NVIDIA introduced GR00T 2, an open frontier world action model for robots, at the Conference on Robot Learning (CoRL) in Austin on November 9, 2026 — a model trained largely on human video that completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations, and averaged 71% success after just one demonstration per task and 95% after five. Why it matters: if robots can pick up delicate manipulation skills from a handful of demonstrations instead of months of teleoperation, the cost of teaching each new factory or warehouse task falls sharply.
November 9, 2026 — Austin, Texas.
What NVIDIA announced
GR00T 2 is a 2-billion-parameter transformer built on NVIDIA’s Cosmos-Predict2.5-2B. It takes in recent camera views, the robot’s own state and a written instruction, then generates expected future views and the next action sequence. A single pretrained checkpoint covers parallel-gripper arms, dexterous hands and a humanoid body. NVIDIA describes the model as open, though its research page says the code will be available in December.
The training mix leans heavily on people rather than robots: 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation. NVIDIA is also releasing 100,000 hours of that human data — video, language annotations and head and hand motion — under a CC BY 4.0 license that permits commercial use.
The results — and the caveat in the fine print
On cross-embodiment tests, GR00T 2 reached 65% success on a filter-paper setup learned from human demonstrations without task-specific teleoperation, and 60% on a toy dump-truck assembly transferred from a handheld gripper to a YAM arm. The headline numbers are the few-demonstration scores: 71% average success across 50 tasks with one demonstration each, rising to 95% with five — where five demonstrations amount to under three minutes of robot data per task, followed by roughly a 30-minute adapter fine-tune on a single RTX 5090.
Read the whitepaper before quoting that 95%, though: NVIDIA’s own technical paper notes the evaluated tasks were selected from among those the model can learn with five episodes at a success rate of at least 80%, so the average is high partly by construction. The honest summary is that GR00T 2 adapts unusually fast on tasks it is already close to being able to do — which is still a meaningful step, but not the same as 95% on arbitrary work.
NVIDIA says the model currently ranks among the top three across all vision-language-action and world-action-model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins. A reasoning agent can sequence its skills, monitor progress and invoke recovery routines when parts drop or work stalls.
Simulation is the other half of the story
Alongside the model, NVIDIA spotlighted Isaac Lab 3.0 — generally available since October 20 — whose tighter integration with the open-source Newton physics engine lets it model deformable objects such as cables and cloth, and which NVIDIA says delivers up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX in a vision-based dexterous-manipulation benchmark. The Isaac Lab-Arena 0.4 developer preview arrives at CoRL, together with two new open benchmarks: DexBench, from RLWRLD and NVIDIA, with 15 tasks across retail, logistics, assembly and hospitality scenes, and UC Berkeley and NVIDIA’s VABAR for industrial tasks.
Why it matters
Nexuni is testing GR00T 2 for laundry handling, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating industrial use cases — testing and evaluation, with no production rollout announced. That gap is the story to watch. The model, the human-video dataset and the simulation stack are all open or opening up, which puts pressure on every closed robotics stack. But as NVIDIA’s own results show, reliability still depends on the full workflow around the model — corrective demonstrations, reinforcement post-training and an agent supervising recovery — and on how much supervision time each task consumes. Everything AI, tested: the December code release is when independent teams can start checking these numbers for themselves.
Frequently asked questions
What is NVIDIA GR00T 2?
GR00T 2 is NVIDIA’s open frontier world action model for robots, introduced November 9, 2026 at CoRL in Austin. Built on Cosmos, it predicts future views and action sequences so robots can learn dexterous manipulation skills from human video and small amounts of robot data.
How well does GR00T 2 actually perform?
NVIDIA reports it completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations, averaged 71% success with one demonstration per task and 95% with five. NVIDIA’s whitepaper notes the tasks were pre-selected from those the model can learn to at least 80% with five episodes, so the 95% figure is partly a result of that selection.
When can developers use GR00T 2?
NVIDIA’s research page says code will be available in December 2026. The accompanying 100,000-hour human egocentric dataset is being released under CC BY 4.0, permitting commercial and noncommercial use.
Who is using GR00T 2 so far?
NVIDIA says Nexuni is testing it for laundry handling, Foxconn is applying it to server-manufacturing automation and Siemens is evaluating industrial use cases. All disclosed work is testing and evaluation; no production rollout has been announced.
Sources: NVIDIA Blog, Unite.AI, TipRanks