lands on the wall if the data-driven model predicts a successful landing. Step 1: Preliminary structural design of the robot. The preliminary design of the robot and its landing strategy are based on Newton mechanics theory. Step 2: Achievement of data sets. An accurate three-dimensional finite element model (i.e., knowledge-driven model) is established to simulate the landing of the robot (Fig. 4a). The computation work of the contact-impact dynamic responses is performed in the software LS-DYNA. Through the computation for 114 cases of maneuvering states, the landing results are obtained. The actual landing experiments with the robot prototypes have also been conducted (Fig. 4b). To define the boundaries of our dataset (Fig. 4c), we identify data points corresponding to invalid states, including out-of-control events. This dataset is composed of the robot’s pitch degree, velocity, lift data, and success/failure outcomes extracted from both computational and experimental results (Fig. 4d and Supplementary Data 1). Step 3: Establishment of the data-driven model. To ensure the machine learning framework achieves effective landing predictions for different lift modes, the dataset is divided into two parts based on lift receding and lift strengthening. Each part contains two features: initial horizontal velocity and initial angle. The dataset is then distributed within a three-dimensional feature space composed of initial horizontal velocity, initial angle, and lift mode (Fig. 4e), representing landing outcomes under various initial conditions. Additionally, there is no noise in the dataset, hence the noise reduction is not required. Step 4: Training. After training with the machine learning models, a decision boundary for Multi-layer Perceptron (MLP) and Random Forest (RF) (i.e. hyperplane for Support Vector Machine) is generated in the landing feature space under different lift conditions (Fig. 4e). The data points representing successful landings are separated from those representing failed landings by the decision boundary. The farther a data point is from the boundary within the decision boundary, the higher the probability of a successful prediction by the model, and the greater the confidence of the robot in a successful landing. By adjusting the model parameters that affect the construction of the decision boundary, we can obtain the optimal model with the highest confidence in its predictions. Step 5: Prediction. According to the training results, when the robot is in a landing situation, the success or failure of landing will be predicted by the data-driven model, and the subsequent behavior of the robot will be determined. In the future actual application of perching, if the prediction result computed by onboard computer indicates failure, the two motors near the wall accelerate their rotational velocity to cause the robot to decelerate, then the robot will adjust its flight states in real-time to fly away from the wall (Fig. 4f) and seek a new landing opportunity (Fig. 4g). Conversely, if the prediction result indicates success, the two motors located away from the wall accelerate their rotational velocity to make the robot flip upwards, and the robot will directly land on the wall (Fig. 4h). The entire computational process is following: After the primary selection of data collected from the robot’s sensors, certain influential factors that affect the success or failure of landing can be determined manually. These influential factors can be employed as dimensions within the dataset. These influences can be interconnected with each other, and there is no limit to the number of dimensions. Both simulated and experimental data samples are utilized to construct datasets. Generally, a larger dataset will yield more accurate prediction results. However, if the selected range of the sample data can encompass all various scenarios and its distribution in the state space is more uniform, even a smaller dataset can still achieve desirable outcomes. The proportion of simulated and experimental data can be adjusted after estimating the time and cost considerations. Generally, simulated data is less expensive and more efficient. When the robot encounters a new scenario, we can instantly predict the outcome, allowing the robot to respond accordingly based on the prediction result. Ultimately, this enables the robot to intelligently handle various new challenges related to perching. This paper employed the Python language to develop the machine learning program. In this study, the datasets were obtained under proper surface conditions, which are sufficient for the robot to land easily. For soft surfaces with low rugosity, the proper condition is when the surface Brinell hardness is less than 2. Many tree bark surfaces can satisfy this condition. For hard wall surfaces, a high level of rugosity is required. Our tests indicate that the surface static coefficient of friction μ approximately needs to be greater than 0.2, and there should be at least a 20% difference between static and dynamic friction. After preparing the dataset, our goal is to ensure that the machine learning method achieves high prediction accuracy. Therefore, we need to optimize the data-driven model. During the model training process, we use the same dataset and continuously adjust the parameters of the generated model in order to identify the best parameter configuration (for SVM and RF) that yields high prediction accuracy. In this way, we are able to find the optimal model. Specifically, the landing success region and failure region are distinguished by the decision boundary. The accuracy of the decision boundary is determined by the ratio of the correctly classified data to the total dataset. Be different from SVM and RF, MLP do not need to identify the best parameter configuration. Figure 5 presents wall-perching results calculated using data-driven models based on MLP, SVM, and RF methods, respectively. Each model predicts outcomes across three dimensions: lift force, pitch angle, and initial velocity. The lift curves illustrated in Fig. 2c, d remain employed here, the hybrid dataset (Fig. 5a, b) used for training combines simulation results from our knowledge-driven model with experimental data. Figure 5c–h displays prediction results from MLP, SVM and RF, respectively. These results demonstrate that regardless of the machine learning method, the decision boundary envelope encompasses a larger successful perching region when lift generation is present, indicating that providing lift enhances landing success rates for flying robots. Comparative analysis reveals considerable differences among the three methods’ predictions, with MLP demonstrating relatively superior performance over SVM and RF. Detailed descriptions of the data-driven models and corresponding machine learning methods are provided in Supplementary Note 3. a Mixed dataset under lift receding obtained using a knowledge-driven model and experiment. b Mixed dataset under lift strengthening obtained using a knowledge-driven model and experiment. c The decision boundary obtained by the MLP model under lift receding. d The decision boundary obtained by the MLP model under lift strengthening. e The decision boundary obtained by the SVM model under lift receding. f The decision boundary obtained by the SVM model under lift strengthening. g The decision boundary obtained by the RF model under lift receding. h The decision boundary obtained by the RF model under lift strengthening. The machine learning-based framework can not only predict the landing, but also guide the design of the flying robot. In the initial design phase, some basic premise need to be considered for achieve the successful landing and perching on the wall, including the following points: 1) The mass of robot should be as light as possible; 2) The structure where the robot collide with the wall needs to have a certain buffering capacity to reduce the impact load; 3) The landing structure is beneficial for the robot to flip up from the horizontal state to the vertical state after collision, so as to achieve successful grasping. Therefore, in view of these three requirements, the overall structure of the robot was preliminarily designed according to the existing experience and knowledge. Compared with other structures, the landing rod is particularly critical for the successful landing of robots. In this paper, the landing rod structure was preliminarily designed as a flexible carbon fiber rod structure. Compared with other forms of buffer structure, its biggest advantage is that it has the lightest weight under the same size, and its structural form is the simplest, which is very conducive to the assembly of the structure. On the basis of the initial design, there are still many parameters to be determined, including the length of the landing rod la, the angle α between the landing rod and the horizontal structure of the robot, etc. These parameters directly determine the success or failure of landing. The length of the landing rod la determines the bending flexibility of the structure and the cushioning performance of the structure; The included angle α determines whether the flying robot can flip up at a reasonable speed. Therefore, the detailed closed-loop optimization design of these key parameters will be carried out below. The process of closed-loop optimization design is shown in Supplementary Fig. 5. During the detailed closed-loop optimization design phase, we focused on optimizing the landing rod structure of the robot while keeping other components unchanged. Building upon the initial design, we developed multiple landing rod configurations with varied parameter combinations. The la was incrementally increased from 2 cm to 4 cm in 1 cm steps, while the α was expanded from 20° to 60° in 5° increments, generating numerous parameter permutations. To identify the optimal parameter combination, we employed an iterative approach. This involved comparing machine learning-predicted successful landing zones across different parameter sets until the resulting hyperplane region achieved maximum coverage in the two-dimensional feature space defined by the robot’s initial horizontal velocity and initial angle—thus accomplishing closed-loop structural optimization. Each iteration integrated FEM simulation analysis, physical landing experiments and machine learning to solve the hyperplane zone in its parameter combination. Through this process, we determined the robot’s optimal structural configuration and geometric dimensions (Supplementary Fig. 6 and Supplementary Table 1). Additionally, this study optimized the lift conditions during the robot’s landing phase. Additionally, for the robot designed in this study, the preceding section determined the optimal set of initial horizontal velocities and initial angles for landing by comparing successful landing regions under two distinct control strategies: lift receding and lift strengthening. Subsequently, based on the data-driven model’s predictions of landing success or failure during actual landings, different control strategies can be implemented for the robot. Optimized control is achieved by adjusting the robot’s flight parameters. Although the prediction framework proposed in this paper has only been validated on the flying robot in this design, the closed-loop design analysis mentioned above for the flying robot is also applicable to other wall-perching robots of different structures. That is, through experiments or simulations, obtain a dataset of the relationship between the influencing factors such as speed, posture, and distance of arbitrary robot during its wall-perching and the success of the perching. Then, use the machine learning method employed in this paper to predict the perching success rate under different initial conditions, and further optimize the design of the robot based on the prediction results. We first discuss the efficiency and accuracy of the three machine learning approaches. A comparative analysis revealed that the MLP training time was only about 5 s, whereas SVM, due to the need to search for optimal parameters, had the longest training time of approximately 180 s—36 times that of MLP. Additionally, by examining the decision boundary results predicted by the three methods, it can be observed that compared to MLP’s predictions, SVM and Random Forest exhibited larger errors. Their success regions contained several characteristic points, which corresponded to failure outcomes in both numerical calculations and experimental tests. Conversely, such contradictions were almost absent in the MLP results. Therefore, for the problem addressed in this paper, the MLP method demonstrates relatively higher prediction accuracy. It is widely acknowledged that traditional methods such as the finite element method or the material point method often require a large amount of computational time to simulate a single landing event. This is primarily due to the nature of the landing event, which presents a typical contact-impact problem. To accurately model this problem, a very small time step is necessary for integrating the high-dimensional dynamic differential equations. Additionally, the iterative computations required to address the contact nonlinearity further contribute to the time-consuming nature of the simulations. Simulating all possible scenarios with arbitrary initial angles and velocities of a robot using traditional methods is clearly impractical. Similarly, conducting experiments to test every case is also unfeasible. The experimental process is not only time-consuming but also expensive. It involves extensive preliminary preparation and equipment debugging, and there is also the risk of potential damage to the robot during experiments. In conclusion, whether through experimentation or traditional numerical simulation, simulating the entire range of coverage states is challenging. The machine learning-based framework employed in this paper allows for predicting the complete space, encompassing all possible cases, using a reduced number of data samples, thereby enhancing the efficiency of landing research. Secondly, the primary goal of AI is generally to achieve “learning to simulation”. This concept entails training models using data, enabling them to perform simulations more efficiently without complex manual modeling and algorithm design. In this paper, we also employ AI to build a data-driven model. By utilizing this model, we can predict the success or failure of a robot’s landing and determine its subsequent behavior. This methodology, known as “learning to simulation”, is widely applicable in diverse domains such as computer graphics, robotics, and physics modeling. However, this paper introduces a concept called “simulation to learning”. Unlike the previous work, the data in this study is not solely derived from experiments but also includes simulation data computed by a knowledge-driven model. The combination of this simulation data and experimental data results in mixed data. The main objective of “simulation to learning” is to obtain mixed data. Subsequently, the mixed data is utilized to establish a data-driven model. Ultimately, this data-driven model enables efficient and accurate simulation of all landing events. Our research validates the effectiveness of the “simulation-to-learning” approach. Finally, we discuss the potential issues in practical applications. The current study demonstrates successful perching of the robotic platform under low-speed flight conditions. However, extending this capability to high-speed operations presents several challenges: sensor delays in gyroscopes and accelerometers may cause attitude control lag during landing, while limited motor response could destabilize rapid flight-to-perch transitions. Accelerometer noise during impact may also interfere with contact detection, though filtering algorithms could mitigate this effect. Additionally, the control strategy must address motor overheating during prolonged stall conditions through optimized power-thermal management. These challenges highlight important directions for future research in high-speed robotic perching. This paper focuses on proposing a machine learning-based prediction framework for wall-perching, where the landing experiments aim to collect samples of successful and failed outcomes. Consequently, our experiments solely examine the process from flight state to perched landing, without yet utilizing prediction results to: assess landing success/failure prior to wall contact, guide the robot’s decision on whether to initiate landing, or direct it to steer away from the wall to reorient both pitch angle and initial approach velocity for subsequent perching attempts. Implementing this comprehensive closed-loop process—using our predictive framework as the decision-making core—remains future work, which will establish complete control strategies for aerial robotic wall-perching and lay the groundwork for developing feedback regulation algorithms. In addition, it must be pointed out that the experimental data we used for prediction were measured under relatively ideal laboratory conditions and are not fully applicable to experimental conditions with high interference or smoother walls. Under high interference conditions (such as when the wind is strong), due to the small size of the robot, it will have a significant impact on the stability of flight, thereby affecting the results of landing prediction. However, the grasping structure used by the robot in this study cannot be successfully attached to smooth walls. Therefore, the prediction model proposed in this study is not applicable to the experimental conditions of smooth walls. The UG software was employed to create a detailed 3D model of the robot (Fig. 1). The robot involves three key systems: flight system, landing system, and attachment system. Ultimately, the Crazyflie 2.1 quadrotor flight platform by Bitcraze was selected. The landing system is composed of two carbon fiber rods, two rubber balls, and two wheels. We have carefully adjusted the thickness and angle of the carbon fiber rods to optimize the energy absorption and conversion capabilities during the landing process. The core components of the attachment system are two spines, which consist of micro-scale curvy metal structures. These spines are connected to curved carbon fiber rods, flexible ropes, a horizontal bar, and a T-shaped bar using revolute joints and holes. The connector parts are 3D printed using resin material. In the flight system, in/out servos and extend/retract servos are employed to control the movement of the two spines. The geometric entities are created within the LS-DYNA software (Fig. 2a). The entity consists of two parts: a vertical wall and a robot. The robot itself is divided into six components: quadcopter, servos, carbon fiber frame, connector, spines, and ropes. Firstly, to capture the transient landing deformation, all parts of the robot are modeled as deformable bodies in terms of the material model. The quadcopter, ropes, and servos are simulated using the 003 PLASTIC isotropic material model to represent plastic behavior. The carbon fiber parts utilize the 054 ENHANCED orthotropic material model to simulate the properties of carbon fiber23. On the other hand, the connectors and spine parts are modeled using the 003 PLASTIC isotropic material model to represent resin material. Lastly, the wall is modeled using the 020 RIGID material model to simulate a rigid body. To facilitate the development of the robot’s dynamics equation, we employ the finite element technique to discretize the structure’s deformation field and create a mesh for the flexible parts. The resulting discrete model in LS-DYNA is depicted in Fig. 2a. Prior to meshing, the parts are divided into simpler shapes, and techniques such as sweeping and mapping are utilized to generate high-quality meshes. Additionally, based on knowledge-driven modeling, a combination of regular and hexagonal meshes is employed to ensure accurate results (Fig. 2a). The total number of elements in the model amounts to 58232. Secondly, in the aspect of constraints, all degrees of freedom of the wall are fully constrained, and an automatic surface-to-surface contact algorithm utilizing a penalty function is employed between the robot and the wall. With consideration for computational convenience, neighboring parts are rigidly connected as constraints. Thirdly, in terms of the loads, the lift of the robot is shown in Fig. 2c and d. The application of the lift exists in two situations: the first is that the lift is always the same as the gravity of the robot. The second is that the lift is the same as the gravity before contacting the surface to simulate normal flight, and then the lift increases to twice the gravity after contact with the wall. The direction of the robot’s initial velocity points towards the wall, ranging from 0 m/s to 1.5 m/s, while the robot’s initial pitch ranges from 5° to −15°. The prestress of the flexible rope has little effect on the collision process, so it is ignored in the simulation calculation. Finally, in terms of numerical integration in calculation, the LS-DYNA central difference method is employed to analyze nonlinear explicit dynamics. The size of time step is determined based on the mesh size and set to be Δt = 7.27 × 10−5s. By utilizing the LS-DYNA software solver, the system’s dynamic response is obtained, which includes the collision contact force, the dynamic stress cloud diagram of the structure, the stress curve for each point, the motion process, and other kinematic parameters. In summary, the dimensions and material properties of the finite element model used in the numerical simulation closely match those of the actual robot. The constraint conditions between components and loading scenarios align with experimental configurations. Meanwhile, the meshing strategy and mesh density account for grid sensitivity in computational results. Furthermore, for the impact dynamics problem of the landing/perching process, the simulation incorporates contact constraints between the robot and the wall while employing explicit dynamics algorithms. This approach ensures computational accuracy alongside enhanced efficiency. Collective implementation of these measures guarantees strong agreement between numerical simulation and experimental results. To further validate the consistency between simulations and experimental results, the landing outcomes of simulation and experiment are compared for selected data points, as presented in Supplementary Table 2. For the establishment of the data-driven model in this paper, three distinct machine learning approaches were employed: Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Random Forest (RF). MLP is a type of feedforward artificial neural network consisting of an input layer, at least one hidden layer, and an output layer. In an MLP, each layer comprises multiple units. Each unit incorporates a differentiable linear function and an activation function, mapping the input vector of the layer to a scalar output unit. This paper pertains to a binary classification problem with two features: initial velocity and initial angle. The combinations of initial velocity and angle yield a complex non-linear decision boundary (Supplementary Fig. 7). Since MLPs can learn complex interactive relationships between features and thereby fit decision boundaries of intricate shapes, it exhibits high adaptability to the research focus of this paper. Consequently, an MLP model composed of 5 fully connected layers was constructed: Input Layer contains 2 neurons, corresponding to the two features— initial velocity and initial angle. Hidden Layers contain 4 layers with 64, 128, 64, and 32 neurons respectively. The ReLU activation function was applied to these layers to introduce non-linear fitting capability. The Output Layer employs the sigmoid activation function to map the final output to a value representing the probability of a successful landing. After training for 150 epochs, the model has achieved satisfactory predictive performance in determining whether the flying robot successfully lands on the wall for given combinations of initial velocity and angle. SVM is a supervised learning algorithm mainly used for classification and regression analysis. Its core concept involves finding an optimal hyperplane in a high-dimensional feature space to separate data points of different classes while maximizing the minimum distance between the hyperplane and samples of both classes (i.e., maximizing the margin). For linearly inseparable data, SVM maps original features to a higher-dimensional space via the Kernel Trick to achieve linear separability. In the binary classification problem studied in this paper, the combination of initial velocity and initial angle forms a complex nonlinear decision boundary (Supplementary Fig. 8). The SVM method used in this paper employs the RBF kernel, which utilizes Gaussian functions to project data into an implicit high-dimensional space. This approach effectively captures complex feature interactions, enabling flexible fitting of intricate nonlinear decision boundaries. The specific implementation process is as follows: Use grid search combined with leave-one-out cross-validation to find optimal parameters c (regularization parameter) and γ (scaling coefficient of RBF kernel) (Supplementary Table 3). Here, c controls the trade-off between classification errors and margin width—smaller c enhances generalization capability while larger c may cause overfitting; determines the influence range of a single sample—larger values yield more complex decision boundaries. The optimal parameters are determined by maximizing cross-validation accuracy rates, subsequently training the final model. RF is an ensemble learning method based on decision trees, comprising multiple independent decision trees. When performing classification in this paper using a random forest, an input sample flows to each decision tree, and these trees independently predict the sample’s outcome. RF uses the majority vote of classification results from multiple decision trees as the final classification result. This research uses the Gini Index as the feature selection criterion. A smaller Gini Index indicates higher sample purity; therefore, node splitting progresses in the direction of decreasing Gini Index. To achieve better classification performance, hyperparameters such as the maximum depth of decision trees, the maximum number of features considered for splitting a node, and the minimum sample count at leaf nodes and internal nodes are configured. As the number of decision trees critically impacts random forest training results, Leave-One-Out Cross-Validation is employed to determine the optimal number of trees (Supplementary Table 3). After setting these hyperparameters, the original data is input into the random forest for training, yielding the model employed in this paper (Supplementary Fig. 9). The dataset of robot landing results used to train the machine learning prediction framework has been deposited in the figshare database at https://doi.org/10.6084/m9.figshare.2963699624. Source data are provided with this paper. The source code is publicly available at https://doi.org/10.5281/zenodo.1747388725. Liu, P., Sane, S. P., Mongeau, J.-M. J., Zhao, J. & Cheng, B. Flies land upside down on a ceiling using rapid visually mediated rotational maneuvers. Sci. Adv. 5, eaax1877 (2019). Article ADS PubMed PubMed Central Google Scholar Siddall, R., Byrnes, G., Full, R. J. & Jusufi, A. Tails stabilize landing of gliding geckos crashing head-first into tree trunks. Commun. Biol. 4, 1020 (2021). Article CAS PubMed PubMed Central Google Scholar Riskin, D. K. et al. Bats go head-under-heels: the biomechanics of landing on a ceiling. J. Exp. Biol. 212, 945–953 (2009). Article PubMed Google Scholar Roderick, W. R. T., Cutkosky, M. R. & Lentink, D. Touchdown to take-off: at the interface of flight and surface locomotion. Interface Focus 7, 20160094 (2017). Article PubMed PubMed Central Google Scholar Pope, M. T. et al. A Multimodal Robot for Perching and Climbing on Vertical Outdoor Surfaces. IEEE Trans. Robot. 33, 38–48 (2017). Article Google Scholar Kovač, M., Germann, J., Hürzeler, C., Siegwart, R. Y. & Floreano, D. A perching mechanism for micro aerial vehicles. J. Micro-Nano Mech. 5, 77–91 (2009). Article Google Scholar Floreano, D. & Wood, R. J. Science, technology and the future of small autonomous drones. Nature 521, 460–466 (2015). Article ADS CAS PubMed Google Scholar KleinHeerenbrink, M., France, L. A., Brighton, C. H. & Taylor, G. K. Optimization of avian perching manoeuvres. Nature 607, 91–96 (2022). Article ADS CAS PubMed PubMed Central Google Scholar Zufferey, R. et al. How ornithopters can perch autonomously on a branch. Nat Commun 13, 7713 (2022). Article ADS CAS PubMed PubMed Central Google Scholar Hang, K. et al. Perching and resting—a paradigm for UAV maneuvering with modularized landing gears. Sci. Robot. 4, eaau6637 (2019). Article PubMed Google Scholar Roderick, W., Cutkosky, M. & Lentink, D. Bird-inspired dynamic grasping and perching in arboreal environments. Sci. Robot. 6, eabj7562 (2021). Article CAS PubMed Google Scholar Hsiao, Y. H. et al. Energy efficient perching and takeoff of a miniature rotorcraft. Commun Eng 2, 38 (2023). Article PubMed Central Google Scholar Liu, H. et al. Electrically active smart adhesive for a perching-and-takeoff robot. Sci. Adv. 9, eadj3133 (2023). Article ADS PubMed PubMed Central Google Scholar Liu, S., Dong, W., Ma, Z. & Sheng, X. Adaptive aerial grasping and perching with dual elasticity combined suction cup. IEEE robot. autom. lett. 5, 4766–4773 (2020). Article Google Scholar Habas, B. & Cheng, B. From flies to robots: inverted landing in small quadcopters with dynamic perching. IEEE Trans. Robot. 41, 1773–1790 (2025). Article Google Scholar Dunlop, D. J. & Minor, M. A. Modeling and simulation of perching with a quadrotor aerial robot with passive bio-inspired legs and feet. Letters Dyn. Sys. Control 1, 021005 (2021). Google Scholar Kaufmann, E. et al. Champion-level drone racing using deep reinforcement learning. Nature 620, 982–987 (2023). Article ADS CAS PubMed PubMed Central Google Scholar Ouahouah, S., Bagaa, M., Prados-Garzon, J. & Taleb, T. Deep reinforcement-learning-based collision avoidance in UAV environment. IEEE Internet Things J 9, 4015–4030 (2022). Article Google Scholar Waldock, A., Greatwood, C., Salama, F. & Richardson, T. Learning to perform a perched landing on the ground using deep reinforcement learning. J. Intell. Robot. Syst. 92, 685–704 (2018). Article Google Scholar de Croon, G. C. H. E., De Wagter, C. & Seidl, T. Enhancing optical-flow-based control by learning visual appearance cues for flying robots. Nat. Mach. Intell. 3, 33–41 (2021). Article Google Scholar Ladosz, P., Mammadov, M., Shin, H., Shin, W. & Oh, H. Autonomous landing on a moving platform using vision-based deep reinforcement learning. IEEE robot. autom. lett.9, 4575–4582 (2024). Deist, T. M. et al. Simulation assisted machine learning. Bioinformatics 35, 4072–4080 (2019). Article CAS PubMed PubMed Central Google Scholar Reuter, C., Sauerland, K.-H. & Troester, T. Experimental and numerical crushing analysis of circular CFRP tubes under axial impact loading. Compos. Struct. 174, 33–44 (2017). Article Google Scholar Shen, Y. et al. Dataset for landing prediction. figshare https://doi.org/10.6084/m9.figshare.29636996 (2025). Article Google Scholar Shen, Y. et al. Machine learning-based framework for wall-perching prediction of flying robot. Zenodo https://doi.org/10.5281/zenodo.17473887 (2025). Article Google Scholar Download references The authors gratefully acknowledge the support from Robotics and Intelligent Machine Lab founded by Yunian Shen, NJUST. This study was sponsored by the Natural Science Foundation of Jiangsu Province grant (BK20221484 to Y.S.) and the Aeronautical Science Foundation of China (2024M058059002 to Y.S.). These authors contributed equally: Yunian Shen, Chenxi Mao. Department of Mechanics and Engineering Science, School of Physics, Nanjing University of Science and Technology, Nanjing, P. R. China Yunian Shen, Chenxi Mao, Zeyu Qi, Kunpeng Liu, Weixu Zhang & An Cao Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Y.S. and C.M. contributed to the conceptualization and methodology of the study. Y.S., C.M., Z.Q., and K.L. conducted the investigation. Y.S. provided supervision throughout the research process. Y.S. and C.M. were responsible for writing the original draft of the paper. Y.S., C.M., W.Z., and A.C. contributed to the review and editing of the manuscript. All authors contributed to the writing of the paper. Correspondence to Yunian Shen. The authors declare no competing interests. Nature Communications thanks Zhongjin Ju, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Reprints and permissions Shen, Y., Mao, C., Qi, Z. et al. Machine learning-based framework for wall-perching prediction of flying robot. Nat Commun 16, 11038 (2025). https://doi.org/10.1038/s41467-025-67386-0 Download citation Received: 14 January 2025 Accepted: 28 November 2025 Published: 11 December 2025 Version of record: 11 December 2025 DOI: https://doi.org/10.1038/s41467-025-67386-0 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Advertisement Nature Communications (Nat Commun) ISSN 2041-1723 (online) © 2026 Springer Nature Limited Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

打開 App 聯合報系 著作權所有 ©2024 太棒了！你沒有錯過任何訊息！ 本文共1118字 美國人型機器人公司Agility Robotics宣布，在C輪融資中獲得輝達（NVIDIA）投資；能率集團旗下能率亞洲、佳能等先前已共同投資Agility Robotics新台幣3億元；Agility創辦人將於24日訪台，首站行程就安排拜訪佳能，雙方有望合作開發量產機器人視覺模組。 Agility Robotics發布新聞稿宣布，與NVIDIA的關係不僅僅是技術合作，也在C輪融資中獲得NVentures（Nvidia的風險投資部門）的投資；AgilityRobotics將繼續與NVIDIA合作開發下一代AI加速機器人技術。 外傳NVIDIA投資Agility Robotics的金額約2000萬至3000萬美元，但並未獲得證實。 Agility Robotics是一家雙足Mobile ManipulationRobot（MMR）公司，創造首款全球商業化應用的人形機器人產品Digit，客戶包含物流巨頭GXO Logistics和電商龍頭Amazon等。 能率集團積極布局機器人領域，旗下能率亞洲先前已宣布，連同佳能等其他關係企業，簽約投資美國機器人公司Agility Robotics約1000萬美元（約新台幣3億元）；除了投資外，佳能也積極爭取Agility Robotics機器人視覺模組的合作開發與生產製造。 此外，Agility Robotics創辦人赫斯特（JonathanHurst）將於24日來台參訪，首發站就安排在佳能企業，預料雙方將針對機器人視覺模組領域討論。 2025年被視為人形機器人元年，能率集團先前宣布，旗下包括能率創新、佳能等，共同攜手投資美國機器人公司Mantis Robotics，佳能也接獲MantisRobotics的感測器訂單，初期約300套，預計第4季出貨，能率創新有望代工組裝機器人手臂。 佳能進一步指出，機器人訂單能見度直達2026年下半年，且人形機器人一台至少8顆鏡頭，視覺模組產品單價（ASP）更高，預估新產品如機器人視覺與無人機相關營收占比，2026年將達5%到10%，2027年挑戰20%，為未來營運添上強勁成長引擎。 Agility Robotics強調，他們目標是製造可在動態環境中與人類一起移動、工作和適應的機器人，多年來NVIDIA幫助Agility Robotics在雙足人形機器人中，實現突破的AI和自主性，透過NVIDIA AI基礎設施和開發框架，可訓練強大的模型，即時處理感測器資料。 Agility Robotics指出，隨著機器人技術進入下一階段，在倉庫、工廠、家庭、醫院、酒店、辦公室和零售店進行大規模部署，NVIDIA的技術平台可提供運算效能和靈活性。 ※ 歡迎用「轉貼」或「分享」的方式轉傳文章連結；未經授權，請勿複製轉貼文章內容 產業熱點 市場焦點 市場焦點 產業熱點 市場焦點 國際專欄 外媒解析 深度報導 科技產業 金融脈動 此為訂戶獨享，每月 10 則，轉傳連結即可邀請朋友免費閱讀全文。 本站使用元件之台股行情、個股基本資料及財務資訊為 凱衛資訊 提供。元件資料來源：台灣證券交易所, 櫃買中心。 本站元件所提供之金融資訊， 係供參考，不能做為投資交易之依據，所有資料以台灣證券交易所、櫃買中心公告為準。 聯合報系 著作權所有 ©2024 前往頁面

Amazon said last week that its Innovation Fund is ramping up its investment in robotics. Complete the form to unlock this article and enjoy unlimited free access to all PYMNTS content — no additional logins required. yesSubscribe to our daily newsletter, PYMNTS Today. By completing this form, you agree to receive marketing communications from PYMNTS and to the sharing of your information with our sponsor, if applicable, in accordance with our Privacy Policy and Terms and Conditions. Δ And as a report Monday (March 4) by Bloomberg News notes, the company already has robots working for it, with a humanoid device called Digit moving bins in one of its warehouses. The Bloomberg report focuses on Digit’s maker, a company called Agility Robotics, saying it has “down-to-earth” ambitions at a time when other robotics startups paint their technology as making science-fiction-worthy accomplishments. Agility, on the other hand, plans to build 10,000 robots a year and for warehouses and storerooms worldwide, with the company taking part in an investment boom that has drawn in $1.6 billion in venture capital in the last five years, the report said. “The humanoid robot is as close to a given as I could imagine from a technology,” Adrian Stoch, chief automation officer at GXO Logistics, told Bloomberg. His company has tested Digit at one of its warehouses, and Stoch praised the robot’s potential flexibility, envisioning bots unloading trucks during the night shift so boxes are ready for human employees on the day shift, while the machines take on other duties. Advertisement: Scroll to Continue The report came the same day that Agility announced the hiring of a new CEO, Microsoft and Qualcomm veteran Peggy Johnson. “In a field cluttered with ‘demo-ware’ and hype, Agility stands apart for having resolutely, steadily, and remarkably made a human-centric robot that actually works — and in demanding customer environments,” Johnson said in a news release covering her appointment. Meanwhile, PYMNTS wrote last week about the way advances in artificial intelligence (AI) are leading to robots with better features and more effective human interaction capabilities. For example, Figure AI has reportedly raised $675 million to develop AI-powered humanoid robots, a project backed by investors like Jeff Bezos’ Explore Investments, as well as technology firms such as Microsoft, Amazon, Nvidia, OpenAI and Intel. “AI can better enable robots to better understand their environments, allowing them to better detect objects and people they come across,” Sarah Sebo, an assistant professor of computer science at the University of Chicago, where she directs the Human-Robot Interaction (HRI) Lab, told PYMNTS in an interview. “AI can allow robots to generate language to more fluidly communicate with people and respond more intelligently to human speech. Finally, AI can help robots to adapt and learn to perform tasks better over time by receiving feedback, making it more likely that the robot behaves in ways that receive positive feedback and less likely to behave in ways that receive negative feedback.” Trust Wallet Integrates Revolut to Facilitate Crypto Purchases in Europe Terraform Labs Co-Founder Do Kwon Sentenced to 15 Years in Prison Costco’s Digital Sales Surge 21% as Members Maintain Spending Appeals Court Says Judge Must Consider Allowing Apple to Collect Commission We’re always on the lookout for opportunities to partner with innovators and disruptors.

Gold prices trade mixed in last session of 2025 Bitcoin price today: climbs to $89k but set for yearly decline, Q4 losses weigh Oil prices set for biggest annual slump since 2020 as supply glut weighs Asia stocks set for stellar yearly gains as AI boom outweighs trade jitters Within the last year, a major milestone was crossed in robotics agility. From Unitree Robotics to Boston Dynamics, humanoid robots have now outpaced many sci-fi movies featuring fictional CGI robotics. Combined with AI agents that give life to agile robotic platforms, fiction is materializing in real time. Citi Global Insights analysts forecast the humanoid robotics industry to scale up to $7 trillion by 2050. Elon Musk is even more optimistic, having noted mid-2024 that Optimus could eventually churn $10 trillion revenue. As with EVs, it is likely that China will have the bulk share in this growth. After all, China is already ahead of the game, having outpaced both Japan and Germany in industrial robot density, according to the International Federation of Robotics (IFR) report. But outside of Tesla (NASDAQ:TSLA) which firms should investors consider for the AI and robotics boom? In 2021, Hyundai Motor Co (OTC:HYMTF) acquired Boston Dynamics from Japanese SoftBank (TYO:9984) for $880 million. To this day, Boston Dynamics remains the envy of the world when it comes to cutting edge humanoid robots. In late March, Hyundai announced a $21 billion investment into the US between 2025 and 2028, boosting domestic supply chains and car manufacturing. Last week, the Trump admin amended the President’s previous executive order (EO) on 25% vehicle tariffs, enabling up to 15% credit if the cars are US-assembled. In short, Hyundai is positioning to be a major player in robotics and EVs domestically, from supplying vehicles to Waymo to scaling up Boston Dynamics’ envelope-pushing work. Year-to-date, Hyundai stock is down 1.2%, currently priced at $51 per share. According to WSJ, analyst ratings for Hyundai shares are overwhelmingly in the “buy” zone with 20 analysts, while zero analysts are recommending to sell. After merging with Vitesco in October 2024, German Schaeffler AG (ETR:SHA0) has become the leading provider of components in the advanced field of motion technology. Investors may have already seen this application with humanoid robot Digit, built by privately held Agility Robotics. In early March, Agility Robotics and Schaeffler Group announced a strategic partnership, enabling investors to have exposure to Agility Robotics in turn. So far, Agility Robotics seems to be at the forefront of deploying Mobile Manipulation Robots (MMRs) at scale. GXO Logistics, with over 1,000 warehouse sites, is the first to deploy Digit as a part of the emerging Robots-as-a-Service (RaaS) business model. According to MarketsandMarkets, the RaaS sector is heading for a CAGR expansion of 17.4% between 2023 and 2028. Operating alongside human workers, Digits will aid in repetitive tasks with more expansion plans. In addition to Digit, GXO Logistics is also looking into Apptronik’s Apollo humanoid robot, but that company is also privately held. In April’s pre-close call for Q1 2025 earnings, scheduled for May 7th, Schaeffler AG reported €5.9 million in sales in Q4 ‘24, slightly lower than €6.1 million in Q1 ‘24. The company expects strong sales growth in e-mobility (electric drives, controls, mechatronics) and vehicle lifetime solutions (repair & maintenance). Even robots are made by robots, more specifically by industrial robots. This is where Swiss-Swedish ABB (ST:ABB) Group comes in, holding 13.5% market share in the industrial robot market. The company is always expanding its list of partnerships, even sponsoring MassRobotics as the US-centric robotics innovation hub, annually supported by AMD (NASDAQ:AMD), Amazon (NASDAQ:AMZN) and other large companies. Most recently in late April, the company announced its plan to split its ABB Robotics division as of Q2 2026, listing it as a separate company. During 2024, ABB Robotics contributed 7% to ABB Group’s revenue, delivering $2.3 billion. If the plan goes ahead with shareholder approval, they will receive “ABB Robotics” shares as dividend in-kind, proportional to their existing stock stake. In the meantime, ABB’s robots are continuing to be deployed across a wide range of businesses. The fast food sector seems a particularly good match, in order to avoid arbitrary minimum-wage increases. BurgerBots is the most recent such restaurant, opened in Los Gatos, California. If BurgerBots takes off, it is likely that global fast food chains will catch up with their much deeper pockets, and ABB Group will be in the middle of it. Year-to-date, ABB stock is up 2.8% to current price of $55 per share. Per WSJ, the average ABB price target is $58.79, while the top is $78.78 per share. If the ABB Robotics split goes through, shareholders could see a substantial split premium not accounted for in the forecasting. *** Neither the author, Tim Fries, nor this website, The Tokenist, provide financial advice. Please consult our website policy prior to making financial decisions. This article was originally published on The Tokenist. Check out The Tokenist’s free newsletter, Five Minute Finance, for weekly analysis of the biggest trends in finance and technology. Join our investing challenges and compete for rewards while you learn!

打開 App 聯合報系 著作權所有 ©2024 太棒了！你沒有錯過任何訊息！ 本文共1253字 美國AI機器人廠Agility Robotic於11）日宣布在C輪融資中獲得輝達（NVIDIA）投資，而國內能率集團（5392）旗下佳能（2374）與能率亞洲先前已投資Agility Robotics約1,000萬美元（約新台幣3億元），未來能率集團有望結盟輝達、Agility Robotics揮軍AI機器人市場。 據悉，Agility Robotics成立於2015年，總部位於美國俄勒岡州奧爾巴尼，該公司前身為俄勒岡州立大學的實驗室，該實驗室以開發雙足機器人成果而著名，並在2022年獲得亞馬遜旗下創投青睞，並在今年接連獲得能率集團、輝達入股。 Agility Robotics主要產品為雙足步行機器人，目前共有Cassie與Digit兩個機型，同領域主要競爭對手為特斯拉旗下機器人公司Optimus；業界人士指出，有別於其他新創機器人公司，Agility Robotics所開發的產品最接近黃仁勳曾提及的AI人形機器人，預料未來在AI巨頭協力下，該公司有望進一步擴大AI人形機器人市場版圖。 對此，Agility Robotics透過新聞稿證實，公司與輝達的關係不僅僅是技術合作，也在C輪融資中獲得NVentures（NVIDIA旗下風險投資部門）的投資；AgilityRobotics將繼續與輝達合作開發下一代AI加速機器人技術。 針對輝達投資金融，儘管Agility Robotics並未在新聞稿中透露，但外界推估金額很可能落在2,000至3,000萬美元，不過尚未獲得雙方證實；Agility Robotics目前的客戶包含物流巨頭GXO Logistics和電商龍頭亞馬遜。 Agility Robotics強調，公司的目標是製造可在動態環境中與人類一起移動、工作的AI人形機器人，而輝達多年以來都在協助公司完成雙足人形機器人的開發，未來透過輝達AI基礎建設與開發框架，公司的技術也將進入下一個階段，有望將AI人形機器人大規模部署在工廠、家庭、醫院、飯店、辦公室等場域。 至於能率集團方面，該公司今年宣布以1,000萬美元投資Agility Robotics後，旗下佳能也正在積極爭取與Agility Robotics共同合作開發、生產製造機器人視覺模組，而Agility Robotics創辦人預計24日訪台，且首站參訪企業就是佳能，估計雙方將會有進一步合作進度將公開。 能率集團今年以來在AI機器人市場布局頻頻，除投資Agility Robotics外，先前也投資另一家美國機器人公司Mantis Robotics，而佳能更因此斬獲Mantis Robotics的感測器訂單，初期出貨量約300套，估第4季開始出貨，而能率則有望為其Mantis Robotics代工機器人手臂。 佳能表示，公司機器人相關訂單能見度已直達2026年下半年，目前人形機器人一台至少要8顆鏡頭，而視覺模組的產品單價更高，對公司營收與獲利有顯著貢獻，預料機器人與無人機營收比重將在明年從0達到10%，2027年則成長至20%。 ※ 歡迎用「轉貼」或「分享」的方式轉傳文章連結；未經授權，請勿複製轉貼文章內容 產業熱點 市場焦點 市場焦點 產業熱點 市場焦點 國際專欄 外媒解析 深度報導 科技產業 金融脈動 此為訂戶獨享，每月 10 則，轉傳連結即可邀請朋友免費閱讀全文。 本站使用元件之台股行情、個股基本資料及財務資訊為 凱衛資訊 提供。元件資料來源：台灣證券交易所, 櫃買中心。 本站元件所提供之金融資訊， 係供參考，不能做為投資交易之依據，所有資料以台灣證券交易所、櫃買中心公告為準。 聯合報系 著作權所有 ©2024 前往頁面

文章 特輯 訂閱 特色頻道 關於我們 讀者服務部電話：(02)2662-0012 服務時間：週一 ~ 週五 9:00～12:30；13:30～17:00 服務信箱：gvm@cwgv.com.tw 傅莞淇 2025-11-03 瀏覽數 喜歡這篇文章嗎 ? 與輝達密切合作、獲國內能率集團投資的美國人形機器人獨角獸Agility Robotics日前訪台，為近來AI機器人熱潮再添柴薪。深研機器人數十年的共同創辦人赫斯特對智慧機器的現況有何觀察？對人形機器人後進又有何建議？ 與輝達密切合作、獲國內能率集團投資的美國人形機器人獨角獸Agility Robotics日前訪台，為近來AI機器人熱潮再添柴薪。深研機器人數十年的共同創辦人赫斯特對智慧機器的現況有何觀察？對人形機器人後進又有何建議？ 在輝達引領、助長的「實體AI」熱潮中，人形機器人是一大焦點。身為最早簽下商用長約的人形機器人供應商，Agility Robotics的共同創辦人赫斯特（Jonathan Hurst）日前訪台。行程受到台廠密切關注，亦帶動投資方能率、興櫃掛牌的能率亞洲股價大漲。 在諸多廠商尋找機器人供應鏈切入點之時，深研機器人領域數十載的赫斯特對近年急起的這波熱潮有何觀察？身為人形機器人市場主要玩家，他又為何數度強調自己原本「不想做人形」？ 「很多人以為，把機器人做得像人形，再結合AI，它就能做所有人類能做的事。」赫斯特表示，「但它不能。」 在大語言模型驅動的聊天機器人打響大眾知名度後，對前沿模型能力的應用及想像延展到其他領域。代理執行數位任務已無法滿足使用者。給予如此聰明的AI模型一個身體，在現實世界中帶來更多價值，似乎是自然的下一步。 然而，必須高度整合軟、硬體的AI機器人研發，遠比純軟體模型來得困難。赫斯特指出，機器人運動數據不如網路圖文資料那樣易得。以人類遙控法累積數據，也僅適用有限用例，如不需用到太大力氣的折餐巾等任務。在跑步、全身體控制等面向，只靠遙控訓練無法達到理想成果。 或許最明顯的是，AI模型能力再強，也需要適合的身體支持，才能在實體世界做到它想做的事。而現在幾乎還不存在專為前沿AI模型打造的硬體供應鏈，需求、規格都還在成形中。諸多原因，使得大眾期待的通用機器人無法在短期付諸實用。 AI機器人的進展雖不如外界想像得簡單、迅速，但確實出現顯著突破。「我們正目睹機器人完成過去無法達成的任務。」赫斯特說。 例如，機器人後空翻這類炫技，是幾年前難以想像的。傳統的機器人動作是由工程師撰寫程式控制。要完成後空翻這樣複雜的動作，需要在極短時間內控制許多關節在正確的時間以適當的角度輸出準確的力度，只要任何一個參數有偏差，機器人就會摔倒在地。再者，也沒有明確的經濟價值驅動工程師研究這樣的動作控制。 但在模擬環境中透過強化學習（RL）等技巧，可以在極大量的反覆試錯中，引導機器人摸索出控制身體完成後空翻的作法。這種以「學習」為基礎的訓練法，為機器人控制開啟了新的可能性。 再加上結合文字、圖像、語音等多模態理解能力的AI模型，為機器人帶入更細膩的語意與環境理解能力，以及更自然的人機互動體驗。目前諸多團隊正在破解的「物理智慧」（physical intelligence），更是機器人能力再上層樓的關鍵拼圖。 因而，赫斯特認為實體AI「不只是炒作」。他表示，「數百年來，人們渴望擁有能完成所有工作的機器。突然之間，人們開始意識到，也許那並非遙不可及的未來，甚至不需要花上五十年，可能就快要達成。因此，我認為人們的興奮是有其理由的。」 這波熱潮也為機器人領域帶進更多資金、人才與關注，對整體進展更是強心劑。赫斯特認為，許多瓶頸有望加速突破。九月底在首爾舉行的機器人學習大會（CoRL），便因為報名人數超額而必須提早關閉登記，「這是前所未見的。」赫斯特指出。 赫斯特認為實體AI「不只是炒作」。黃菁慧攝 為Agility Robotics打響知名度的機器人「Cassie」，便是學習式訓練策略的先驅。控制Cassie的神經網路，是花了一週的現實時間，在模擬環境裡進行大約1年的訓練而成。 訂閱遠見電子報，掌握國內外大事 只有雙腿、沒有上身的Cassie在2022年創下雙足機器人跑百米的金氏世界紀錄。Cassie以站立姿勢起跑、越過終點線，全程沒有摔倒，共花了24.73秒完賽，至今是赫斯特最喜愛的動作紀錄。「我認為那看起來非常、非常自然。」他驕傲地說。 影／Cassie以站立姿勢起跑、越過終點線，全程沒有摔倒。 Agility Robotics在2015年脫胎自奧勒岡州立大學的動態機器人實驗室，關鍵技術就是雙足機器人的動態平衡與運動能力。在研究領域上，擁有卡內基美隆大學機器人博士學位的赫斯特專精於生物力學。宗旨是先理解動物運動的原理，再以馬達等機械硬體重現。這解釋了為什麼Cassie、Digit的步態在人類眼中看起來特別「自然」。 在研發機器人時，赫斯特強調，團隊採用的進路是「仿生啟發」（bio-inspired），而非「仿生模仿」（biomimetic）。「我們想要了解人類或動物運動的原理，再應用到機械上。」他解釋，「而不是直接複製動物。」 長久以來，Agility Robotics想打造出能去到任何人類可去的地方的機器人，在人類為自己打造的環境中，與人們共存、協作。雙腿行走能力是邁向這個願景的基礎，如人的外觀亦有其社會價值。 例如，Digit的頭部除了可以放置攝影機與光達，也有助於與人類互動。透過眨眼這類的信號，讓周圍人類可以輕易察知機器人行動的意圖。不須專業訓練，就可促成人機自然協作。 然而，只有雙腿的Cassie逐步進化為有軀幹、雙臂的Digit的過程，實則充滿著工程計算，而非對人形機器人簡單、直覺的期待。 Digit的頭部除了可以放置攝影機與光達，也有助於與人類互動。Agility Robotics提供 赫斯特解釋，Cassie早期的一大問題是偏航控制能力不足，這導致機器人跨步時容易水平旋轉而摔倒。這是因為機器人跨出一側腿部時，會產生一個反向作用力。Cassie的腳掌太小，摩擦力不足以抵抗這個反作用力，所以容易原地旋轉後側向倒下。 為了解決這個問題，團隊想過安裝一個反作用輪或陀螺儀來抵消反作用力。但這個裝置除了旋轉沒有其他功用，還會增加機器人的重量與耗能。 更好的作法似乎是安裝另一個可移動的質量來施加反作用力，例如說尾巴。尾巴裡的空間還可以安置其他元件。而要控制偏航角度，裝設尾巴較好的位置不會是正後方，而是在對稱中心軸的兩側，可以輪流抵銷兩腿的反作用力。 若尾巴從兩腿旁側伸出，移動時容易撞到周圍事物。藉由將兩隻尾巴的位置上移，機器人也有了可以協助啟動步伐的軀幹，可以透過前傾來改變身體重心，利用重力助力移動。最後，兩隻裝在軀幹兩側的尾巴正好就像是雙臂。 賦予機器人雙臂，也可提升操控物品的能力，以及在摔倒時支撐自己的能力。赫斯特表示，「就像是有三條完全彼此獨立的思路，最終都把你引導到雙臂連結軀幹的同一個位置。」 「我們試過把手臂安裝在別的位置，但（肩部）才是正確答案。」赫斯特說，「所以最後它看起來就像是人形了。」 赫斯特解釋，起初他刻意避免打造人形機器人的原因，是為了防止自己陷入「先入為主」的窠臼，預先認定人形會是理想解，而失去探索其他可能性的機會。 「如果我們做出一個像是人的機器人，我們希望那是一個巧合。」他表示，「我們希望有充足的工程學理由支持這個設計。如果最終的解決方案正好是人形，那就採用，但那不是預設的目標。」 有趣的是，目前Digit最「不像」人的地方，也即將在2026年的改版中大幅調整得更像人類，而這背後也有著審慎的工程考量。 Digit使用鳥腿式設計已久，身體下降時，看似膝蓋的踝部是向後彎曲。這利於走路、跑步，但也使得Digit無法蹲得非常低，因為踝部會撞到地面。為了能從地面抬起重物，Digit也需要能將軀幹前傾，借用髖部和膝蓋的馬達來分擔重量。團隊發現，人類的腿部設計更適合進行負重的操控。 赫斯特預告，除了快充、可換手等更新，第五代Digit在外貌上也將會看起來相當不同，但希望能同樣討人喜愛。 赫斯特初次訪台，對台灣硬體供應鏈的觀察具體而微。他指出，台灣擁有一群生產零組件的中小企業，深知製造流程，也具有客製化經驗。在發展新一代機器人的研發團隊眼中，台廠這樣的能力相當令人興奮。 赫斯特也表示，從機器人學家觀點來說，要把零件、模組做得品質精良、穩定發揮應有功能，可能是最難的環節。相對來說，組裝一台Digit的時間不到一周，佔總成本不到5％。若台灣擁有蓬勃發展的零組件生態系，他相信也能孕育出整機機器人供應商。 對此，赫斯特給機器人新創後進們的建議是：打造出產品，而不只是人形機器人。他舉例，宇樹能以低成本生產可跳舞、打拳擊的機器人，但它們並不適合長時間在倉庫裡搬運重物。即使外型都與人類相近，H1與Digit並非競爭相同市場的產品。 他強調，較好的方式是打造解決方案，為客戶提供商業價值，而不是「為做而做」人形機器人。 早在2020年就投資Agility Robotics的創新工業技術移轉股份有限公司（ITIC）現任總經理余宛如，是這次赫斯特訪台的重要推手。選在Agility期待提升量產能力的時刻，強力推銷台廠作為硬體策略夥伴，是說動赫斯特親自來台的一大原因。 余宛如觀察，台廠的優勢除了價格相對有競爭力、英文溝通能力佳，生產管理的軟實力也可以是一大特色。面對新一波機器人商機，要如何在既有優勢上疊加出新的市場區辨度，企業們皆相當積極。 創新工業技術移轉股份有限公司（ITIC）現任總經理余宛如（左）、Agility Robotic共同創辦人Jonathan Hurst（右）。黃菁慧攝 Agility Robotics與能率集團的合作可視為先例。在供應鏈多元化的趨勢中，台廠如何挑對題目、提供可靠解方，將是在新一代AI機器人硬體戰中脫穎而出的關鍵能力。 延伸閱讀 人形機器人Digit已上工一年！為何最怕它「摔倒」？ 早鳥報名｜國際級師資，嚴長壽 53 年經營智慧親授！ 加好友掌握趨勢 關注FB熱門議題 AI治理有方 新北市打造智慧城市資安新典範 洞察趨勢的新北市政府，以超前部署、積極布局的策略，致力打造宜居、創新的智慧城市，聚焦「基礎建設」、「關鍵技術」、「智慧應用」三大面向，新北市已同步展現具體成果，體現地方政府的前瞻思維與執行力。 熱門 訂遠見再享25+20吋雙行李箱 「有選擇」才是真正的財富自由 《阿凡達：火與燼》史詩年度大作 走進嘉義：藝術如何改變鄉村風貌 螢幕亮起新北把希望送孩子手中 BMS翻轉職場打造人才跨域成長 三重變局下的台灣永續新航道 請往下繼續閱讀 熱門 訂遠見再享25+20吋雙行李箱 「有選擇」才是真正的財富自由 《阿凡達：火與燼》史詩年度大作 走進嘉義：藝術如何改變鄉村風貌 螢幕亮起新北把希望送孩子手中 BMS翻轉職場打造人才跨域成長 三重變局下的台灣永續新航道 近期熱門文章 曹西平驟逝享壽66歲！家屬不認領恐成「有名無主屍」乾兒子：盼後事全權給我處理 Google AI Pro打折快截止！Gemini付費或免費好？ 2026長照3.0全新上路！服務項目、補助新增什麼？申請一次看 看此文章的人也看了... 加好友掌握趨勢 關注FB熱門議題 登入網站會員 享受更多個人化的會員服務 加入網路會員，享受讀者專屬權益 關於我們 文章 讀者服務部電話：(02)2662-0012 服務時間：週一 ~ 週五 9:00～12:30；13:30～17:00 服務信箱：gvm@cwgv.com.tw Copyright© 1999~2026遠見天下文化出版股份有限公司. All rights reserved. 請登入會員 此為會員限定功能立即登入 今天準備好明天的競爭力！精選國際、產經時事，快速掌握最新趨勢。

Amazon testuje wykorzystanie w swoich magazynach humanoidalnych robotów o wysokości ponad sześciu stóp, co zwiększa obawy co do zwolnień pracowników. Eksperyment zostanie przeprowadzony w ośrodku eksperymentalnym BFI1 firmy Amazon w Sumner w stanie Waszyngton. Robot, znany jako Digit, to dwunożny robot, który jest w stanie chwytać i podnosić przedmioty w magazynach Amazon. Digit ma dwie ręce, dwie nogi, niebieską klatkę piersiową i dwa kwadratowe światła zamiast oczu. Może poruszać się do przodu i do tyłu, obracać się i schylać. Może dosięgnąć, chwycić i podnieść żółte pudełka Amazon. Więcej na temat wprowadzania tego typu systemów w zakładce "Przyszłość pojazdów bezzałogowych": Firma twierdzi, że Digit pomoże pracownikom magazynu w recyklingu toreb, zbierając i przenosząc puste torby po usunięciu wszystkich znajdujących się w nich przedmiotów. Aby stworzyć Digit, spółka zależna Amazona zajmująca się robotyką, Amazon Robotics, nawiązała współpracę ze start-upem technologicznym Agility Robotics. Ten start-up z siedzibą w Oregonie koncentruje się na tworzeniu robotów dla firm logistycznych i magazynowych. Amazon za pośrednictwem funduszu Amazon Industrial Innovation Fund przekazał Agility Robotics kwotę 150 milionów dolarów na pomoc w stworzeniu Digit. Pod koniec okresu próbnego Amazon planuje wdrożyć robota „w nowatorski sposób w przestrzeniach i narożnikach magazynów”. „Wierzymy, że istnieje duża szansa na skalowanie rozwiązania manipulatora mobilnego, takiego jak Digit, które może współpracować z pracownikami” dodała firma Amazon. Link do oryginalnego artykułu: LINK | Humanoidalne roboty przybywają by przejąć pracę w magazynach. | Południowokoreańscy naukowcy stworzyli robo-pilota wykorzystującego Ai. | Po raz pierwszy armia USA pomyślnie wykonała lot myśliwcem wykorzystującym AI. | Dron kontrolowany przez sztuczną inteligencję „zabił” ludzkiego operatora w symulowanym teście sił powietrznych USA. | Ekspansja generatywnych systemów sztucznej inteligencji może znacząco zakłócić rynek pracy. | Ai w samochodach? Robo-taksówki powodują chaos w ruchu drogowym w Austin. | Microsoft włączy obowiązkowe systemy AI w systemie Windows 11, aby szpiegować co robisz na komputerze. | Firmy wprowadzają do szkół opaski kontrolujące myśli uczniów i monitoring AI. | Znikające zawody do roku 2030. 50% zawodów ulegnie komputeryzacji. | Czy aktorzy głosowi stracą pracę na rzecz sztucznej inteligencji? | Scenariusz rozwoju rynku pracy do roku 2050. | Bill Gates: Ludzie nie zdają sobie sprawy, ile miejsc pracy zostanie wkrótce zastąpionych przez boty. | Dlaczego przyszłości nas nie potrzebuje. | Implikacje zmian demograficznych do roku 2050. | Miliarderzy biorą populację na celownik. | Telemarketer robot, który zaprzecza, że jest robotem. | Czy Robo-Reporterzy zastąpią dziennikarzy głównego nurtu? | Sztuczny mózg przeszedł podstawowy test IQ. | Pierwsza bezzałogowa flota ciężarówek rusza w Australii. | Pentagon będzie budował roboty z "prawdziwym" mózgiem. | Armie wirtualnych przyjaciół promują propagandę poprzez sieci społecznościowe. | NBIC- zabawa w Boga. Obraz świata w 2025 roku. | Przyszłość pojazdów bezzałogowych. | ONZ przewiduje transhumanistyczną przyszłości, w której człowiek będzie zbędny. | Pola bitwy 2050. | Dlaczego przyszłości nas nie potrzebuje. | Elity Doliny Krzemowej założyły kult religijny mający czcić sztuczną inteligencję jako Boga. | Twórca firmy Apple: “Roboty podbiją świat, i będą nas trzymać jako zwierzęta domowe.” | ONZ przewiduje transhumanistyczną przyszłości, w której człowiek będzie zbędny. | Chatboty AI zastąpią wyszukiwarki i wirtualnych asystentów. I twoich przyjaciół. | Elon Musk: ludzie muszą połączyć się z maszynami lub stać się nieistotnymi w wieku sztucznej inteligencji. | Samoświadoma sztuczna inteligencja może pewnego dnia kontrolować każde urządzenie na świecie, przewiduje wynalazca Andy Rubin. | Największy na świecie fundusz hedgingowy zastąpi menedżerów sztuczną inteligencją. | Sztuczna inteligencja Microsoftu stała się nazistą w mniej niż 24 godziny. | NSA i sztuczna inteligencja do wykrywania przed-zbrodni. | Sztuczna inteligencja: "Homo sapiens zostanie podzielony na kilku bogów i resztę". | Eric Schmidt: "Sztuczna inteligencja powinna mieć systemy weryfikacji, aby uniknąć niepożądanych efektów. | Sztuczna inteligencja będzie decydowała o życiu i śmierci ludzi. | Eric Schmidt: Sztuczna Inteligencja Google nie do odróżnienia od ludzkiej w przeciągu dekady. | Wyłączniki robotów i legalny status dla maszyn: posłowie EU poparli wniosek dotyczący sztucznej inteligencji. | Nowe oprogramowanie "sztucznej inteligencji" posiada niemalże ludzkie zdolności rozpoznawania obrazów. | Skynet, robotyka i globalne trendy strategiczne. | Bogaci Amerykanie szukają na czarnym rynku możliwości wszczepienia implantów mózgowych by podłączyć się do sztucznej inteligencji. | Pola bitwy 2050. | Znikające zawody do roku 2030. 50% zawodów ulegnie komputeryzacji. | Człowiek postindustrialny. |

Oops, something went wrong MercadoLibre, Inc. (NASDAQ:MELI) is one of the 14 Most Promising Fintech Stocks to Invest In. On December 10, MercadoLibre, Inc. (NASDAQ:MELI) and Agility Robotics announced a commercial agreement. Agility Robotics is the creator of the leading humanoid robot called Digit. This agreement will bring Agility Robotics’ Digit humanoid robot into MercadoLibre, Inc.’s (NASDAQ:MELI) facility in San Antonio, Texas. Initially, Digit will help with tasks that support commerce fulfillment. MercadoLibre, Inc. (NASDAQ:MELI) and Agility Robotics also plan to find more use cases where AI-powered humanoids can support and add value to logistics operations in MercadoLibre, Inc.’s (NASDAQ:MELI) warehouses across Latin America. The main goal here is to automate jobs that are hard to recruit for, especially those that are extremely repetitive and physically difficult. This could help improve safety for the company’s workers and also reduce labor gaps. By using robots and automating these tasks, MercadoLibre, Inc. (NASDAQ:MELI) hopes to increase productivity and allow employees to focus on more value-added work. In other news, on December 5, TipRanks reported that Citi analyst Joao Soares reiterated a Buy rating on MercadoLibre, Inc. (NASDAQ:MELI). The research firm has a price target of $2,500 on the stock. On December 3, TipRanks reported that BTIG analyst Marvin Fong reaffirmed a Buy rating on MercadoLibre, Inc. (NASDAQ:MELI) with a price target of $2750.00. MercadoLibre, Inc. (NASDAQ:MELI) is the leading e-commerce and financial technology company in Latin America with a presence in 18 countries. While we acknowledge the potential of MELI as an investment, we believe certain AI stocks offer greater upside potential and carry less downside risk. If you’re looking for an extremely undervalued AI stock that also stands to benefit significantly from Trump-era tariffs and the onshoring trend, see our free report on the best short-term AI stock. READ NEXT: 15 Best Technology Penny Stocks to Buy and 15 Best Aggressive Growth Stocks to Buy Right Now. Disclosure: None. This article is originally published at Insider Monkey. Sign in to access your portfolio