A robot can watch a person complete a task and still fail when asked to do it with a different hand, camera or body. NVIDIA's GR00T 2 is designed to narrow that gap. Announced at the Conference on Robot Learning (CoRL) on November 9, it is a robot foundation model that predicts what the scene will look like next at the same time as it predicts the robot's actions. NVIDIA calls this a world-action model. The research is notable for trying to connect human video, demonstrations from different robots and physical action inside one learning system.
What makes GR00T 2 a world-action model?
Many robot policies map a camera view and instruction to an action. GR00T 2 adds a prediction of how the observed world should change as the action happens. Its diffusion transformer predicts future video frames and an action chunk together; the robot executes that chunk, receives new observations and repeats. NVIDIA's paper describes a 0.8-second visual history, robot-state input and a language instruction as inputs. The future frames are model predictions, not a literal replay or a promise that the physical robot will follow them exactly.
The shared prediction objective gives the system a way to learn from video that has no robot commands attached. A person reaching for a tool shows how the scene changes during a task, even though the footage does not say which motor values a robot should use. Separate robot-specific action heads translate the learned plan into commands for different bodies. That is the proposed bridge between seeing a skill and physically performing it.
What data does it learn from?
NVIDIA reports a pretraining mixture of 402,730 hours of first-person human work video, 10,184 hours of handheld UMI-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation. The robot data spans parallel-gripper arms, dexterous hands and a Unitree G1 humanoid. NVIDIA says a 100,000-hour subset of its Human-1M footage is openly available under CC BY 4.0; that data release is separate from access to the GR00T 2 model itself.
This mixture is intended to help skills cross what NVIDIA calls embodiments: different physical forms and sensors. In one direction, the model can learn a task from human video and transfer it to a robot that has not practiced that task. In another, it can learn from a handheld gripper and perform the work with a robot arm whose gripper is different. The target robot still contributes task-agnostic movement data so the model can learn how that body looks and moves. Transfer does not mean the robot needs no calibration, robot data or adaptation at all.
What do NVIDIA's transfer results show?
NVIDIA reports testing a dexterous robot on 50 tasks that were present in human video or other robots' data but not in that target robot's task demonstrations. Without task-specific demonstrations on the robot, GR00T 2 averaged 40% success across trials; it succeeded at least once on 40 of the 50 tasks. One demonstration per task raised the reported average to 71%. Five demonstrations, taking less than three minutes of robot data per task, raised it to 95% after a 30-minute LoRA fine-tune on one RTX 5090. NVIDIA says 37 of the 50 tasks then succeeded in every trial.
These figures describe a particular evaluation, not a universal measure of robot intelligence. The 40% average across trials and the 40-of-50 result answer different questions: the former counts repeated attempts, while the latter counts tasks with at least one success. The five-demonstration figure is also after task-specific fine-tuning. It is evidence that this system adapted quickly in NVIDIA's test setting, not evidence that a buyer can point any robot at an arbitrary job and expect a 95% success rate.
Where do agents and humanoids fit?
GR00T 2 handles contact-rich movements, while a higher-level reasoning model can choose and sequence tasks, check progress and call recovery tools. NVIDIA reports an agent-connected YAM robot completing at least 97 of 100 assembly trials for each of three toys, with the agent able to recover from stalls and errors. This is a system result: the policy, robot, task setup and supporting agent all contribute. It should not be attributed to the foundation model acting alone.
A smaller variant, GR00T 2 Turbo, is paired with NVIDIA's SONIC whole-body controller on a Unitree G1 humanoid. NVIDIA reports 124-millisecond prediction latency on the robot's Jetson Thor and all 48 popcorn-serving trials completed in its reported test. The layered design matters: Turbo proposes short-horizon motion, SONIC converts it into joint-level control, and a reasoning model can choose the task. That is different from asking one language model to directly control every motor.
What remains uncertain for developers?
The results and demonstrations come from NVIDIA's research program and white paper. They are not an independent replication across arbitrary work sites, robot vendors or safety conditions. NVIDIA's paper describes its own benchmarks and notes that its real-robot ratings are suite-specific and based on a limited number of distinct tasks. Its open-problem discussion also calls out safe physical action: language-model safety does not automatically transfer to a robot policy, whose movements need checks at the speed they execute.
Availability needs the same care. NVIDIA calls GR00T 2 an open world-action model and has published the research paper, but its research page says code will be available in December. As of November 14, that is a future availability statement; it does not establish that model weights, code or a production API can already be downloaded and deployed. Teams should check the actual release, license, supported hardware and safety documentation before planning a build around it.
For software makers, the broader signal is that the agent stack is moving beyond screens. A high-level agent may plan and verify, while a specialized model handles the physical interaction and a controller keeps movement stable. The hard engineering questions are still familiar: what each component can access, how it detects a failed step, when it stops, and what evidence supports a claimed success rate. Robotics demos show a direction; dependable products will need repeatable evaluation and safeguards around the complete system.
NVIDIA Research: model architecture, data mixture, transfer experiments and stated code timing
NVIDIA white paper: evaluation design, system integration and open problems
NVIDIA's November 9 CoRL announcement
Human-1M open dataset subset and CC BY 4.0 terms
How agent skills can carry workflows across AI clients
How do developers self-host Claude Managed Agents?
VIBPS is the social network for software, where makers can publish apps and build an audience around what they create. Robotics is another software frontier: its progress depends not only on models, but on the people building, testing and sharing tools around them.