Nvidia has introduced GR00T 2, an open artificial intelligence model designed to equip robots with the ability to execute intricate hand movements. The system draws training data from robot interactions and extensive human video material, specifically 100,000 hours of first-person perspective footage that Nvidia is making publicly available. During internal testing, the model achieved a 95% success rate when robots were shown a task five times before attempting it. The company made the announcement at the Conference on Robot Learning held in Austin, Texas on Monday.
Nvidia describes GR00T 2 as an open "world action model" that forecasts how its movements will alter its visual perception before executing those movements. The training approach enables robots to absorb techniques from human demonstrations and transfer learned abilities across robots with different physical configurations, according to Nvidia. Companies like Maxinsights have adopted comparable strategies, compensating individuals to record everyday activities for machine learning purposes.
Testing results showed that GR00T 2 completed 40 out of 50 tasks at least once without any prior training on those specific tasks. When presented with a single demonstration, the model succeeded 71% of the time on average. Nvidia's published research acknowledges that the tested tasks were selected from a pool of activities the model was already capable of learning.
Three major manufacturers are currently evaluating GR00T 2 in real-world scenarios. Foxconn, which derives the majority of its revenue from AI server manufacturing, is implementing the model to support server assembly operations. Siemens is conducting trials for manufacturing environments, while Nexuni is exploring applications in laundry operations. None of these organizations have announced plans to deploy the technology in production systems at this time.
Alongside GR00T 2, Nvidia released Isaac Lab 3.0, an open-source platform for simulating robot training environments. The company reports that the tool achieves training speeds up to 2.5 times faster based on one benchmark test. Walden Robotics, which secured $300 million in funding in July, is among the partners utilizing the platform.
To me, agentic robotics is the most exciting technological development since the internet in 1995
Ken Goldberg, robotics professor at UC Berkeley who collaborated with Nvidia on two of the projects
Earlier iterations of the GR00T model are already integrated into robots currently in development. LG has a humanoid robot scheduled for release in 2027 that incorporates GR00T technology. Siemens previously demonstrated an Nvidia-powered humanoid robot completing an eight-hour factory shift.
Source: The Next Web
