CLAW

CLAW is a text-to-motion dataset for humanoid robot motion generation, with all trajectories designed for the Unitree G1 robot. The dataset comprises 14,970 motion samples, each recorded at 50 Hz.

Each sample includes a kinematic trajectory, a corresponding dynamic trajectory, and a natural-language annotation file. The source repository is available at JianuoCao/CLAW.

Repository Structure

CLAW/
├── README.md
├── dataset_summary.json
├── requirements-render.txt
├── g1_xml/
│   ├── g1_29dof_rev_1_0.xml
│   └── meshes/
├── script/
│   └── render_video.py
└── motion/
    ├── 00000_right_hook_kneel_one_idle_boxing_right_jab/
    │   ├── trajectory_kinematic.csv
    │   ├── trajectory_dynamic.csv
    │   ├── annotation.json
    │   └── trajectory_dynamic.mp4
    ├── 00001_random_punches_crouch_left_hook/
    └── ...

Each motion directory follows the same core structure:

File	Description
trajectory_kinematic.csv	Reference kinematic trajectory
trajectory_dynamic.csv	Dynamic trajectory associated with the sample
annotation.json	Segment-level and full-trajectory language annotations

Trajectory Format

Trajectory files are stored as CSV files at 50 FPS. Each row corresponds to one frame and contains a 36-dimensional robot state vector.

Columns	Dimension	Description
0:3	3	Root position (x, y, z) in meters
3:7	4	Root orientation quaternion (x, y, z, w)
7:36	29	Robot joint positions in radians

The CSV files store root quaternions in xyzw order. Downstream simulation or tokenization pipelines may convert this representation to wxyz internally.

Annotation Format

Each sample includes an annotation.json file containing segment-level timing, action labels, and natural-language descriptions.

Field	Description
recipe_name	Canonical action-sequence name
total_duration_sec	Total motion duration in seconds
data_fps	Frame rate of the trajectory files
trajectory_file	Primary dynamic trajectory referenced by the annotation
segments	Timed action segments with frame ranges, action labels, movement direction, duration, and speed metadata
full_trajectory	Natural-language descriptions of the complete motion
per_segment	Natural-language descriptions for each individual segment

Video Example

Example motion: 00001_random_punches_crouch_left_hook

Dynamic trajectory:

Open dynamic MP4

Kinematic trajectory:

Open kinematic MP4

Annotation: annotation.json

Field	Value
Recipe	random_punches_crouch_left_hook
Duration	14.82 s
FPS	50
Summary	The robot throws random punches forward for 3.1 seconds, then crouch-walks forward for 3.1 seconds, and finally throws a left hook forward for 8.7 seconds.

Segment	Time	Frames	Action
0	0.00-3.11 s	0-156	random_punches forward
1	3.11-6.17 s	156-308	crouch forward
2	6.17-14.82 s	308-741	left_hook forward

Video Rendering

The repository includes the Unitree G1 MuJoCo XML model used for visualization:

g1_xml/g1_29dof_rev_1_0.xml

Video previews can be rendered with the repository-provided MuJoCo EGL renderer:

script/render_video.py

The renderer is self-contained in this repository and does not require the original development Conda environment. A minimal Python environment is sufficient.

Minimal Python dependencies are listed in:

requirements-render.txt

Create a minimal environment with:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements-render.txt

sudo apt-get install ffmpeg

Example command:

python script/render_video.py \
  --csv motion/00001_random_punches_crouch_left_hook/trajectory_dynamic.csv \
  --output motion/00001_random_punches_crouch_left_hook/videos_hd_grid/trajectory_dynamic.mp4 \
  --model g1_xml/g1_29dof_rev_1_0.xml \
  --fps 50 \
  --width 1920 \
  --height 1080 \
  --track-body pelvis \
  --track-distance 4.0 \
  --track-azimuth 90 \
  --track-elevation -8 \
  --white-grid-scene \
  --crf 18 \
  --preset medium