NVIDIA Research introduces robot foundation model GR00T 2, which learns actions from human videos

NVIDIA Research said GR00T 2 achieved a 40% zero-shot success rate on 50 tasks the target robot had not encountered. With five demonstrations per task, the rate rose to 95% after 30 minutes of fine-tuning on a single RTX 5090. GR00T 2 is a foundation model that predicts future video and robot actions simultaneously, and is designed for use with parallel grippers, dexterous hands and humanoid robots.

In NVIDIA Research's evaluation, GR00T 2 succeeded on 40% of 50 tasks without demonstrations on the target robot, and achieved a 95% success rate with five demonstrations and fine-tuning. For adopters, a key question is how far its design—which transfers skills from human videos to robots—can reduce the need for robot demonstrations for each task.

GR00T 2's defining feature is that it predicts not only robot actions but also the future video resulting from those actions. NVIDIA Research says it uses a diffusion transformer to generate video and actions, then converts them into motor commands using a small head for each robot. Its use of videos of people at work to test whether skills can transfer without task demonstrations on the target robot sets it apart from approaches trained only on operational data from individual robots. The pretraining data is led by 402,730 hours of first-person video of people, followed by 23,500 hours of robot teleoperation data and 10,184 hours of demonstrations with UMI grippers. This mix puts human videos at the center of the training material while combining them with robot-specific body and operation data. However, task progress in video and success rates on physical robots are separate measures. The increase from 40% to 95% across 50 tasks also depends on five demonstrations and 30 minutes of fine-tuning. It does not mean the same performance will be achieved across physical robots in general. The model is not limited to a single robot type. Its training data includes arms such as Franka and YAM, hands from Sharpa and Wuji, and the Unitree G1; examples of tasks on other robots, including Dexter and Tianji, are also shown. GR00T 2 Turbo runs on a Jetson Thor aboard the G1 and is said to take 124 milliseconds per prediction. This points to a structure combining a shared model, conversion of commands for each robot and onboard execution. However, the reported success rates alone do not show the burden of robot-specific adjustments or how well performance holds across different hands and sensors. Compared with Hugging Face releases robot learning platform LeRobot v0.7.0, adding interactive control and large-scale data training, which this newspaper published on November 5, 2026, the two reports highlight different issues in the foundations of robot learning: software platforms for handling data and training functions, and mechanisms for transferring skills from a specific model to different robots. The materials do not establish whether the two are compatible or can be used together. Further tests should examine how the reported number of trials and success rates change across tasks and robot platforms, and how human videos used in training affect failures and safety on physical robots. The materials identify 100,000 hours of public data, but do not provide details on the terms for using GR00T 2's full weights or training data, or on the adjustments needed to deploy it on each robot. Adoption costs and generalizability should not be judged by the small number of demonstrations alone; reproducibility across robot platforms and operating conditions should be assessed separately.

The pretraining data consists of 402,730 hours of first-person video of people at work, 10,184 hours of demonstrations across 2,897 tasks using handheld UMI grippers, and 23,500 hours of robot teleoperation data. The robot data includes YAM, Franka, hands from Sharpa and Wuji, and the Unitree G1. Of the human video data, 100,000 hours are available on Hugging Face as Human-1M-100K. The model uses a diffusion transformer to generate future video and actions simultaneously. A small head for each robot converts the plan into motor commands. GR00T 2 Turbo is said to take 124 milliseconds per prediction on a Jetson Thor aboard the Unitree G1. In NVIDIA's evaluation, Turbo completed all 48 trials of scooping and serving popcorn through SONIC whole-body control, and succeeded in 93% of trials when clearing a table with multiple objects.