NVIDIA Physical AI Dataset

NVIDIA's open collection of robotics and autonomous-driving datasets hosted on Hugging Face.

The NVIDIA Physical AI Dataset is an open collection of datasets that NVIDIA announced at its GTC conference on March 18, 2025, hosted on Hugging Face, aimed at developing physical AI for robotics, autonomous driving, and related fields, combining both real and synthetic data. The initial release totaled about 15TB, including more than 320,000 robot training trajectories and up to 1,000 OpenUSD assets. It isn't a single file but a series of datasets under NVIDIA's Hugging Face account, all prefixed PhysicalAI-, such as the simulated trajectories used for GR00T N1 post-training, GR00T teleoperation data, SimReady asset repositories, about 1,700 hours of multi-sensor autonomous-driving data, and, added in 2026, synthetic world-model scene data. As of September 2026, more than 40 datasets exist under this prefix, each with its own license, so licenses need to be checked individually before use.

ExampleThe GR00T-X-Embodiment-Sim subset provides the simulated trajectories used for GR00T N1 post-training; the GR-1 humanoid's tabletop manipulation portion alone has 240,000 trajectories, downloadable directly for fine-tuning GR00T.

Also called

PhysicalAI Dataset

Related

NVIDIA · NVIDIA Isaac GR00T N1 · Synthetic Data · SimReady Assets · Autonomous Driving · Physical AI

Sources

NVIDIA 博客：NVIDIA Releases Open Physical AI Dataset (Chinese)
nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim（Hugging Face）

As of

2026-09