LocoMuJoCo

A MuJoCo-based imitation-learning benchmark for whole-body locomotion, with humanoid, quadruped, and human musculoskeletal models.

LocoMuJoCo is an open-source benchmark released in 2023 by Jan Peters's group at TU Darmstadt in Germany (Firas Al-Hafez and colleagues), specifically for evaluating imitation-learning algorithms aimed at locomotion. Earlier locomotion benchmarks were mostly simplified toy tasks; this one provides quadruped, humanoid, and human musculoskeletal models, paired with real motion-capture data, expert data, and suboptimal data. Later versions switched to JAX, adding MJX and MuJoCo Warp for parallel simulation; it currently contains 12 humanoid and 4 quadruped environments, retargeting over 22,000 motion-capture clips from AMASS, LAFAN1, and other sources onto various humanoids, along with baselines including PPO, GAIL, AMP, and DeepMimic, plus a domain-randomization interface.

ExampleUsing ImitationFactory to create a Unitree H1 environment, loading dance and walking clips from LAFAN1, and then training a policy to imitate those motions with the built-in AMP algorithm.

Also called

Imitation Learning Benchmark for Whole-Body Locomotion

Related

MuJoCo (Multi-Joint dynamics with Contact) · MuJoCo XLA · Imitation Learning · Motion Retargeting · AMASS (Archive of Motion Capture as Surface Shapes) · Adversarial Motion Priors

Sources

LocoMuJoCo: A Comprehensive Imitation Learning Benchmark for Locomotion (arXiv 2311.02496)
loco-mujoco GitHub 仓库 (Chinese)

As of

2026-09