GR00T 2: Nvidia’s robot brain runs on-board, and its weights and 100K hours of data are open

Jim Fan, a senior researcher at Nvidia, introduced GR00T 2 on X, describing it as the company’s most advanced open “System One” model for robotics. It runs fully on a robot’s onboard compute, and Nvidia is releasing the weights.

A release that goes well beyond model weights

According to Fan’s post, GR00T 2 is built around three features. The first is long-horizon work: dexterous manipulation and whole-body control chained over more than 20 steps and more than five minutes. Fan’s examples are running chemistry lab experiments and assembling model airplanes from start to finish.

The second is a technique called X-Adapt. Nvidia says it lets the model learn from cheap data sources, namely human wearables and simulation, and then deploy zero-shot to a range of real robots. The third, Unified Cognition, is described as native integration with frontier agents for physical reasoning, error recovery and safety monitoring.

Fan called the release “an open-source buffet”. Besides the weights, Nvidia says it is publishing its agentic framework, simulation engine, annotation engine, a hardware design and training data totaling 100,000 hours, which Fan equates to roughly 11 years of experience. All of this is available today, with one exception: the X-Adapt recipes are due in December, so the cross-robot transfer claim cannot be reproduced yet.

Fan closed the post with a line that sums up the pitch: “Everything that moves will be autonomous.” A longer thread with a deep dive accompanies the announcement.

Why running on the robot itself is the hard part

In the dual-system framing Nvidia has used since its first open humanoid model, System Two is the slow, deliberate reasoner and System One is the fast controller that turns perception into motion. GR00T N1, released in March 2025, paired a vision-language module with a diffusion transformer for actions. According to its paper, the public checkpoint had about 2.2 billion parameters, 1.34 billion of them in the vision-language part. GR00T 2 keeps the System One role, but the post does not say what size or architecture it has.

The untethered claim matters because a robot cannot wait on a network round trip while holding an object. Nvidia’s related VFlash vision model, covered here on November 5, was pitched at the same constraint. Nvidia reported 5.1x faster whole-robot inference on small on-device GPUs such as Jetson Thor, a board with 128 GB of memory and 273 GB/s of bandwidth, according to our earlier report. Those figures describe the hardware class Fan’s “entirely onboard” claim has to fit within. The GR00T 2 post names no target device.

Chaining 20-plus steps over five minutes is also a different problem from a short pick-and-place demo. Small errors compound, and a robot has to notice and recover from a failed grasp rather than continue blindly. That is where Nvidia says the agent integrations come in, though the post gives no success rates for any of the long tasks.

Training on human wearables and simulation instead of teleoperated robot data addresses the main bottleneck in the field, which is that real robot demonstrations are expensive to collect. If X-Adapt works as described, it would let a lab with one robot benefit from data gathered on other bodies. That is the part that cannot be judged until December.

Who can use it now, and what is still unclear

Anyone can start with what is already out: the model weights, simulation and annotation tools, the agentic framework, hardware design files and the dataset. Fan’s framing, “Anyone can become a roboticist”, points at labs, startups and hobbyists rather than only large robot makers. The post is the only source we could read, so details that would matter to a team deciding whether to build on it are missing.

• License terms for the weights, data and hardware design are not stated in the post.
• No benchmark numbers, model size or inference speed are given.
• The compute needed to run the model on a robot is not specified.
• The 20-step, five-minute tasks are Nvidia’s own description; no independent evaluation exists yet.

For context, this is a research lead’s announcement on X, and the thread with the deep dive was not available to us when we wrote this. The scale of the open data release, 100,000 hours, will be the first thing outside teams can verify. The cross-robot results will follow once the December recipes land.