Models and Datasets

Use these Hugging Face assets when you want to start from NVIDIA-provided data or checkpoints instead of collecting every demonstration and training every model from scratch.

Simulation Workflow

Asset	Hugging Face Link	Use It When
Simulation dataset	nvidia/Arena-G1-Static-PickNPlace-Task	You want to skip simulation teleoperation data collection and start from a provided HDF5 dataset.
Model trained on simulation dataset	nvidia/GN1x-Tuned-Arena-G1-Static-PickNPlace	You want to skip simulation post-training and evaluate a provided GR00T checkpoint in Isaac Lab-Arena.

Real Robot Workflow

Asset	Hugging Face Link	Use It When
Real robot dataset	nvidia/GR00T-N1.7-AppleToPlate	You want to inspect or train from NVIDIA-provided Unitree G1 apple-to-plate demonstrations in LeRobot format.
Model trained on real dataset	nvidia/GR00T-N1.7-ApplePnP-V1	You want to start from a provided real-robot apple pick-and-place model instead of fine-tuning one yourself.

Format Notes

• The simulation dataset starts from HDF5 recordings and is converted to LeRobot format inside the simulation workflow.

• The real robot dataset is published in LeRobot format for the Unitree G1 apple-to-plate task.

• The model assets are intended as starting points or skip paths. Match the model, dataset, embodiment, and deployment configuration before running a policy on hardware.