NVIDIA’s GR00T 2 transfers robot skills, but reliability needs more than a model

• GR00T 2 combines human video, handheld-gripper demonstrations and robot data to transfer skills between bodies.
• NVIDIA’s 95% average success with five demonstrations applies to tasks selected for being learnable with that amount of data.
• Recovery agents improve assembly reliability, while industrial applications remain in testing or evaluation and code is scheduled for December.

NVIDIA, the accelerated-computing company behind the Isaac robot-learning tools, has introduced GR00T 2, a foundation model that predicts future video and robot actions together. Its central promise is to reduce the need to demonstrate each task on the particular robot that will perform it.

Announced at the Conference on Robot Learning in Austin, Texas, the model transfers some skills from human demonstrations or handheld devices to robots that have not practiced those tasks. But NVIDIA’s results also show why that is only one step toward industrial automation: success varies by task, and its most reliable assembly workflows add corrective demonstrations and a separate agent that detects and repairs failures.

Head- and wrist-camera observations from a GR00T 2 truck-assembly rollout. Source: NVIDIA Research. Source: NVIDIA Research.

What changes from NVIDIA’s earlier robot models

NVIDIA calls GR00T 2 a world-action model. It jointly predicts what cameras will see and what the robot should do; small components specific to each robot translate the shared model’s output into motor commands. The design supports gripper arms, five-fingered hands and humanoids, according to its research description.

It builds on Cosmos, NVIDIA’s platform of generative world models and video-processing tools for developing robots and autonomous vehicles. When NVIDIA introduced Cosmos in January 2025, it described models that generate physics-based synthetic data for training and evaluation. GR00T 2 couples prediction of the physical scene with prediction of actions.

Nor is combining high-level reasoning with robot movement new to this program. NVIDIA introduced GR00T N1 in March 2025 as a customizable humanoid foundation model with a vision-language reasoning system and a separate action system. A post-trained version powered a domestic-tidying demonstration on humanoid developer 1X’s robot. GR00T 2’s case for progress rests on joint video-and-action prediction, transfer across bodies and measured adaptation gains, rather than the first use of reasoning to direct movement.

The data mixture is larger than the public release alone suggests. NVIDIA lists a pretraining mixture of 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. It is releasing 100,000 hours of the human footage as Human-1M-100K; that subset is not the full human-video pretraining corpus.

In an experiment varying the amount of human video, NVIDIA reports that real-robot task progress across 72 downstream dexterous-robot tasks rose from 50% without it to 70% with the full human-video mixture. That is a 20-percentage-point gain in progress through tasks, not a 70% rate of completing entire jobs.

Company-reported adaptation results from NVIDIA on 50 selected tasks. Tasks qualified by reaching at least 80% success with five demonstrations. Source: NVIDIA Research.

Transfer succeeds unevenly, and the 95% result has a filter

NVIDIA’s human-to-robot experiment used demonstrations by people wearing cameras, plus task-agnostic data to teach the target robot about its own body. With no task-specific teleoperation, a two-handed robot succeeded at filter-paper setup in 65% of 20 trials. Across the other three tasks, success ranged from 10% for microscope-slide dabbing to 50% for conveyor-package tossing. Human demonstrations transferred useful behavior, but precision remained uneven.

In a separate test, demonstrations collected with a handheld gripper transferred to a YAM robot arm, yielding 60% end-to-end success on toy dump-truck assembly. YAM supplied task-agnostic teleoperation data, but no demonstrations of that task. These experiments test transfer without target-robot task demonstrations, not learning without robot data.

For a 50-task evaluation on a bimanual robot with two Sharpa hands, NVIDIA reports 40% average trial success without target-robot task demonstrations. The model succeeded at least once on 40 of those tasks—a different measure from the average success rate. One demonstration per task raised average success to 71%; five raised it to 95%. Each adaptation used a 30-minute fine-tune of small adapter weights on one RTX 5090 GPU, with the five demonstrations supplying under three minutes of robot data per task.

The selection rule limits that headline figure. A report examining NVIDIA’s paper notes that the tasks were chosen from those the model could already learn well. A technical account of the same whitepaper specifies that selected tasks reached at least 80% success with five episodes. NVIDIA’s 95% average therefore describes adaptation within that selected set, rather than an expected success rate on arbitrary new work.

There is competitive precedent for transferring knowledge between robots. Physical Intelligence, the developer of the π0 family of vision-language-action models that turn visual and language inputs into robot actions, introduced π0.5 in April 2025. Its recipe combines data from different robot bodies and environments with multimodal web data. The company demonstrated cleaning tasks in previously unseen homes, while explicitly distinguishing that goal from learning new skills or achieving high dexterity.

NVIDIA’s own comparisons report 60% success for GR00T 2 against zero for π0.5 on handheld-gripper-to-YAM dump-truck transfer. That supports a specific transfer advantage in NVIDIA’s test, not a universal ranking across the different capabilities the two developers emphasize.

Company-reported NVIDIA assembly results with supervised fine-tuning, added corrective data, and corrective data plus an agent. The final condition measures the coordinated system. Source: NVIDIA Research.

Reliability comes from correction and recovery

NVIDIA’s continuous assembly tests show how much the surrounding system changes outcomes. Each trial began with toy parts dumped onto a table in a cluttered layout absent from post-training data.

In these company tests, the reasoning agent monitored progress and called recovery routines when parts dropped or work stalled. GR00T 2 performed most manipulation. The final column measures that coordinated system, rather than the pretrained model alone.

Precision also benefited from a distinct post-training phase. NVIDIA reports that reinforcement learning on the real robot raised success on three contact-rich insertion tasks from 40–76% after supervised post-training to 88–96%. Those gains required additional learning on physical hardware.

Agent guidance has limits and overhead. Marking grasp and release targets on camera views helped YAM complete 17 of 20 trash-sorting episodes, against none with step instructions alone. Yet the same approach completed only two of 20 trash-sorting episodes on the dexterous robot and no full board-disassembly episodes there. The technical account of NVIDIA’s whitepaper also reports that the robot waited for the agent during a median 38% of an episode in visually steered trash sorting.

NVIDIA is also promoting alternatives for coordinating skills. Its CoRL announcement describes UC Berkeley’s Graph-as-Policy framework, which uses agents to assemble robot programs, rehearse them in simulation and revise them after failures. Reliable orchestration is a broader development effort, not a capability supplied solely by GR00T 2.

Computing and simulation shape the deployment tradeoff

GR00T 2 Turbo, a smaller variant with 2 billion parameters, addresses onboard computing. NVIDIA reports a median prediction latency of 124 milliseconds on the Unitree G1 humanoid’s Jetson Thor computer. Turbo plans chunks of motion; a separate whole-body controller executes them. Everything except the reasoning model runs onboard.

Speed does not preserve performance uniformly. Turbo trails the larger model on five of six disclosed dexterous-hand post-training tasks, tying on the sixth, while beating it on two of six YAM tasks. On the humanoid, Turbo completed all 48 popcorn-serving trials but only 14 of 30 spill-cleanup trials. Onboard operation and reliable execution across tasks remain separate tests.

NVIDIA’s commercial stake spans the workflow: its GPUs support adaptation and inference, while its Isaac software supports simulation and evaluation. Isaac Lab 3.0 is an open-source robot-learning framework; integration with the Newton physics engine adds support for deformable objects such as cables and cloth. NVIDIA says in its simulation announcement that this combination achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX on a vision-based dexterous-manipulation benchmark. That comparison does not establish a reduction in total deployment cost.

Simulation also supplies a way to screen changes before using physical robots. NVIDIA reports a 0.92 Pearson correlation between simulated and real-world matched-task success rates for 14 GR00T 2 policies on a 25-task YAM benchmark. It used simulated evaluations to screen data-processing changes and inference settings. The result supports simulation as a development aid within that benchmark; it does not establish reliability across arbitrary industrial conditions.

The next test is sustained production work

NVIDIA says in its announcement that Nexuni is testing GR00T 2 for laundry handling, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating industrial uses. The disclosed applications do not establish a production rollout. Although the announcement calls GR00T 2 open, its research page schedules code availability for December.

The immediate questions are what developers can reproduce when code arrives and how much adaptation, recovery and supervision industrial users need. The disclosed trial success rates and single-GPU fine-tuning time do not establish cost per successful job, sustained throughput or human intervention across a factory shift. Those measurements would determine whether transferring skills from people lowers the full cost of deploying robots, beyond reducing the demonstrations needed to start teaching them.