NVIDIA GR00T X-Embodiment Sim
NVIDIA (GEAR Lab / Isaac) · 2025 · datasets.bot · datasets.bot page
One-liner. NVIDIA's open simulated humanoid/arm manipulation dataset used to post-train the GR00T N1 foundation model; LeRobot format, CC-BY-4.0.
Setup
- Datasets / benchmarks: The public, downloadable dataset behind NVIDIA's Isaac GR00T N1 humanoid foundation model. It bundles simulated post-training trajectories across multiple embodiments (Fourier GR1 humanoid in arms-only/full-upper-body/arms+waist configs, single-arm and bimanual Franka Panda with gripper and dexterous-hand variants, and a Unitree G1 loco-manipulation set) covering cross-embodied bimanual manipulation, humanoid tabletop manipulation, and robot-arm kitchen manipulation. Released on Hugging Face under CC-BY-4.0 in a LeRobot-compatible per-task layout (data/meta/videos), it has surpassed 10M downloads. The related DreamGen / GR00T-Dreams release (arXiv 2505.12705, Apache-2.0) is a synthetic-trajectory generation method plus code/benchmark, not a standalone downloadable trajectory dataset. License: CC-BY-4.0. Download: https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim.
- Hardware / simulator: Embodiment: franka_panda, humanoid_other, unitree_humanoid. Environment: simulation, kitchen, tabletop. Realness: simulated.
Schema
Multiple per-embodiment/per-task LeRobot datasets (data/meta/videos) -> episodes -> steps -> {observation (rgb images + state/proprioception), action, language task annotation}
Links