GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation
NVIDIA's GR00T 2 is a robot foundation model that learns by predicting future video and actions together, so it can pick up skills shown only by humans or on other robots. The write-up backs this with success numbers across dexterous hands, arms, and a humanoid.
- Robots performed tasks they never practiced after seeing only human, handheld-gripper, or wearable-device demonstrations, e.g. a lab funnel-lining task at 65% success with no robot teleoperation.
- Zero-shot success on 50 unseen tasks was 40%, rising to 95% after five demonstrations and a 30-minute LoRA fine-tune on one RTX 5090.
- A smaller variant, GR00T 2 Turbo, runs on the Unitree G1 humanoid itself at 124 ms per prediction.
GR00T 2 is NVIDIA's new generation of what it calls foundation world-action models: a single diffusion transformer that denoises future video and robot actions together. The idea is that predicting how a scene should change, and not just which motor commands to send, lets the model learn physics from human video. Small per-robot output heads then turn that plan into motor commands for parallel-gripper arms (YAM, Franka), dexterous hands (Sharpa, Wuji), and the Unitree G1 humanoid. Pretraining mixes 402,730 hours of egocentric human video, 10,184 hours of handheld UMI gripper demos across 2,897 tasks, and 23,500 hours of robot teleoperation. The team says human video is the biggest lever: going from no human video to 100% lifts real-robot task progress on 72 Dexter tasks from 50% to 70%.
The headline capability is what they call cross-adaptation, or X-Adapt. A robot can learn a task it never performed itself, from a different embodiment. A two-hand robot with five-fingered hands filled a lab funnel with filter paper in 65% of trials after seeing only humans do it. A YAM arm assembled a toy dump truck end to end in 60% of trials after seeing only a handheld gripper demo, a task that runs up to four minutes. A two-handed Tianji robot with XHAND1 hands learned bimanual kitchen and lab tasks from DexUMI wearable-hand demos alone. On the adaptation side, GR00T 2 succeeds zero-shot 40% of the time on 50 unseen tasks. Five demos per task, under three minutes of robot data, and a 30-minute fine-tune on one RTX 5090 push that to 95%. Reinforcement learning on the real robot lifts three contact-rich insertion tasks to 88–96%.
The more practical-sounding pieces are agentic control and a faster variant. Because the model predicts the outcome of its actions, a reasoning model can direct it through an API: YAM completed 97 or more of 100 back-to-back assemblies of each of three toys, recovering from stalls along the way. When a reasoning model marked grasp and release points on the camera view, instruction-following rose from 40% to 92%. GR00T 2 Turbo is a smaller transformer built to run on the G1's own Jetson Thor at 124 ms per prediction, and paired with the SONIC whole-body controller it scooped and served popcorn in all 48 trials. Most of these results are reported by NVIDIA on its own hardware and tasks, so the trial counts are small and worth reading as company-run benchmarks rather than independent evaluations.