README.md

metadata

title: Human to G1 Robot Motion Retargeting
emoji: 🤖
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.23.0
app_file: app.py
pinned: false

Human → G1 Robot Motion Retargeting

Neural motion retargeting from human motion capture (SMPL-X / AMASS format) to Unitree G1 humanoid robot.

How to Use

1. Upload an AMASS-format .npz file containing human motion data
2. The model will retarget the motion to G1 robot in ~5-15s (CPU)
3. View the 3D skeleton animation and download the result PKL file

Input Format

AMASS NPZ file with the following fields:

• trans or transl: (T, 3) root translation
• root_orient or global_orient: (T, 3) root orientation (axis-angle)
• pose_body or body_pose: (T, 63) body joint rotations (21 joints × 3)
• mocap_frame_rate (optional): source frame rate (auto-downsampled to 30 FPS)

Coordinate system: Z-up (AMASS convention), automatically converted to Y-up internally.

Output Format

PKL file containing:

• dof: (T, 29) — G1 joint angles in radians
• root_trans: (T, 3) — root position (Y-up, meters)
• root_rot_quat: (T, 4) — root rotation quaternion (wxyz format)

Model

• Architecture: LLaMA-style Transformer (70M parameters) with VQ-VAE encoder
• Training data: paired SMPL-X ↔ G1 motion sequences
• Inference: non-autoregressive forward pass with sliding window (4s chunks, 1s overlap)