| --- |
| license: cc-by-4.0 |
| task_categories: |
| - robotics |
| language: |
| - en |
| size_categories: |
| - n<1K |
| tags: |
| - lerobot |
| - Unitree G1 |
| --- |
| # Dataset Card for HuMI |
| [[Project Page]](https://humanoid-manipulation-interface.github.io/) | [[Paper]](http://arxiv.org/abs/2602.06643) |
|
|
| ### Dataset Summary |
| This dataset was collected using the [HuMI](https://humanoid-manipulation-interface.github.io/) data collection pipeline and converted into the LeRobot format. It provides robot-free demonstrations for humanoid whole-body manipulation. |
|
|
| **Task Description:** This repository features a `walk-to-clean-table` task in which a humanoid robot navigates to a desk and cleans the tabletop using a lint roller to remove scattered paper scraps. |
|
|
| The dataset consists of 105 demonstrations collected within a single environment. |
|
|
| ## Dataset Structure |
|
|
| ### Data Instances |
|
|
| Each instance in the dataset represents a single frame from an episode. The data includes synchronized camera observations, end-effector poses, joint positions, and task instructions. |
|
|
| ### Data Fields |
|
|
| The dataset contains the following features: |
|
|
| - `task` (`string`): The natural language description of the task. |
| - `observation.images.camera{cam_id}_rgb` (video, `(H, W, 3)`): RGB observations from the cameras. |
| - `{robot_name}_eef_pos` (`float32`, `(3,)`): End-effector position (x, y, z). |
| - `{robot_name}_eef_rot_axis_angle` (`float32`, `(3,)`): End-effector rotation in axis-angle representation. |
| - `{robot_name}_gripper_width` (`float32`, `(1,)`): Gripper width (only present for gripper end-effectors). |
| - `{robot_name}_demo_start_pose` (`float32`, `(6,)`): Initial 6DoF pose of the end-effector for the demonstration. |
| - `{robot_name}_demo_end_pose` (`float32`, `(6,)`): Final 6DoF pose of the end-effector for the demonstration. |
| - `joint_pos` (`float32`, `(29,)`): Full joint positions of the Unitree G1 robot, computed via inverse kinematics using the end-effector poses as targets. |
|
|
| For the HuMI setup, we recorded images from two wrist-mounted cameras, where `camera0` corresponds to the right gripper and `camera1` corresponds to the left. |
| Additionally, we recorded trajectories for five end-effectors, adopting the naming convention from [UMI](https://umi-gripper.github.io/): |
|
|
| - `robot0`: right gripper |
| - `robot1`: left gripper |
| - `robot2`: pelvis |
| - `robot3`: right foot |
| - `robot4`: left foot |
|
|
| ### Citation Information |
| ```bibtex |
| @article{nai2026humanoid, |
| title={Humanoid Manipulation Interface: Humanoid Whole-Body Manipulation from Robot-Free Demonstrations}, |
| author={Nai, Ruiqian and Zheng, Boyuan and Zhao, Junming and Zhu, Haodong and Dai, Sicong and Chen, Zunhao and Hu, Yihang and Hu, Yingdong and Zhang, Tong and Wen, Chuan and others}, |
| journal={arXiv preprint arXiv:2602.06643}, |
| year={2026} |
| } |
| ``` |
