Is GR00T 2 state of the art?Pending benchmarks.Announced Nov 9, 2026; no independent results yet.
NVIDIA GEAR's foundation world-action model (WAM) that jointly predicts future video and robot actions. Built on NVIDIA Cosmos; one family covers parallel grippers, dexterous hands and humanoids. Announced at CoRL with an NVIDIA blog and research page on Nov 9, 2026. NVIDIA's research page says code is available in December — weights are not published yet. GR00T 2 Turbo is the smaller on-robot variant (about 2B parameters) that runs on a Unitree G1's Jetson Thor.
The details so far
- Lab
- NVIDIA
- Announced
- Nov 9, 2026
- API
- Not yet · No public cloud API announced; WAM API described for agent control in lab demos
- Open weights
- Announced, not yet published · Not released. NVIDIA GEAR research page states Code Available in December
- Type
- Foundation world-action model: one diffusion transformer denoises future video and robot action chunks together; small per-robot heads map the plan to each embodiment
- Embodiments
- Parallel-gripper arms (YAM, Franka), dexterous hands (Sharpa, Wuji / XHAND), and Unitree G1 humanoid (NVIDIA research page)
- Base
- Built on NVIDIA Cosmos (NVIDIA CoRL blog)
- Pretraining data
- Mixture of about 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 (NVIDIA research page)
- Open human video
- 100,000 hours released as Human-1M-100K on Hugging Face (dataset; not the GR00T 2 weights)
- GR00T 2 Turbo
- Smaller joint video-action transformer (about 2B parameters) with a higher-compression video encoder; onboard Unitree G1 via Jetson Thor at 124 ms per prediction (38 ms on RTX 5090); steers SONIC whole-body control
- Agent control
- Reasoning agents can call GR00T 2 through a WAM API and steer grasps by marking camera views (lab demos)
- Related
- Separate from Flora, MotionBricks, Cosmos3-H-Embodied-Nano and earlier GR00T N1 / N1.5 / N1.6 naming
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on 50 tasks never seen on the target robot, GR00T 2 succeeds 40% of the time zero-shot; with five demonstrations per task (under 3 minutes of robot data) and a 30-minute LoRA fine-tune on one RTX 5090, average success rises to 95% (NVIDIA research page).
Lab claim: with minimal post-training on two 22-DoF Sharpa hands, seating GB300 memory modules reaches 85% and toy-truck assembly 65%; real-robot RL lifts three contact-rich insertion tasks to 88–96% success.
Lab claim: with a reasoning model calling the WAM API, YAM completes 97 or more of 100 back-to-back assemblies of each of three toys; visual steering reaches 92% what-and-where follow-through (up from 40% with the task instruction alone).
Lab claim: NVIDIA's CoRL blog reports GR00T 2 among the top three models across the vision-language-action and world-action categories it cites for RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, with multiple leaderboard wins on those boards.
Sources
LABPractice Makes Precise: CoRL robotics blog (NVIDIA)blogs.nvidia.com/blog/robotics-simulation-advancements-corl/ · Nov 9, 2026LABGR00T 2 research page (NVIDIA GEAR)research.nvidia.com/labs/gear/gr00t-2/ · Nov 9, 2026LABJim Fan on X: GR00T 2x.com/yukez/status/2119880324075377094 · Nov 9, 2026LABYunze Man on X: GR00T 2x.com/yunzeman/status/2119910338724298966 · Nov 9, 2026Tracking ends Feb 7, 2027 unless the API or weights become available.