NVIDIA Robotics announced GR00T 2 at CoRL 2026, and NVIDIA GEAR lab lead Yuke Zhu posted the technical details. GR00T 2 is a new generation of world action model (WAM) built on Cosmos: one diffusion transformer denoises future video and an action chunk together, and one model drives parallel grippers, dexterous hands and humanoids. Pretraining mixes 402,730 hours of egocentric human video, 10,184 hours of handheld UMI gripper demos and 23,500 hours of robot teleoperation. By NVIDIA's numbers, it reaches 40% zero-shot success on 50 tasks the target robot never practiced, rising to 95% with five demos per task (under 3 minutes of robot data) and a 30-minute LoRA fine-tune on one RTX 5090. Reasoning models can call GR00T 2 as a tool through the WAM API. Code arrives in December; 100,000 hours of the human video data are already on Hugging Face.
Key Takeaways
- ✓Architecture: one diffusion transformer predicts future video and action chunks together; small per-robot heads let one model cover grippers, dexterous hands and humanoids
- ✓Data: 402,730 h of egocentric human video + 10,184 h of UMI demos (2,897 tasks) + 23,500 h of teleoperation; going from 0% to 100% of the human video lifts real-robot task progress from 50% to 70% (NVIDIA figures)
- ✓Adaptation: 40% zero-shot on 50 unseen tasks, 71% with 1 demo, 95% with 5 demos, about a 30-minute LoRA fine-tune on one RTX 5090
- ✓Agents: a reasoning model calling GR00T 2 via the WAM API completes 97+ of 100 back-to-back assemblies for each of three toys; marking grasp points on the camera view lifts instruction following from 40% to 92%
- ✓GR00T 2 Turbo is a 2B-parameter variant running at 124 ms per prediction on the Unitree G1's Jetson Thor; code ships in December, and 100k hours of human video are already on Hugging Face
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the industry pain point
Robot foundation models are bottlenecked by data: teleoperation is slow and expensive, and switching robots often means collecting again. GR00T 2, announced by NVIDIA at CoRL 2026, first learns to predict how the world will change, which lets it use large amounts of robot-free human video.
Architecture and mechanism
Built on Cosmos, GR00T 2 is a world action model (WAM). It reads recent video, the instruction and robot state, and one diffusion transformer denoises future frames and an action chunk together; it executes and repeats. Each robot needs only a small action head. Pretraining uses 402,730 h of egocentric human video, 10,184 h of UMI demos and 23,500 h of teleoperation. A 2B-parameter GR00T 2 Turbo runs in real time on a humanoid.
Benchmarks and measured results
All figures are NVIDIA's own and not yet independently reproduced: 40% zero-shot on 50 tasks the target robot never practiced, 95% after five demos; 65% on filter-paper setup learned from human demos only, with zero teleoperation; real-world RL lifts three insertion tasks to 88-96%; Turbo takes 124 ms per prediction on Jetson Thor and served popcorn in 48 of 48 trials.
Getting started for developers
NVIDIA says code arrives in December; parameter count of the main model and license are not disclosed. For now, browse the Human-1M-100K dataset on Hugging Face, review the open data-collection hardware, and watch the WAM API, which lets reasoning agents plan and verify while GR00T 2 handles manipulation.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.