NVIDIA Unveils GR00T 2 World Action Model That Teaches Robots From a Single Demonstration

Roboticists have spent years chasing a stubborn gap: a machine that performs flawlessly in a lab demo often collapses when the lighting shifts, the object changes, or the task moves to a different arm. The usual fix is more data, and more data usually means more teleoperation, more labeled episodes and more engineering hours per task. That economics has kept general purpose robot learning confined to well funded labs and a handful of narrow factory cells.

NVIDIA has argued for three years that the missing ingredient is not hardware but a foundation model that can watch people, watch other robots, and then generalize. Its Isaac simulation stack handles the rehearsal, and a family of robot foundation models handles the transfer. At the Conference on Robot Learning in Austin, Texas, the company moved that argument forward on both fronts at once.

The centerpiece is GR00T 2, which NVIDIA describes as an open frontier world action model built on NVIDIA Cosmos. A world action model predicts future video frames and the robot actions that produce them in the same pass, which lets a policy reason about what a scene will look like after it moves. The company also shipped Isaac Lab 3.0 and two open industrial benchmarks, turning a research demo into something closer to a platform.

The headline number is not a benchmark record. It is a data efficiency claim, and it reframes how much human effort a new robot skill should cost.

Key Facts

NVIDIA published the announcement on November 9, 2026, under the headline Practice Makes Precise, framing the release as a study in how robots rehearse delicate work in simulation. The Robot Report reported on November 9, 2026, that the company introduced GR00T 2, its updated open frontier world action model, alongside Isaac Lab 3.0 at the Austin conference.

The evaluation is the part that matters. Across 50 tasks never seen on the target robot, GR00T 2 succeeds 40 percent of the time zero-shot and completes 40 of the 50 at least once with no task-specific demonstrations. One demonstration per task lifts average success to 71 percent, and five demonstrations, totaling under 3 minutes of robot data, lift it to 95 percent, with 37 of the 50 tasks succeeding in every trial. The adaptation runs as a 30 minute LoRA fine-tune on a single RTX 5090.

The pretraining mixture is unusually broad. NVIDIA Research's project page, published on November 9, 2026, lists 402,730 hours of egocentric video of people at work, 10,184 hours of demonstrations with a handheld UMI gripper across 2,897 tasks, and 23,500 hours of robot teleoperation spanning YAM and Franka gripper arms, Sharpa and Wuji dexterous hands, and the Unitree G1 humanoid. Of the human video, 100,000 hours are released openly as Human-1M-100K on Hugging Face.

NVIDIA said on November 9, 2026, that the model 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. The company listed four headline capabilities: out of the box performance, skill transfer between robot bodies without robot demonstrations, generalization with little data, and agent control, in which AI agents call the GR00T 2 API to sequence skills and recover from errors.

Isaac Lab 3.0 adds simulation ready assets, switchable PhysX and Newton physics solvers, a new reinforcement post-training library, and modelling of deformable objects such as cables and cloth. NVIDIA reports up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX in a vision based dexterous manipulation benchmark. The evaluation stack grew as well: an Isaac Lab-Arena 0.4 developer preview, DexBench from RLWRLD with 15 tasks across retail, logistics, manufacturing and hospitality scenes, and VABAR from UC Berkeley with six industrial tasks.

Analysis

What this really means is that the bottleneck in general purpose robotics is shifting from demonstration collection to demonstration quality and simulation fidelity. The five demonstration result is the one to watch: under 3 minutes of robot data plus a 30 minute fine-tune is a cost profile a mid sized integrator can absorb, not just a corporate research lab with a dedicated data team.

The partner roster shows where the demand is. The Robot Report reported on November 9, 2026, that the list includes Ambi Robotics, Bosch Research, Dexmate, Dyna Robotics, FieldAI, Flexion, Noitom Robotics, Siemens, Skild AI, RLWRLD, UC Berkeley, Walden Robotics and the University of Washington. Nexuni is testing GR00T 2 for robotic laundry, Foxconn is applying it to factory automation for server manufacturing, and Siemens is evaluating industrial use cases. MicroAGI is post-training the model with reinforcement learning to swap Ethernet cables on a YAM arm.

Those deployments map onto three different economic arguments. Laundry is a low margin, high variability task where value comes from uptime rather than cycle speed. Server manufacturing is a high mix, short cycle business where retooling cost dominates and a generalist policy saves engineering time on every variant. Industrial evaluation at Siemens is about whether a generalist model can survive qualification regimes written for deterministic machines, which is a harder test than any leaderboard.

The technical results also hint at where the model is strong and where it is not. Cross adaptation included filter paper setup in a chemistry lab at 65 percent success from human video with no teleoperation, end to end toy dump truck assembly on a YAM arm at 60 percent from a handheld UMI gripper, and 85 percent success seating GB300 memory modules with 22 degree of freedom Sharpa hands. A smaller variant, GR00T 2 Turbo, runs on the Unitree G1 onboard Jetson Thor at 124 milliseconds per prediction and, through the SONIC whole body controller, scooped and served popcorn in all 48 trials and cleared a table of objects in 93 percent of trials. Agentic control reached 97 or more of 100 back to back toy assemblies per item, and 92 percent adherence to what and where visual steering instructions versus 40 percent from task instructions alone.

Why It Matters

On the simulation side, the AlphaNIST result is the strongest argument that the stack is maturing. The Robot Report reported on November 9, 2026, that AlphaNIST's single policy reached more than 98 percent training success against 81 percent for per task policies while using about 39 times less GPU time. If one policy can replace a fleet of task specific ones, the cost of standing up a new robot cell falls in a way that changes who can afford automation at all.

Scaling evidence points the same direction. NVIDIA Research reported on November 9, 2026, that raising human video from 0 percent to 100 percent of the pretraining mixture increased real robot task progress on 72 tasks from 50 percent to 70 percent. That is a direct measurement of the value of human data, and it justifies the decision to release 100,000 hours of it openly rather than keeping it behind a commercial license.

Openness is also the strategic play. By releasing weights, data and benchmarks together, NVIDIA makes its simulation and compute stack the default for anyone building a robot foundation model. That position pays off through GPU demand long before it pays off through robot revenue, and it puts pressure on rivals to match a release cadence they did not set.

Next Up

The near term test is whether partners publish results from production settings rather than curated demos. Foxconn's server lines and Nexuni's laundry operations will be the first places where a 71 percent or 95 percent lab figure meets throughput targets, uptime requirements and the messy failure modes that benchmarks rarely capture.

NVIDIA has framed the release as a platform moment, with open weights, open data and open benchmarks arriving together, so the next data point will come from the community rather than the launch. The partner evaluations, the AlphaNIST follow ups and the first independent reproductions of the 95 percent claim are what the field will argue about next.