introvoyz042/SEW-TWIST

SEW-TWIST G1 Teleoperation Dataset

This dataset contains offline teleoperation trajectories for the Unitree G1 humanoid robot generated using the SEW-MIMIC controller \[1\] and LaFAN1 BVH motion capture data \[2\].

The dataset was generated by replaying BVH motion capture sequences through a MuJoCo simulation of the G1 robot and logging the resulting robot state trajectories in a format compatible with TWIST-style imitation learning pipelines \[3\].

Each trajectory is stored as a .pkl file containing joint states, root pose, and body positions for each simulation frame.

Dataset Structure

dataset/
│
├── *.pkl              # Motion trajectories
├── metadata.csv       # Optional metadata per sequence
└── README.md

Each .pkl file contains a single trajectory sequence recorded from BVH playback.

Data Format

Each .pkl file contains a Python dictionary with the following structure:

{
    "fps": float,
    "root_pos": np.ndarray,
    "root_rot": np.ndarray,
    "dof_pos": np.ndarray,
    "local_body_pos": np.ndarray,
    "link_body_list": list[str]
}

Fields

fps

float

Frames per second of the recorded trajectory.

root_pos

shape: (T, 3) dtype: float32

World-space position of the robot root (pelvis mocap frame).

[x, y, z]

Units: meters

root_rot

shape: (T, 4) dtype: float32

Root orientation quaternion in (x, y, z, w) format.

This is converted from MuJoCo's internal (w, x, y, z) ordering during logging.

dof_pos

shape: (T, 21) dtype: float32

Joint configuration vector for the robot.

The DOF vector is concatenated as:

[leftleg, rightleg, torso, leftarm, rightarm]

DOF Breakdown

Total: 23 DOF

localbodypos

shape: (T, N, 3) dtype: float32

Local body positions for each link relative to the root frame.

Positions are computed as:

local = Rroot^T * (xworld - root_pos)

Where

• —R_root is the root rotation matrix
• —x_world is the world position of the body

linkbodylist

list[str] length = N

Names of the robot bodies corresponding to the local_body_pos array.

Example entries include pelvis, hip joints, knees, ankles, torso, shoulders, elbows, wrists, and hands.

Example Usage

import pickle

with open("trajectory.pkl", "rb") as f:
    data = pickle.load(f)

print(data.keys())

Output:

dict_keys([
    'fps',
    'root_pos',
    'root_rot',
    'dof_pos',
    'local_body_pos',
    'link_body_list'
])

Intended Use

This dataset can be used for:

• —humanoid motion imitation learning
• —motion retargeting research
• —policy learning from motion capture
• —trajectory prediction
• —humanoid control benchmarking

Citation

This dataset is released as part of the SEW-MIMIC project

@misc{sew_mimic,
  title={A Closed-Form Geometric Retargeting Solver for Upper Body Humanoid Robot Teleoperation},
  author={Kong, Chuizheng and Cho, Yunho and Jung, Wonsuhk and others},
  year={2026},
  note={Project website: https://sew-mimic.com/}
}

Acknowledgements

• —LaFAN1 Motion Capture Dataset
• —MuJoCo Physics Engine
• —SEW Geometric Teleoperation Framework

References

\[1\] A Closed-Form Geometric Retargeting Solver for Upper Body Humanoid Robot Teleoperation, Project website: https://sew-mimic.com/ arXiv preprint arXiv:2602.01632, 2026

\[2\] LaFAN1 Motion Capture Dataset Ubisoft La Forge. https://github.com/ubisoft/ubisoft-laforge-animation-dataset

\[3\] Y. Ze, Z. Chen, J. P. Araújo, Z. Cao, X. B. Peng, J. Wu, and C. K. Liu, "TWIST: Teleoperated Whole-Body Imitation System," Project website: https://yanjieze.com/TWIST/

Data Sources and Licensing

This dataset is generated from the LaFAN1 Motion Capture Dataset provided by Ubisoft La Forge.

The original dataset is licensed under:

Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)

See the original repository for details: https://github.com/ubisoft/ubisoft-laforge-animation-dataset

Users of this dataset must comply with the license terms of the original dataset.