01 · Dataset
This LAFAN1 benchmark uses all 40 G1 sequences from the LAFAN1 Retargeting Dataset. Motions are used as supplied, without filtering by difficulty.
| Property | Specification |
|---|---|
| Source | lvhaidong/LAFAN1_Retargeting_Dataset |
| Pinned revision | ce1572906efe6157840e8474d5a0d7aa87481e74 |
| Coverage | 40 clips · 8,822.1667 seconds · first frame to exact final timestamp |
| Integrity | The manifest fixes filenames, file sizes and SHA256 checksums. |
- This measures tracking on the supplied reference set; it does not establish that every motion is physically feasible or unseen during training.
02 · Evaluation setup
| Setting | Specification |
|---|---|
| Robot | Unitree G1 · 29 actuated joints · canonical benchmark asset |
| Engine | MuJoCo 3.17.0 · 2 ms physics step |
| Initialization | First reference pose and canonical finite-difference velocities; histories reset independently per clip. |
| Environment | Nominal: no observation noise, no domain randomization; sliding friction 1. |
| Controller | Author-native control frequency, observations, history, normalization, gains and action processing. |
| Scoring | Every 0.02 s after physics, plus the exact reference endpoint. |
| State access | Ideal-proprioceptive-no-position-v1: joint state, full orientation and ideal root velocities; no actual position, contacts or explicit odometry. |
- Scored body origins (14): pelvis, left_hip_yaw_link, right_hip_yaw_link, torso_link, left_knee_link, right_knee_link, left_ankle_roll_link, right_ankle_roll_link, left_shoulder_yaw_link, right_shoulder_yaw_link, left_elbow_link, right_elbow_link, left_wrist_yaw_link, right_wrist_yaw_link.
03 · Failure configuration
Root height and inclination follow SuperMimic Appendix A. End-effector deviation is not a failure condition.
| Condition | Body mapping | Failure threshold |
|---|---|---|
| Root height | pelvis | > 0.25 m |
| Root inclination | pelvis | > 0.8 · dimensionless |
- End-effector deviation does not trigger failure. No XY or yaw alignment is applied to failure checks.
04 · Reproduction
Clone a model repository with its core submodule, run uv sync, then uv run evaluate.py. This evaluates the complete motion set and verifies the saved trajectories.