PhysicalAI Robotics GR00T Teleop G1

The PhysicalAI Robotics GR00T Teleop G1 dataset is a real-robot teleoperation dataset collected by NVIDIA using the Unitree G1 humanoid robot. Released in 2025 under the CC BY 4.0 license for commercial and research use, it contains 187 teleoperation episodes recorded at 20 frames per second covering upper-body manipulation tasks with language-conditioned prompts. The dataset is formatted in Parquet and distributed via Hugging Face, with 124,000 total rows across training splits. It is part of NVIDIA's broader GR00T ecosystem for training general-purpose humanoid robot policies and is designed for behavior cloning and language-conditioned policy learning. The Unitree G1 is a priority embodiment in the Geppetto robot directory with direct supply chain documentation across reducers, motors, sensors, and screw components. This dataset provides one of the first open, commercially licensed teleoperation datasets specifically for a production humanoid platform available at consumer scale.

Dataset specifications
Year2025
Episodes187
Frame rate20 fps
EmbodimentsUnitree G1
Modalitiesrgb, proprioception
View typethird-person
Task categoriesmanipulation, pick-and-place
Data formatparquet, hdf5, mp4
LicenseCC BY 4.0
Accessopen — commercial use permitted
MaintainerNVIDIA
Origin countryUS

What is it?

PhysicalAI Robotics GR00T Teleop G1 is a real-robot teleoperation dataset released by NVIDIA in 2025 for training manipulation policies on the Unitree G1 humanoid robot. It contains 187 teleoperation episodes recorded at 20 frames per second, distributed in Parquet format via Hugging Face under the CC BY 4.0 license. The dataset is part of NVIDIA's broader GR00T ecosystem for developing general-purpose humanoid robot policies and is designed for behaviour cloning and language-conditioned policy learning on production humanoid hardware.

Who is it for?

This dataset is designed for robotics researchers and engineers working on humanoid robot manipulation using the Unitree G1 platform. It is particularly valuable for teams working with NVIDIA's Isaac GR00T framework and those seeking commercially licensable teleoperation data specifically for a production humanoid platform available at consumer scale. Researchers building on the GR00T foundation model can use this dataset for fine-tuning on new manipulation tasks.

Key specifications

How it compares

GR00T Teleop G1 is modest in scale (187 episodes) compared to EgoVerse (57,761 episodes) and DROID (76,000 episodes), but it is one of the only open datasets collected specifically on a production humanoid robot platform (Unitree G1) rather than tabletop arm hardware. Its value is specificity rather than scale — it provides commercially licensed, real-hardware demonstration data for the most widely deployed affordable humanoid robot. PHUMA provides complementary locomotion data for the same platform.

Limitations and access notes

The dataset covers upper-body manipulation only — it does not include full-body locomotion or whole-body coordination tasks. 187 episodes is a small dataset by modern standards; it is best used for fine-tuning pre-trained policies rather than training from scratch. The CC BY 4.0 license permits commercial use with attribution. Access is open via Hugging Face with no registration required.

Linked professions

Frequently asked questions

What robot is the GR00T Teleop G1 dataset collected on?

The dataset is collected on the Unitree G1 humanoid robot, a full-size bipedal humanoid available for direct purchase starting at $16,000. The Unitree G1 is one of the most widely deployed affordable humanoid robots and is the target platform for NVIDIA's GR00T foundation model.

Can GR00T Teleop G1 be used commercially?

Yes. The dataset is licensed under CC BY 4.0, which permits commercial use. You must provide attribution to NVIDIA as the dataset maintainer.

How do I access the GR00T Teleop G1 dataset?

The dataset is freely available on Hugging Face at huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-Teleop-G1. No registration is required. It is distributed in Parquet format compatible with the Hugging Face Datasets library.

What is the GR00T ecosystem?

GR00T (Generalist Robot 00 Technology) is NVIDIA's initiative for developing foundation models for humanoid robots. The GR00T Teleop G1 dataset is one of the training data sources for this ecosystem, alongside the Isaac GR00T framework for sim-to-real transfer and robot policy training.

Is 187 episodes enough to train a manipulation policy?

187 episodes is typically insufficient to train a policy from scratch but is well-suited for fine-tuning a pre-trained foundation model such as GR00T or OpenVLA. For training from scratch, combining this dataset with larger datasets such as Open X-Embodiment or EgoVerse is recommended.