GR1 Robot Teleoperation Dataset

The GR1 Robot Teleoperation Dataset is a large-scale humanoid robot teleoperation dataset collected using the Fourier GR-1 (GR1T1) robot, containing 22,209 episodes of egocentric video observations paired with full-body robot states and actions across 44 degrees of freedom. Released in May 2026 under MIT license for commercial use, it provides one of the most comprehensive full-body humanoid teleoperation datasets available — covering all 44 DoF of the GR-1 including finger joints, wrist, arm, torso, and leg actuation. The dataset is collected in real-world environments using a whole-body teleoperation interface, capturing natural coordination between upper and lower body during manipulation tasks. The MIT license permits unrestricted commercial use, making it more accessible than Fourier ActionNet's research-only license.

Dataset specifications
Year2026
Episodes22,209
EmbodimentsFourier GR-1, Fourier GR-1T1
Modalitiesrgb, proprioception
Task categoriesmanipulation, bimanual, dexterous-manipulation, loco-manipulation
Data formathdf5, mp4
LicenseMIT
Accessopen — commercial use permitted
MaintainerNVIDIA, Fourier Intelligence
Origin countryCN

What is it?

The GR1 Robot Teleoperation Dataset contains 22,209 episodes of egocentric video observations paired with full-body robot states and actions across all 44 degrees of freedom of the Fourier GR-1 humanoid robot. Released in May 2026 under MIT license for commercial use, it provides one of the most comprehensive full-body humanoid teleoperation datasets available — capturing finger joints, wrist, arm, torso, and leg actuation simultaneously during real-world manipulation tasks.

Who is it for?

Researchers training whole-body humanoid manipulation policies on the Fourier GR-1 platform, and those studying full-body coordination between upper and lower body during manipulation. The 44-DoF coverage is significantly higher than datasets focused on arm-only manipulation, enabling training of policies that coordinate locomotion with manipulation.

Key specifications

How it compares

The GR1 dataset complements Fourier ActionNet — which covers bimanual manipulation on GR series robots but with a research-only license. The GR1 Teleoperation Dataset's MIT license makes it commercially accessible. At 22,209 episodes it is smaller than AgiBot-Alpha (1,979 hours) but covers full 44-DoF body state rather than arm-focused data.

Limitations and access notes

GR-1 specific — policies may require adaptation for other humanoid platforms. MIT permits unrestricted commercial use.

Linked professions

Frequently asked questions

What is the Fourier GR-1 robot?

The Fourier GR-1 is a full-size bipedal humanoid robot developed by Fourier Intelligence, featuring dexterous multi-finger hands and 44 degrees of freedom. It is the platform used in the GR1 Teleoperation Dataset, with data captured across all joints simultaneously including finger, wrist, arm, torso, hip, and leg actuation.

Can the GR1 Teleoperation Dataset be used commercially?

Yes. MIT licensed, permitting unrestricted commercial use, modification, and redistribution.

How does 44 DoF compare to typical manipulation datasets?

Most manipulation datasets capture 6-8 DoF from a single arm. The GR1 dataset captures all 44 DoF simultaneously — fingers (~16 DoF), wrists (~2 DoF), arms (~14 DoF), torso (~4 DoF), legs (~12 DoF). This full-body state coverage enables training of whole-body coordination policies that arm-only datasets cannot support.

How do I access the GR1 Teleoperation Dataset?

Available on Hugging Face at huggingface.co/datasets/GR1-robot/GR1-Teleoperation-Dataset. No registration required.

How does this differ from Fourier ActionNet?

Fourier ActionNet is licensed for non-commercial research only and focuses on bimanual tabletop manipulation. The GR1 Teleoperation Dataset is MIT licensed (commercial use permitted) and captures full 44-DoF whole-body state during real-world teleoperation, making it more accessible and broader in scope.