LocoMuJoCo: A Comprehensive Imitation Learning Benchmark for Locomotion
LocoMuJoCo is an imitation-learning benchmark for locomotion in MuJoCo, presented at a NeurIPS 2023 robot learning workshop. It contains 12 humanoid and 4 quadruped environments, plus biomechanical human models, each with noisy motion-capture reference data mapped to the embodiment, expert demonstrations and sub-optimal demonstrations. The repository adds MJX support for parallel simulation.
How well imitation-learning methods reproduce agile locomotion from motion-capture references, scored with handcrafted, task-specific metrics. The repository's documentation names trajectory comparison metrics such as dynamic time warping and Fréchet distance.
Environments run in MuJoCo for single instances and in MJX or MJWarp for parallel instances. Each environment ships with reference motion data, expert datasets and handcrafted metrics, and baseline algorithms are included. The repository also supports custom reward classes, so pure reinforcement learning is possible. Lower trajectory distance means the motion is closer to the reference.
Humanoid walking learned from human motion data is a central problem in humanoid control. The benchmark includes humanoid environments (the README names Unitree H1 and G1) with motion-capture data, so it targets that problem directly.
Scoring
| Metric | Key | Unit | Direction |
|---|---|---|---|
| Trajectory distance (DTW / Fréchet, handcrafted per task) | trajectory_distance | Unitless | Lower is better ↓ |
Hardware
The README names Unitree H1 ('UnitreeH1') and Unitree G1 ('MjxUnitreeG1') among its humanoid environments. Quadruped and biomechanical models (MyoSkeleton) are also included. Other humanoids, such as Atlas and Talos, are not named in the sources opened.
Recorded outcomes
No result files yet
Evidence
Related evaluations