Nvidia unveils GR00T 2 world-action model, with models, data and simulation to be open-sourced

At the CoRL conference in Austin, Nvidia's GEAR team unveiled GR00T 2, a new world-action model, with team lead Yuke Zhu and Jim Fan presenting on stage. Eighteen months after the previous generation, the release is less a single model than a full stack stretching from data collection to deployment.

A 90-second official demo shows a wide task range: trimming tags with scissors, installing a graphics card, striking matches, slicing a banana for breakfast, wiping up coffee, clearing a table and selling popcorn. From fine two-handed manipulation to whole-body movement, the clips cover much of the span of robot capability.

Zhu said the GR00T 2 model, its data, simulation environments and data-collection hardware will be open-sourced in stages, with 100,000 hours of human data already released. For researchers, that means following the official toolchain to capture their own data, adapt it to their own tasks, validate skills in simulation and then deploy to real robots.

Architecturally, GR00T 2 takes language instructions, multi-view camera frames and robot state, and uses a single diffusion transformer to jointly predict future video and action sequences, learning how the world will change alongside how the robot should act. Fan positioned it as a fast, locally runnable System 1 physical-skill model that is easy to teach by demonstration and easy to steer with reasoning models.

The core method is X-Adapt, a cross-embodiment adaptation scheme Zhu described on social media. Teaching a robot a new task used to require fresh robot demonstration data tailored to its body and control stack; X-Adapt aims to transfer skills demonstrated by a human or a handheld device directly to robots that never trained on that task.

The evidence comes in three paths. Researchers wearing motion-capture gear demonstrated placing and wetting filter paper, and the two-handed Dexter robot reached a 65% success rate over 20 trials without any teleoperated robot data for the task. A handheld UMI gripper demonstrated assembling a toy dump truck, and the skill transferred to a YAM arm, lifting average task progress from 50% to 70%. The conclusion is direct: large-scale human experience does not just enrich a model's visual knowledge, it converts into execution progress on real tasks.

The companion toolchain connects capture, data, simulation and execution. Capture uses devices including UMI and DexUMI; data includes Human-1M, which supplies first-person human video with a 100,000-hour public subset, and Flora for curation, captioning and automated evaluation; SimFoundry reconstructs task environments from real scenes for policy testing and data generation. On the execution side, ASPIRE organizes models with code skills for error recovery, Turbo cuts prediction cost with a lighter network and came in at 124 milliseconds per prediction on the G1 humanoid's onboard Jetson Thor, and SONIC runs on a separate Jetson Orin at 50Hz for whole-body control.

Anthropic's Claude Fable 5.2, released the same day, added a competitive edge to the moment. On Anthropic's RoboDojo Sim dual-arm test, run under the same harness and each model's highest available reasoning effort, Fable 5.2 scored 64.9% against 29.8% for GPT-6 Astra. Fan told the CoRL audience that this is not a winter but a robot summer, and the scoring contest between foundation-model teams and embodied-model teams has only just begun.