Topic
Robotics
58 stories, newest first.
-
Disney's research team has built three small Star Wars characters, the Anzellans, that, Disney says, use AI running on Amazon's cloud computers to decide what to say and how to move when someone waves or talks to them. Disney says its older Audio-Animatronics figures perform the same show every time. Disney says every line and personality quirk they can choose from was written by its designers. They are a prototype: people attending an Amazon conference in Las Vegas from Nov. 30 to Dec. 4 can meet them, and Disney says only that guests can one day meet characters like them, without naming a park or a date.
-
The companies say chipmaker GlobalFoundries will be the first customer outside the car industry to test Boston Dynamics' Atlas humanoid robot, with trials at GlobalFoundries' plants starting in early 2027. The deal also widens the chipmaker's use of Boston Dynamics' four-legged Spot inspection robot across its plants in New York, Vermont, Germany and Singapore. The announcement does not say how many Atlas robots will be tested, at which plants or on which tasks.
-
Genesis AI, a robot maker based in Paris and California, has raised $500 million at a $3.5 billion valuation, in a round led by Premji Invest, Khosla Ventures and Eclipse, with NVIDIA and AMD's venture arm among the other investors. It plans to put its first robot, Eno, a wheeled machine with two arms and hands shaped like a person's, to work in factories, warehouses and laboratories first. Genesis is testing only a small number of robots with customers so far, according to a news report, and has not said what Eno will cost.
-
Engineers at the University of Pennsylvania made thin artificial muscles that sense how far they have moved, so they can stop partway and hold still. Driven by them, robot fingers pressed piano keys softly or loudly and a pair of robot eyes shifted their gaze. So far the muscles have only been shown in lab models, and the researchers say future work could focus on how the muscles handle heat and use energy.
-
ASUS, which has spent decades making motherboards and PCs, has shown Gallop, its first humanoid robot, which is learning to pack motherboard accessories at an electronics factory in the ASUS group, mostly by watching the workers who do the job. ASUS plans trials with customers in 2027 and aims to start selling the robot in the second half of 2028. It has given no price, and has not said how close Gallop is to its target of packing correctly 99.5% of the time within 20 seconds, the standard ASUS holds a person to.
-
Agile Robots, a Munich company that makes robot arms, robot hands and humanoid robots for industrial customers, has raised $500 million at an $8.5 billion valuation, in a round led by SoftBank with NVIDIA and Hitachi taking part. It says the money will pay for growth in the United States and more production, including in Germany, where a factory it opened early in 2026 can, it says, build up to 10,000 humanoid robots a year. The company has not said how many of the more than 20,000 industrial robot systems it has installed since 2018 are humanoids.
-
Ring, Amazon's home security brand, is taking orders for Rover Cam, a $999.99 wheeled robot that patrols an office, store or warehouse on its own and ships in the U.S. on Dec. 16. Ring's $249.99 Drone Cam, a camera that flies from room to room inside a home, is due to ship in the U.S. on Jan. 27, 2027, but cannot be sold until the U.S. Federal Communications Commission authorizes it. Ring says both draw their own map of the space and head out when another Ring device is triggered.
-
Google DeepMind released Gemini Robotics 2.5, the AI that controls its partners’ robots, and says it succeeds 30% to 100% more often than Google’s earlier models in its own tests, depending on the type of task, with big jumps on fiddly jobs like packing pastries with tongs. In a trial on Mercedes-Benz production lines in Hungary and Germany, Google says Apptronik’s Apollo 2 robots running it sorted car parts into trays with a 96% success rate. Only Google’s robotics partners are getting it so far, and others can join a waitlist. Every result so far comes from tests by Google and its partners.
-
NVIDIA unveiled GR00T 2, an AI model for robots trained mostly on hundreds of thousands of hours of video of people at work, which NVIDIA says lets a robot carry out skills it was shown only by people or on other robots. In NVIDIA's tests on 50 tasks its robot had never practiced, the model succeeded 40% of the time with no practice and 95% after five demonstrations of each task. NVIDIA calls the model open, but its project page says the code arrives in December, and so far the results come only from NVIDIA.
-
Boston Dynamics bolted scooter wheels onto its Atlas humanoid robot, and software that learned in a computer simulation drives it, rolling, sliding sideways and even jumping and spinning on them. The company calls the wheeled Atlas a research test, not a product, and says its main lesson, that the robot learns better when its simulated motors behave like the real ones, is already going into the walking Atlas. It has not said what specific changes that brings.
-
Apple researchers trained a Unitree humanoid robot mostly on practice runs a computer generated in 1,000 made-up rooms, with no one steering the robot to collect them. They say it then succeeded in 91% of 360 tries at single tasks, such as sitting on a chair or crawling under a table, in six real homes it had never seen. Chains of two or three tasks, tested in one real room, fully succeeded 70% of the time, and the results come only from the researchers' own tests.
-
Robot company LiberAI says its new AI, Liber-0, learned to handle objects mostly from more than 2 million hours of recordings of people, some made with a glove that tracks finger movements. On tasks it had never practiced on a robot, it succeeded about a third of the time in LiberAI's own tests, and long chores with several steps were the hardest. Anyone can try a simulated version online, but LiberAI hasn't said whether it will offer the model itself or sell robots that run it.
-
Agility Robotics, which makes Digit, a robot built in a human shape for warehouses and factories, began trading on Nasdaq on Nov. 5 as AGLT, after its merger with Churchill Capital Corp XI and related share sales raised about $465 million before costs. Agility calls itself the first U.S.-listed company that only makes humanoid robots and already has them at work for customers. Agility expects its newest robot, Digit 5, to reach its first customers in the first half of 2027.
-
Shield, whose software flies drones and other military aircraft on its own, has agreed to buy Echodyne, whose radars spot and track drones, in a deal valued at up to $5.3 billion. To help pay for it, Shield is raising $2.5 billion, including a $1 billion investment that values Shield at $28.8 billion, more than double its value in March. The deal needs regulators' approval and is expected to close in early 2027.
-
A Chinese government office that screens foreign investment for security risks has barred SoftBank from buying the China part of ABB's robot business, ABB said on Nov. 5. The Swiss company agreed in 2025 to sell its whole robotics division, which makes factory robots, to the Japanese investor for $5.375 billion. ABB says the two companies are still working on the deal, but ABB has not said what the decision means for the rest of the sale.
-
Boston Dynamics has unveiled Stretch Gen 2, a new version of its robot that unloads boxes from truck trailers, rebuilt around the most common causes of unplanned stops on the first one. The company says it cut most of the robot's internal wiring and made one big repair take about an hour instead of a day. Boston Dynamics has not given a price or said when the new robot ships.
-
Researchers at the RAI Institute have taught a Spot robot to stand a 15-kilogram car tire upright, roll it and drag it to a goal a few meters away outside the lab, seeing the tire only through cameras built into its own body. RAI's earlier version needed cameras in the ceiling and a desktop computer, so it worked only in the lab. Anything blob-shaped, even a person walking past, can still look like the tire to it, and it has only been tried with one tire on flat ground.
-
Agility Robotics, which makes a robot built in a human shape for warehouses and factories, said its shares would start trading on Nasdaq on Nov. 5 and that it expected to raise about $465 million. The deal closed on Nov. 4, the companies said. In June the companies expected more than $620 million, but only about $164 million of the $420 million held by Churchill Capital Corp XI, the listed company Agility is merging with, is now expected to reach Agility, after Churchill's shareholders asked for their money back. The companies said Churchill's shareholders would vote on the deal on Nov. 2.
-
President Trump signed a proclamation on Wednesday agreeing with a Commerce Department finding that robots, factory machines and their parts are coming into the country in amounts and ways that threaten U.S. national security. Instead of setting tariffs now, it tells U.S. trade officials to negotiate with other countries and report back within 180 days (by late April 2027), and says the president may consider tariffs, quotas or fees later, depending on how the talks go. It does not name any country or say how high any tariff might be.
-
Tesla says it has started building Optimus, its humanoid robot, on what it calls "high-volume lines" at its factory in Fremont, California, and calls them the first production-quality robots it has made. Chief executive Elon Musk said production will grow slowly and the robot's AI still needs a lot of work, according to a news report. Tesla's letter does not say how many it has built or what jobs they can do on their own.
-
Two German ministries signed a declaration of intent with NVIDIA to develop AI models that run factory machines and robots, starting from NVIDIA's open models and drawing on German industrial data. NVIDIA also plans to invest in four German startups. Neither side has said how much money the joint work involves or when the first models will be ready.
-
Robot maker Unitree has released UniUni, a free app for Windows and Mac computers that plans and runs group performances by its humanoid robots, from each robot's path across the stage to the moves it makes on cue. Users can also type what they want, such as six robots in two rows doing a dance, and Unitree says the app's AI maps out the paths and picks the moves. It is built for Unitree's own H2 and G1 humanoids, and so far only Unitree's own examples show how well the AI works.
-
Kyutai's robot model remembers the past few seconds, which the lab says helps an arm on longer tasks
Kyutai, an AI research lab in Paris, has released REKI, a model that turns what a robot's camera sees and a typed instruction into arm movements, along with all the code used to build it. Unlike most such models, the researchers say, it remembers the past few seconds, which they say made arms in simulated tests get stuck moving back and forth about a quarter as often, and helped a real arm on longer tasks. Its files are free for non-commercial uses such as research, but not for business use, and so far it has been tried on only one kind of real arm, in the researchers' own tests.
-
Google DeepMind says its new Gemini 4 Argon model watched a video of a person playing cornhole, worked out the moves for Apptronik's Apollo robot and wrote the computer simulation for them on its own, and the robot then sank seven throws in a row. Google says this can shrink weeks of hand-built simulation work to a few hours. It was an internal test, and so far the results come only from Google.
-
Nature Will, a robotics AI startup, says its new AI, Natura-1, first learned from more than a million hours of video filmed from people's own point of view, and that a limited amount of robot practice was then enough to adapt it to a pair of robot hands. Its videos show the hands folding wontons, serving tea and following odd requests, but the six long chores, wontons and tea included, got extra training for those jobs, and the lab has not said whether or how anyone else can use it.
-
Anthropic has updated the rules for using its Claude AI, spelling out bans on building weapons software, arming drones and tracking people without consent, and adding safety demands for Claude when it controls machines that could hurt someone. It also bars users from sustained, needless cruelty toward Claude itself. Anthropic said the new rules would take effect on Nov. 12, and it is not yet clear how Anthropic will police the cruelty ban beyond letting Claude end a chat.
-
Amazon Web Services has released an open-source kit that strings together the steps of building a robot's AI, from a robot's recorded work to testing its AI in a computer simulation and loading the finished AI onto the robot. The code is free, but it runs on Amazon's paid cloud and NVIDIA's robot software. Amazon has not said what a typical project costs to run.
-
Boston Dynamics has named Rohit Prasad, who spent 12 years at Amazon helping build the Alexa voice assistant and later its Nova AI models, as chief executive from Oct. 7. The robot maker says the move is meant to pair its machines with advanced AI and sell them at scale. Its announcement gives no product plans or timelines.
-
Boston Dynamics has a new four-finger hand for Atlas, its human-shaped robot, built to use tools like drills and to turn objects around in its grip. The company left out the pinky on purpose to keep the hand simpler and sturdier, and it has not said when the hand will reach working robots.
-
The U.S. Army announced on Oct. 2 that its acting secretary has signed a memo ordering the creation of the Army Futures and Autonomous Systems Command and of a single office to buy the Army's priority autonomous systems. The memo tells the Army to make it a priority to buy autonomous weapons, combat vehicles, boats that carry supplies and reconnaissance systems and get them to troops no later than the end of the federal fiscal year 2028, in September 2028. The Army has since said a three-star general will lead the command from Fort Eustis, Virginia, but not who, what it will cost, or how much control people will keep over these systems.
-
Praxis-1 is Runway's new model for controlling real robots, built on what its video AI learned about how objects and hands move. Runway says that gives robots a head start, since real-world robot data is scarce and costly to collect.
-
Sparse-WAM is a shortcut for robot AI that predicts what its moves will do: it works out which parts of the predicted scene matter for the robot's next move and skips the rest, with no extra training. Its authors say it nearly doubled the speed of two such robot models while the robots still did their tasks about as well. No one outside the team has tested it yet.
-
On Sept. 29, DYNA Robotics launched DYNA 2.1, an AI system for its wheeled, semi-humanoid robot Taku, and said the robot can work an entire commercial laundry shift without a person stepping in. DYNA is betting that businesses want robots that finish whole shifts, not single tasks. It is not yet clear how that holds up outside the company's own demos.
-
Flourish Robots, a startup that had kept its work under wraps, introduced Flourish 1 on Sept. 29: a $3,555 home robot that it says owners can teach a household task in about 30 minutes by guiding it from a phone. Pre-orders for the first 50 numbered units opened the same day, with deliveries due to begin in December 2026. How well that teaching works outside the company's own demos has not been shown.
-
Researchers at Beihang and Peking University posted a paper on Sept. 28, 2026, describing RTP, a method that lets a robot's AI keep the short video it predicts of what should happen next and revise it as the robot acts, instead of always making a new one from scratch. In their tests on two sets of robot tasks, RTP scored higher than every other method they compared. No outside lab has tested it yet.
-
Humanoid robots, not electric vehicles, are shaping up as the first commercial market for solid-state batteries, according to a news report citing industry sources. Samsung SDI has picked robots as the first target for its SolidStack batteries, with mass production aimed at the second half of 2027. Cost could slow the shift: analysts put early prices at $400 to $600 per kWh, against roughly $350 per kWh needed for robot batteries to pay off.
-
On Sept. 25, 2026, InternRobotics researchers posted InternW0-Delta, a robot AI that predicts how a scene will change and feeds those predictions straight to the part that chooses the robot's moves, without first generating video of the future. They trained it on more than 20,000 hours of robot and human data and plan to release the data, the code and the model where licenses allow. No one outside the team has tested it yet.
-
AD-WM is a way of training a robot's built-in "what if" simulator so it can tell apart the results of different moves. The researchers say a real robot arm using it picked up and placed objects about seven times in ten, up from about four.
-
Rolling-WAM lets a robot keep picturing what comes next without redrawing that picture from scratch each time it updates its plan. The researchers say that makes its plan updates more than four times faster than the standard approach.
-
East Palo Alto startup Feather is selling a two-armed humanoid robot on wheels, built for outside developers to program. It listed the robot from $29,990 in September, and by Nov. 16 its site gave a starting price of $32,990. An update from Seed Innovations, one of its investors, says Gradient Ventures led a $7.6 million pre-seed round. Feather says it has already booked more than $1 million in revenue, a figure that comes only from the company.
-
Tokyo startup O-ID raised $1.2 million in pre-seed funding to build factory humanoids, robots built in a human shape, whose joints, limbs and other parts it says can be swapped out on the factory floor in minutes. O-ID says Chicago's TAWANI Ventures led the round, and that it signed a letter of intent to discuss using Sumitomo Electric's wire-harness technology. The robots are still prototypes.
-
InternW0, from Shanghai AI Laboratory, lets a robot keep forecasting what comes next while it reacts quickly to what it sees and feels. The lab tested it on science chores such as preparing a chemical solution and using a pipette.
-
On Sept. 22, 2026, four Unitree humanoid robots walked the Galleria Vittorio Emanuele II runway during Act IV "The Age of Intelligence" at Vogue World 2026: Milano.
-
Qualcomm Technologies said on Sept. 23, 2026, that it has agreed to buy PickNik Inc., the Boulder company behind MoveIt, widely used open-source software that plans and controls the movements of robot arms. The companies say MoveIt 1 and 2 will stay open source and keep working with any hardware, while Qualcomm ties it more closely to its Dragonwing robotics platforms and Arduino boards. Terms were not disclosed, and the deal has not yet closed.
-
XPENG said on Sept. 8, 2026, that it had put its humanoid robot production lines in Guangzhou into operation, and that its first IRON robot walked off the line on its own once it was built. The company says more than 80% of core processes on the lines are automated. Starting the lines is not mass production, which XPENG still targets by year-end, with deliveries in China and overseas planned for 2027.
-
Humanoids Summit Seoul opened at COEX on Sept. 22, 2026, with South Korea's Ministry of Science and ICT (MSIT) on the Day One program for congratulatory remarks and a briefing on Korea's policy for AI that runs robots and other machines. Companies pitched plans to put humanoid robots to work in industry, but the sales head of Unitree's South Korean distributor, quoted in a news report from the show floor, said fully autonomous humanoids practically do not exist yet. Whether the policy leads to robots doing real factory work is not yet clear.
-
Boston Dynamics opens a center at Hyundai's Georgia plant to train Atlas humanoids for car factories
Boston Dynamics said on Sept. 21, 2026, that the first phase of its Robotics Metaplant Application Center is running inside Hyundai Motor Group Metaplant America near Savannah, Ga., training Atlas humanoid robots to prepare car parts and put them in order for the assembly line. Hyundai Motor Group plans 25,000 Atlas robots at Hyundai and Kia plants over the next few years, but Boston Dynamics declined to tell reporters how many are at the center now.
-
Unitree said on Sept. 21, 2026, that it was selling the Dex5-S, a five-finger robot hand the size of a human hand with 22 motor-driven joints, from $6,500 before tax and shipping. Its product page now lists the hand from $6,350 and says touch sensing comes only on a Pro version. An industry newsletter noted that Unitree had not yet published force, speed or durability data for the hand.
-
DeformSmith, from Nankai University and Rightly Robotics, turns a short description or a single image into a soft virtual object that squashes and springs back. Robots could then practice on many more soft things, without measuring each real one first.
-
OpenAI's San Francisco robotics job listings offer base pay of up to $500K for a machine-learning engineer working on data systems, among about 26 core robotics openings. Many of the roles involve collecting robot training data, and a news report describes the push as building a data operation for robots, not a product launch. OpenAI declined to comment on the hiring, another news report said.
-
The National University of Singapore, Tsinghua University and partner universities released OpenWAM, a free kit for building robot AI on top of video generators. The team says publishing all of it lets other labs test which design choices make robots better.
-
Chinese regulators are quietly slowing humanoid-robot companies' plans to go public after Unitree's volatile stock market debut, according to a news report on Sept. 21, 2026. The report's sources describe informal guidance from regulators, not a published ban, and they disagree on whether it amounts to a freeze. It is not yet clear whether the China Securities Regulatory Commission (CSRC) will publish a rule for the sector.
-
Unitree says its UnifoLM-X2-1.0 AI drove a G1 humanoid robot through a sparring match with a person, with no one steering it remotely and no scripted moves. The company released no test results, and a news report says the footage hints that some of the computing may happen on a separate computer rather than on the robot.
-
Researchers at X-Humanoid, a Beijing robotics center, have built Pelican-Sim 1.0, an AI that predicts what a robot's camera will see as the robot carries out a series of moves. Robot software can then be trained and tested on that predicted video, with fewer trials on real machines. In the team's own tests, adding its practice runs raised a robot control program's success rate from 70% to 93%, and the model has not been released.
-
UBTECH started production on Sept. 12, 2026, at its new robot factory in Liuzhou, China, which the company and city officials say is designed to finish one robot every 10 minutes and to make more than 10,000 Walker S humanoids and Cruzr robots a year. That is a plan, not robots shipped: according to a news report, UBTECH delivered 921 full-size humanoids in the first half of 2026.
-
XPENG Robotics built XPACE, an AI that teaches its IRON humanoid partly from videos of people and lets it rehearse fixing mistakes in imagined scenes. That could cut the costly work of steering robots by hand to show them every task.
-
Figure says a humanoid robot running its new Helix 2.5 AI finished whole household chores in 30 Bay Area homes it had never seen 56% of the time, in the company's own tests. Figure credits early training on its large collection of human-behavior data: in its comparison, a version trained from scratch succeeded 9% of the time. No one outside Figure has repeated the tests, and the robot still failed on about 44% of attempts.
-
Odyssey's Odyssey-3 is an AI trained on huge amounts of video to learn how the world works, and the company says it can steer robots, cars, drones and game characters. If that holds, teaching a new machine could take hours of examples, not mountains of data.