Fourier ActionNet — Humanoid Manipulation Dataset
Fourier ActionNet is a real-robot bimanual manipulation dataset developed by Fourier Intelligence in collaboration with Shanghai Jiao Tong University for training dexterous manipulation policies on the Fourier GR-1 and GR-2 humanoid robot platforms. Released in 2024, the dataset covers bimanual tabletop manipulation tasks requiring coordinated dual-arm dexterous control including object transfer, assembly operations, and fine manipulation with multi-finger hands. Data is collected at 50 frames per second with multi-view cameras and wrist cameras capturing detailed hand and finger state information. The dataset is gated and requires acceptance of the Fourier Research License restricting use to non-commercial research. Fourier Intelligence is a Chinese humanoid robot manufacturer known for the GR series humanoid robots with full-size dexterous hands. Fourier ActionNet is one of the only publicly available datasets for full humanoid hand manipulation as opposed to parallel-jaw gripper datasets.
| Year | 2024 |
|---|---|
| Embodiments | Fourier GR-2, Fourier GR-1 |
| Modalities | rgb, proprioception |
| View type | third-person, wrist-cam |
| Task categories | manipulation, bimanual, dexterous-manipulation, pick-and-place |
| Data format | hdf5, lerobot |
| License | Fourier Research License — non-commercial research use |
| Access | gated |
| Maintainer | Fourier Intelligence, Shanghai Jiao Tong University |
| Origin country | CN |
What is it?
Fourier ActionNet is a real-robot bimanual manipulation dataset developed by Fourier Intelligence in collaboration with Shanghai Jiao Tong University for training dexterous manipulation policies on the Fourier GR-1 and GR-2 humanoid robot platforms. Released in 2024, it covers bimanual tabletop manipulation tasks requiring coordinated dual-arm dexterous control with multi-finger hands. Data is collected at 50 frames per second with multi-view and wrist cameras capturing detailed hand and finger state. Access requires acceptance of the Fourier Research License.
Who is it for?
Fourier ActionNet targets researchers working on full humanoid hand manipulation — specifically the challenge of coordinating bimanual dexterous hands with multiple fingers rather than parallel-jaw grippers. It is the primary open dataset for Fourier GR series robots and relevant to all researchers studying multi-finger dexterous manipulation on humanoid platforms.
Key specifications
- Embodiments: Fourier GR-2, Fourier GR-1
- Frame rate: 50 fps
- Tasks: Bimanual coordination, dexterous multi-finger manipulation, assembly operations
- View types: Multi-view, wrist camera
- Format: HDF5, LeRobot
- License: Fourier Research License — non-commercial research only
- Access: Gated — requires license acceptance on Hugging Face
- Origin: China (Fourier Intelligence)
How it compares
Fourier ActionNet is one of the only open datasets capturing multi-finger dexterous hand manipulation on full humanoid platforms. Most bimanual datasets use parallel-jaw grippers (ALOHA, DROID). DexWild and DexUMI cover dexterous manipulation but on standalone hand platforms rather than integrated humanoids. Fourier ActionNet is distinctive for its humanoid-integrated multi-finger hand context.
Limitations and access notes
The Fourier Research License prohibits commercial use. The dataset is gated and requires explicit acceptance of terms. Episode count is not confirmed from primary sources. The dataset is Chinese-origin; geopolitical considerations apply.
Linked professions
Frequently asked questions
What robots does Fourier ActionNet use?
Fourier ActionNet is collected on the Fourier GR-1 and GR-2 humanoid robots manufactured by Fourier Intelligence, a Chinese robotics company. The GR series features full-size bipedal humanoid form with dexterous multi-finger hands.
Can Fourier ActionNet be used commercially?
No. The Fourier Research License restricts use to non-commercial academic research only. Commercial use requires a separate license from Fourier Intelligence.
What makes Fourier ActionNet different from ALOHA?
ALOHA uses parallel-jaw grippers on a bimanual arm platform. Fourier ActionNet captures multi-finger dexterous hand manipulation on full humanoid robots. This makes Fourier ActionNet more relevant for humanoid hand policy research but less accessible due to the non-commercial license restriction.
How do I access Fourier ActionNet?
Fourier ActionNet is available on Hugging Face with gated access. You must accept the Fourier Research License, which requires providing institutional affiliation and agreeing to non-commercial use restrictions.
What is the difference between GR-1 and GR-2?
The Fourier GR-1 is the first generation of Fourier Intelligence's humanoid platform. The GR-2 is an improved second generation with enhanced dexterity, payload, and sensing. Both appear in the Fourier ActionNet dataset.