Repolode
Submit
robfiras

robfiras/loco-mujoco

An imitation learning benchmark for whole-body control, with MuJoCo and MJX environments for humanoids, quadrupeds and musculoskeletal human models.

View on GitHubWebsite
Stars
1,472
Forks
152
Open issues
43
Last commit
2 days ago

What is it?

LocoMuJoCo is an imitation learning benchmark for whole-body control. It provides a set of locomotion environments covering humanoids, quadrupeds, and (musculo-)skeletal human models, each paired with motion capture datasets such as AMASS, LAFAN1, and a native LocoMuJoCo dataset.

Environments run on MuJoCo for single-instance simulation or on MJX and MJWarp for parallel simulation on compatible GPUs. The project includes single-file JAX implementations of PPO, GAIL, AMP and DeepMimic, and it supports custom reward functions for pure reinforcement learning. It also exposes a Gymnasium interface.

Beyond training, the project offers robot-to-robot motion retargeting, trajectory comparison metrics such as dynamic time warping and discrete Fréchet distance, and built-in domain and terrain randomization. Observation types, rewards, terminal handlers and randomization components can be swapped and reused.

Key features

  • 01Humanoid, quadruped and biomechanical human environments built on MuJoCo
  • 02Parallel simulation via MJX and MJWarp backends
  • 03Single-file JAX implementations of PPO, GAIL, AMP and DeepMimic
  • 04Robot-to-robot motion retargeting for existing datasets
  • 05Built-in domain and terrain randomization
  • 06Gymnasium interface

Use cases

  • Benchmarking imitation learning algorithms on locomotion tasks
  • Training humanoid walking policies from motion capture data
  • Running reinforcement learning with custom rewards on legged robot models
  • Comparing motion trajectories with DTW or Fréchet distance metrics

Quick start

  1. 1

    Clone the repository

  2. 2

    Install in editable mode from the repo folder

    pip install -e .
  3. 3

    Optionally install JAX with GPU support

    pip install jax["cuda12"]

Things to know

  • Installation requires cloning the repository rather than a simple PyPI install in the README.
  • GPU use of JAX needs an extra CUDA install step.
  • The MyoSkeleton environment requires accepting a license and downloading the model separately.
  • AMASS datasets must be downloaded and installed separately due to licensing.

In collections

Reinforcement learning frameworks & training environmentsRobotics & embodied AI projects

Topics

#imitation-learning#reinforcement-learning#mujoco#robotics#humanoid#locomotion#jax

FAQ

What is loco-mujoco?
An imitation learning benchmark for whole-body control, with MuJoCo and MJX environments for humanoids, quadrupeds and musculoskeletal human models. LocoMuJoCo is an imitation learning benchmark for whole-body control. It provides a set of locomotion environments covering humanoids, quadrupeds, and (musculo-)skeletal human models, each paired with motion capture datasets such as AMASS, LAFAN1, and a native LocoMuJoCo dataset.
Is loco-mujoco free to use? What license does it have?
loco-mujoco is open source under the MIT license. Check the license terms before using it in commercial products.
Is loco-mujoco still maintained?
Actively maintained. The last commit to loco-mujoco was on Nov 5, 2026.
What can I use loco-mujoco for?
Benchmarking imitation learning algorithms on locomotion tasks Training humanoid walking policies from motion capture data Running reinforcement learning with custom rewards on legged robot models Comparing motion trajectories with DTW or Fréchet distance metrics
How do I install loco-mujoco?
Clone the repository Install in editable mode from the repo folder: pip install -e . Optionally install JAX with GPU support: pip install jax["cuda12"]
Where can I get loco-mujoco?
The source code and installation instructions are on GitHub: https://github.com/robfiras/loco-mujoco

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects