Welcome to LocoMuJoCo!

Latest News

🚀 Since Release: v1.0

LocoMuJoCo now supports MJX and includes new JAX algorithms, expanded environments, and over 22,000 datasets!

Overview

LocoMuJoCo is an imitation learning benchmark tailored for whole-body control. It includes a diverse range of environments—quadrupeds, humanoids, and (musculo-)skeletal human models—each equipped with comprehensive datasets (22k+ samples per humanoid).

While designed for imitation learning, it also supports pure reinforcement learning with custom reward classes.

Key Advantages

• ✅ Supports MuJoCo (single) and MJX (parallel) environments

• ✅ Includes 12 humanoid + 4 quadruped environments, with 4 biomechanical human models

• ✅ Clean, single-file JAX algorithms: PPO, GAIL, AMP, DeepMimic

• ✅ 22,000+ motion capture datasets (AMASS, LAFAN1, native)

• ✅ Robot-to-robot retargeting

• ✅ Trajectory comparison metrics (e.g., DTW, Fréchet distance) implemented in JAX

• ✅ Gymnasium interface

• ✅ Built-in domain and terrain randomization

• ✅ Modular design: easily swap components (observations, rewards, terminal handlers, etc.)

• ✅ Comprehensive [documentation](https://loco-mujoco.readthedocs.io/)