Human2LocoMan: Learning Versatile Quadrupedal Manipulation with Human Pretraining

Yaru Niu^1,*    Yunzhe Zhang^1,*    Mingyang Yu¹    Changyi Lin¹    Chenhao Li¹    Yikai Wang¹
Yuxiang Yang²    Wenhao Yu²    Tingnan Zhang²    Zhenzhen Li³    Jonathan Francis^1,3    Bingqing Chen³   
Jie Tan²    Ding Zhao^1   
1Carnegie Mellon University    ²Google DeepMind    ³Bosch Center for AI   
^*Equal contributions

Robotics: Science and Systems (RSS) 2025
Website | Paper | Code

---

Data Description

This dataset includes both human and LocoMan robot data for six manipulation tasks featured in our work: unimanual toy collection, bimanual toy collection, unimanual shoe organization, bimanual shoe organization, unimanual scooping, and bimanual pouring.

Data Structure

The human and robot data share a unified format and are stored in HDF5 files with the same structure as follows. The accompanying MP4 video files are provided for reference only.

root
├── observations
│   ├── images
│   │   ├─ main          # images from main camera: [h * w * c] * traj_length
│   │   └─ wrist         # images from wrist cameras (at most two, right first): [h * w * c] * traj_length
│   └── proprioceptions       
│       ├─ body          # 6d pose of the rigid body where the main camera is mounted: [6] * traj_length
│       ├─ eef           # 6d pose of the end effectors (at most two, right first): [12] * traj_length
│       ├─ relative      # relative 6d pose of the end effector to the rigid body where the main camera is mounted (at most two, right first): [12] * traj_length
│       ├─ gripper       # gripper angle (at most two, right first): [2] * traj_length
│       └─ other         # other prorioceptive state, e.g., robot joint positions, robot joint velocities, human hand joint poses, ...
├── actions       
│   ├── body             # 6d pose of the rigid body where the main camera is mounted: [6] * traj_length
│   ├── delta_body       # delta 6d pose of the rigid body where the main camera is mounted: [6] * traj_length
│   ├── eef              # 6d pose of the end effectors (at most two, right first): [12] * traj_length
│   ├── delta_eef        # delta 6d pose of the end effectors (at most two, right first): [12] * traj_length
│   ├── gripper          # gripper angle (at most two, right first): [2] * traj_length
│   └── delta_gripper    # delta gripper angle (at most two, right first): [2] * traj_length
└── masks                # embodiment-specific masks to mask out observations and actions for training and inference (the mask for observations and actions of the same modality could be different)
    ├── img_main         # mask for the main camera image input: [1]
    ├── img_wrist        # mask for the wrist camera image input: [2]
    ├── proprio_body     # mask for the body 6d pose input: [6]
    ├── proprio_eef      # mask for the eef 6d pose input: [12]
    ├── proprio_gripper  # mask for the gripper angle input: [2]
    ├── proprio_other    # mask for other proprioception input (n components): [n]
    ├── act_body         # mask for the body 6d pose output: [6]
    ├── act_eef          # mask for the eef 6d pose output: [12]
    └── act_gripper      # mask for the gripper angle output: [2]

Citation

If you find this work helpful, please consider citing the paper:

@inproceedings{niu2025human2locoman,
  title={Human2LocoMan: Learning Versatile Quadrupedal Manipulation with Human Pretraining},
  author={Niu, Yaru and Zhang, Yunzhe and Yu, Mingyang and Lin, Changyi and Li, Chenhao and Wang, Yikai and Yang, Yuxiang and Yu, Wenhao and Zhang, Tingnan and Li, Zhenzhen and Francis, Jonathan and Chen, Bingqing and Tan, Jie and Zhao, Ding},
  booktitle={Robotics: Science and Systems (RSS)},
  year={2025}
}