NVIDIA introduces GR00T 2 robot model

Sole source: NVIDIA. The company says it is introducing GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas. The company describes it as an open world action model built on NVIDIA Cosmos: a model that, it says, helps robots perceive their surroundings, anticipate the effects of actions and carry out multistep tasks.

In its announcement, NVIDIA said the model completed 40 of 50 evaluated tasks at least once without demonstrations specific to those tasks. That measures how many tasks it completed at least once, not its success rate on each attempt. With one demonstration per task, the model achieved 71% average success across 50 tasks, rising to 95% with five demonstrations per task, NVIDIA said.

The company did not specify the tasks, robot bodies, trial counts or success criteria behind the 50-task results, or say which evaluations used simulation and which used physical robots. NVIDIA also did not give a public release date or access, licensing, hardware or pricing terms for the model’s weights, code or application programming interface (API).

From demonstrations to robot skills

NVIDIA says GR00T 2 was trained on robot and human data, including 100,000 hours of its open-source first-person human data. It says the model can learn from human demonstrations and transfer skills between different robot bodies without robot demonstrations. Artificial intelligence (AI) agents can use the model’s API and supporting tools to run skills in sequence and recover from errors, the company says.

NVIDIA says Nexuni is testing the model for robotic laundry handling, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating industrial uses. MicroAGI is using a version further trained through reinforcement learning—a method that uses feedback to improve behavior—for Ethernet cable swapping with a YAM robotic arm. NVIDIA says MicroAGI automates the creation and calibration of simulation environments using Isaac Lab 3.0 and Newton, where it trains the robot’s policy, the system that chooses its actions, and plans to demonstrate the workflow at CoRL.