NVIDIA introduced GR00T 2 on November 9. The model predicts future video and robot actions together, executing action chunks in a repeated control loop.
The big change
- The change: Prediction informs motion.
- Why it matters: Human demonstrations can provide examples without repeating every task on a target robot.
- What to watch: Physical success requires testing with real equipment and surroundings.
Learning from demonstrations
Robot manipulation involves choosing a movement and coping with how objects respond. NVIDIA’s earlier DreamZero paper studied joint video and action prediction and transfer from humans or other robots. In that earlier study, adding human video raised average progress on unseen tasks from 38.3% to 54.3%; the researchers describe it as an early signal, with moderate success rates. A demonstration can help describe a skill without providing a trajectory for the target robot’s own joints. Different hands and surroundings can still make execution difficult: a plausible predicted scene does not prove that a real grasp will succeed.
What the test found
The project page reports 40% average zero-shot success across 50 tasks seen in human or other-robot data but absent from the target two-handed robot’s task data. Five demonstrations per task, each set under three minutes, raised the reported average to 95% after a 30-minute LoRA fine-tune per task on one RTX 5090. Without those demonstrations, 40 of the 50 tasks succeeded at least once.
The company-reported physical-robot test is not independent validation or evidence of routine commercial deployment.
What developers can use
NVIDIA plans code for December; current GR00T 2 weight availability is not established on the project page. The existing Isaac GR00T repository provides code and weights for the earlier N1.7 model. That is a separate release.
