Nvidia unveils GR00T 2 to teach robots delicate tasks

NVIDIA ↘ 4.30% used the opening day of the Conference on Robot Learning (CoRL) in Austin, Texas, on Monday, November 9, to introduce GR00T 2, an open "world action model" for robot manipulation, along with new simulation tools and benchmarks, the company said. UC Berkeley researchers working with NVIDIA also presented "agentic robotics," an approach in which AI agents write robot-control programs, test them in simulation and revise them when they fail, according to the university.blogs.nvidia.com+1

The announcements reflect a broader move in the field from teaching robots to walk to teaching them delicate tasks such as connecting cables and assembling small parts. The conference runs through November 12.nvidia.com

Agents that write robot code

Berkeley's Graph-as-Policy framework splits a robot task into separate skills, such as finding an object, planning a grasp and moving without collisions. AI agents then link these skills into a program and rehearse it across task variations using NVIDIA's Isaac simulation libraries, Berkeley said. In one test, OpenAI's Astra model built a physics simulation from a short video of a robot wiping ink off a metal bar with a sponge, then produced control code that worked on a real robot.techxplore.com

"Agentic robotics provides a missing link between model-based and model-free methods. It's a way forward that doesn't require any additional data," said Ken Goldberg, the Berkeley roboticist leading the work. Berkeley-based Ambi Robotics used the methods to update its parcel-sorting robots, adding capacity for more than 15,000 packages per robot each year, and has deployed the change across 30% of its U.S. fleet, the university reported. Berkeley and NVIDIA also released VABAR, a benchmark covering six industrial tasks with standardized variations and difficulty tiers, NVIDIA said.blogs.nvidia.com+1

GR00T 2 and simulation tools

GR00T 2 is built on NVIDIA Cosmos and was trained partly on 100,000 hours of open-source human egocentric video. According to NVIDIA, it averaged 71% success across 50 tasks with one demonstration per task, and 95% with five. Foxconn is applying it to server manufacturing, and Siemens is evaluating it for industrial uses.blogs.nvidia.com

Isaac Lab 3.0, which became generally available at GTC Berlin in October, lets agents set up training environments. With the Newton physics engine, it delivered up to 2.5 times the training throughput of Isaac Lab 2.3 on a vision-based dexterity benchmark, NVIDIA said. The company also highlighted DexBench, a dexterity benchmark built with RLWRLD, and noted that two NVIDIA papers are among six finalists for CoRL's Best Paper Awards.investor.nvidia.com

Limits remain

Goldberg said the gains are clearest in repetitive industrial work and cautioned that household robots that fold laundry are still some way off. Even so, he is enthusiastic. "To me, agentic robotics is the most exciting technological development since the internet in 1995," he said in NVIDIA's announcement.blogs.nvidia.com