Good and fast for equipment/factory automation, but yes, sounds dangerous if using fake data for AI vaccine development.
Good and fast for equipment/factory automation, but yes, sounds dangerous if using fake data for AI vaccine development.
•
NVIDIA CEO Jensen Huang introduces GR00T-Dreams, an artificial intelligence-powered humanoid robot trained on synthetic data, at the Computex conference on May 19, 2025 in Taipei, Taiwan. (Screenshot of NVIDIA presentation at Computex 2025)
NVIDIA says its latest general-purpose robot update can teach a humanoid machine using software entirely, bypassing the months of motion capture sessions most rivals still depend on.
The chipmaker unveiled “GR00T-Dreams” at the Computex trade show in Taipei on Monday, pitching it as a shortcut that turns a single photo into short simulation videos and, from there, into the step-by-step instructions a robot needs to move. In effect, engineers feed the robot a “dream” reel instead of hours of recorded factory walks, cutting development time and cost.
“Physical AI and robotics will bring about the next industrial revolution,” Jensen Huang, CEO of NVIDIA, said in a press release. “From AI brains for robots to simulated worlds to practice in or AI supercomputers for training foundation models, NVIDIA provides building blocks for every stage of the robotics development journey.”
NVIDIA’s research team claims the approach already trimmed training its previous GR00T model to 36 hours — work that previously took three months of of live motion capture and annotation. The company generates the footage inside its Omniverse and Cosmos simulators, then runs the best trajectories through Cosmos Reason, an AI filter that tosses out awkward or impossible moves before they ever reach a motor.
In short, the robot watches thousands of ultra-realistic practice rounds that only exist on a server, learns what succeeds, and steps onto the factory floor already knowing how to stack boxes or sort parts.
“The age of generalist robotics has arrived, with breakthroughs in mechatronics, physical AI and embedded computing — just in time, as labor shortages limit worldwide industrial growth,” a promotional video during NVIDIA’s Computex presentation said. “A major challenge for robot makers is the lack of large-scale, real and synthetic data to train models. Human demonstrations aren’t scalable; limited by the number of hours in a day.”
That stands in stark contrast to China’s humanoids, which still gorge on real-world footage. Unitree, a Shanghai-based firm, furnished its H-series robots with an open-source trove of full-body motion capture data so the machines could nail a folk dance routine at this year’s Spring Festival Gala. AgiBot goes even further: the Shanghai startup runs its sandwich-making, shirt-folding prototypes in 17-hour warehouse shifts, records every step, then feeds footage into DeepSeek overnight so the robots can relearn their chores the next morning — an approach Beijing is backing with billions in subsidies. Neither NVIDIA, Unitree nor AgiBot responded to the Daily Caller News Foundation’s requests for comment.
By eliminating the training bottleneck, NVIDIA could speed past its hardware-rich Chinese competitors and set the de-facto operating standard for humanoids — a market Goldman Sachs pegs at more than $38 billion by 2035.
Archive (archive.today)
>NVIDIA CEO Jensen Huang introduces GR00T-Dreams, an artificial intelligence-powered humanoid robot trained on synthetic data, at the Computex conference on May 19, 2025 in Taipei, Taiwan. (Screenshot of NVIDIA presentation at Computex 2025)
>NVIDIA says its latest general-purpose robot update can teach a humanoid machine using software entirely, bypassing the months of motion capture sessions most rivals still depend on.
>The chipmaker unveiled “GR00T-Dreams” at the Computex trade show in Taipei on Monday, pitching it as a shortcut that turns a single photo into short simulation videos and, from there, into the step-by-step instructions a robot needs to move. In effect, engineers feed the robot a “dream” reel instead of hours of recorded factory walks, cutting development time and cost.
>“Physical AI and robotics will bring about the next industrial revolution,” Jensen Huang, CEO of NVIDIA, said in a press release. “From AI brains for robots to simulated worlds to practice in or AI supercomputers for training foundation models, NVIDIA provides building blocks for every stage of the robotics development journey.”
>NVIDIA’s research team claims the approach already trimmed training its previous GR00T model to 36 hours — work that previously took three months of of live motion capture and annotation. The company generates the footage inside its Omniverse and Cosmos simulators, then runs the best trajectories through Cosmos Reason, an AI filter that tosses out awkward or impossible moves before they ever reach a motor.
>In short, the robot watches thousands of ultra-realistic practice rounds that only exist on a server, learns what succeeds, and steps onto the factory floor already knowing how to stack boxes or sort parts.
>“The age of generalist robotics has arrived, with breakthroughs in mechatronics, physical AI and embedded computing — just in time, as labor shortages limit worldwide industrial growth,” a promotional video during NVIDIA’s Computex presentation said. “A major challenge for robot makers is the lack of large-scale, real and synthetic data to train models. Human demonstrations aren’t scalable; limited by the number of hours in a day.”
>That stands in stark contrast to China’s humanoids, which still gorge on real-world footage. Unitree, a Shanghai-based firm, furnished its H-series robots with an open-source trove of full-body motion capture data so the machines could nail a folk dance routine at this year’s Spring Festival Gala. AgiBot goes even further: the Shanghai startup runs its sandwich-making, shirt-folding prototypes in 17-hour warehouse shifts, records every step, then feeds footage into DeepSeek overnight so the robots can relearn their chores the next morning — an approach Beijing is backing with billions in subsidies. Neither NVIDIA, Unitree nor AgiBot responded to the Daily Caller News Foundation’s requests for comment.
>By eliminating the training bottleneck, NVIDIA could speed past its hardware-rich Chinese competitors and set the de-facto operating standard for humanoids — a market Goldman Sachs pegs at more than $38 billion by 2035.
[Archive](https://archive.today/0TZVO)
You are currently inside a comment thread.
Click here to see all the comments (3).