NVIDIA introduces GR00T 2, an open robot model built on Cosmos, at CoRL

NVIDIA introduced GR00T 2, a robot model it calls open, on 9 November 2026 at the Conference on Robot Learning (CoRL), held in Austin, Texas, together with Isaac Lab 3.0, a preview of Isaac Lab-Arena 0.4 and two new open benchmarks.

NVIDIA describes GR00T 2 as a frontier model of the world-action type, built on its Cosmos platform. It says training drew on both robot and human recordings, among them 100,000 hours of NVIDIA's openly released first-person human video, and that the model teaches robots dexterous manipulation. NVIDIA's research page says the model forecasts upcoming video together with the robot's moves, and lists its code as arriving in December.

NVIDIA reports that GR00T 2 succeeded at least once on 40 of the 50 tasks tested, without any demonstrations for those tasks. With one demonstration per task it averaged 71% success across the 50 tasks, and with five it reached 95%. The research page adds that the model succeeded in 40% of zero-shot trials, and that each adaptation is a half-hour fine-tune on one RTX 5090 graphics card. NVIDIA also says the model can learn from human demonstrations and transfer skills between robot bodies, and that AI agents can drive it through its programming interface, chaining skills and recovering after mistakes.

NVIDIA states that GR00T 2 currently places among the best three models on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, among all world-action and vision-language-action models, and has won several leaderboards. It names Nexuni as testing the model for laundry handling, Foxconn as using it in factory automation where servers are made, and Siemens as evaluating it across industrial uses.

Simulation and benchmarks

Isaac Lab 3.0 is the latest release of NVIDIA's open-source framework for training robots. NVIDIA says it lets researchers use agents to set up training environments, choose between the PhysX and Newton solvers, and use a new library for reinforcement-learning refinement of models such as GR00T. NVIDIA says tighter links to Newton, an open-source physics engine, let Isaac Lab simulate soft items like cables and cloth, and robot hands with many linked joints. On a benchmark of camera-guided dexterous manipulation, NVIDIA reports, Isaac Lab 3.0 running Newton reached up to 2.5 times the training throughput of version 2.3 running PhysX.

NVIDIA says a developer preview of Isaac Lab-Arena 0.4, its open-source framework for running policy evaluations at scale, is arriving at the conference. DexBench and VABAR are two new open benchmarks built with it and Newton. RLWRLD developed DexBench with NVIDIA; NVIDIA says it has 15 tasks set in four scenes: hospitality, retail, logistics and manufacturing. UC Berkeley and NVIDIA developed VABAR, which NVIDIA says covers six industrial tasks, each with set variations, tiers of difficulty and success measures.