Fourier ActionNet

傅利叶 ActionNet 数据集Advanced

Fourier's open-source teleoperated dataset of about 30,000 humanoid-plus-dexterous-hand manipulation trajectories.

Fourier ActionNet is a real-robot dataset open-sourced in 2025 by the humanoid robot company Fourier Intelligence. The data was collected by operators performing VR teleoperation while wearing an Apple Vision Pro headset, driving Fourier's GR1-T1, GR1-T2, and GR2 humanoid robots, fitted with either 6-DoF or 12-DoF dexterous hands. It totals over 30,000 trajectories and about 140 hours, mostly tabletop bimanual manipulation such as pick-and-place, pouring water, opening and closing cabinet doors, and precise placement. Each trajectory's language instruction was first auto-generated by Qwen2.5-VL-7B and then manually checked in full. The data uses the LeRobot dataset format and can be used directly to train methods like ACT, Diffusion Policy, and iDP3; the dataset itself is released under a non-commercial license, while the training code is Apache 2.0.

ExampleOne water-pouring trajectory: a GR2 robot picks up a container with its dexterous hand and pours water into a cup, while first-person camera footage, joint states, and a language instruction are all recorded together.

Also called: ActionNet
Related: Fourier · Fourier GR-2 · VR Teleoperation · Dexterous Hand · LeRobotDataset · Real-Robot Data
Sources: Fourier ActionNet 官网 (Chinese)
As of: 2025