Overview

An IsaacLab DirectEnv for quadrupedal locomotion under motor failures, with support for multiple quadruped robots, sim-to-sim, and sim-to-real pipelines.

Features:

• Concurrent State Estimator
• Rapid Motor Adaptation
• Mixture-of-Experts
• Morphological Symmetries
• Sim-to-Sim in Mujoco
• Sim-to-Real in ROS2

Real-world deployment via:

• muse for state estimation (if no concurrent state estimation is used)
• unitree-ros2-dls for unitree robot communication

A list of robots and environments available are described below:

Robot Model	Environment Name Pattern
Aliengo, Go2	FaultLocomotion-RobotModel-Flat-Blind FaultLocomotion-RobotModel-Rough-Blind FaultLocomotion-RobotModel-Rough-Vision

Installation and Runs

If you want only to deploy a trained policy on your robot, continue on README_deploy otherwise on README_train.

For the train, check first the compatibility with IsaacLab and rsl-rl at the top of this readme. They indicate the releases that we tested.

Citing this work

If you find the work useful, please consider citing one of our works:

Mixture-of-Experts RL for Fault-Tolerant Legged Locomotion (ArXiv)

@inproceedings{turrisi2026moefault,
  author={Turrisi, Giulio and Pali, Ozan and Oneto, Luca and Semini, Claudio},
  booktitle={arXiv}, 
  title={Mixture-of-Experts RL for Fault-Tolerant Legged Locomotion}, 
  year={2026},
  doi={arXiv:2606.25965}
}

Maintainer

This repository is maintained by Giulio Turrisi