zhouyikai/FDDC-single-leg-balance

FDDC — Stratified Single-Leg-Balance Motions (900)

The motion set used by FDDC (First Deployable Dynamic-CoM), a unified policy + method-agnostic sim2sim benchmark for humanoid single-leg balance. Paper: arXiv:2608.00500. Code & benchmark: https://github.com/zhouyikai888/FDDC

Contents

sample_<support>_<id>_mj.npz — Unitree G1 29-DoF, 50 Hz, 249-frame clips, each built around a single-support balance phase (double → single → double).

• train/ — 720 clips · val/ — 90 clips · test/ — 90 clips
• dataset_info.json, manifest.csv — per-clip class / bin / support / split labels
• Split: balanced across 9 pose classes (3×3 grid of pelvis height × swing-foot height) and both support sides, seed 42. Checkpoints are selected on val; all numbers are reported on test.

Each .npz holds joint_pos, joint_vel, body_pos_w, body_quat_w, body_lin_vel_w, reference_support_phase, reference_support_state, body_names, joint_names, fps.

Provenance & license

These motions are a derivative of the AMS synthetic balance-motion dataset (Pan et al., "Agility Meets Stability", arXiv:2511.17373, HF: ruoyiqiao/AMS). From AMS's ~9,814 valid synthetic single-leg clips we (1) retargeted to the Unitree G1 29-DoF MuJoCo model and (2) distilled a compact, difficulty-stratified, balanced set of 900. The AMS motions are synthetic (trajectory-optimization generated), not motion capture.

License: GPL-3.0, inherited from AMS. Please cite both FDDC (arXiv:2608.00500) and AMS (arXiv:2511.17373) if you use this set.