NVIDIA GR00T Teleop Datasets

Humanoid robot teleoperation data for the GR00T foundation model. Real + sim, multiple embodiments, 13M+ total rows.

NVIDIA Open Model License Parquet + MP4 Humanoid

Dataset Variants

VariantRobotRowsSizeDownloads
GR00T Teleop GR1NVIDIA GR1 humanoid7.55M~50 GB65.2K
GR00T Teleop SimSimulated humanoid5.82M~40 GB6.5K
GR00T Teleop G1Unitree G1 humanoid124K~2 GB526
GR00T X-Embodiment SimMultiple (sim)345K~5 GB--
Manipulation KitchenRobot arm405K~5 GB1.5K
Open-H-EmbodimentHumanoid (cross)--~0.1 GB36.9K
GraspGenHand/gripper25.5K~3 GB1K

What is NVIDIA GR00T?

GR00T (Generalist Robot 00 Technology) is NVIDIA's foundation model initiative for humanoid robots. The GR00T teleop datasets are the real-world and simulated teleoperation recordings used to train these models. The GR1 variant alone contains 7.55 million rows of humanoid teleoperation data, making it the most downloaded NVIDIA robotics dataset with 65.2K downloads.

The collection spans multiple embodiments (NVIDIA GR1, Unitree G1, simulated humanoids) and tasks (manipulation, kitchen interactions, grasping), providing a cross-embodiment foundation for humanoid policy learning. NVIDIA also released Cosmos world model integrations with RoboCasa and LIBERO environments, bridging the gap between simulation and real-world deployment.

Access

License note: NVIDIA Open Model License allows redistribution with attribution but restricts use for competing model development. Review full terms before commercial use.

View on HuggingFace All NVIDIA Datasets

Related datasets

Deploying humanoid robots?

We lease Unitree G1 and H1 humanoids with teleoperation rigs for data collection.