This repository provides a benchmark for continuous multi-agent robotic control, extending OpenAI's Mujoco Gym environments. It is designed for researchers and practitioners in multi-agent reinforcement learning (MARL), offering a standardized platform for evaluating decentralized cooperative control algorithms. The primary benefit is a collection of pre-configured multi-agent scenarios based on popular robotic simulations.

How It Works

The library implements multi-agent configurations by partitioning single-agent Mujoco environments. It supports customizable agent observations based on proximity (agent_obsk) and specific properties (k_categories, global_categories). This approach allows for flexible and scalable MARL experiments, enabling agents to perceive their local or global environment state as needed for cooperative tasks.

Quick Start & Requirements

• Install by cloning the repository and adding ./src to your PYTHONPATH.
• Requires OpenAI Gym version 0.10.8 and Mujoco 2.1.
• Set LD_LIBRARY_PATH to ~/.mujoco/mujoco210/bin and LD_PRELOAD to /usr/lib/x86_64-linux-gnu/libGLEW.so for rendering.
• Official documentation: https://robotics.farama.org/

Highlighted Details

• Supports various multi-agent configurations for Ant, HalfCheetah, Hopper, Humanoid, Reacher, Swimmer, and custom coupled scenarios.
• Introduces a "coupled_half_cheetah" environment with agents linked by an elastic tendon.
• Offers configurable observation spaces based on agent proximity and property visibility.
• Includes a detailed example demonstrating environment interaction and rendering.

Maintenance & Community

This repository is a fork of OpenAI's original Mujoco Gym environments. The README notes that a maintained version with fixes and broader support is available in Gymnasium Robotics (https://github.com/Farama-Foundation/Gymnasium-Robotics).

Licensing & Compatibility

The README does not explicitly state a license. However, it is based on OpenAI Gym, which was typically released under the MIT license. Compatibility with commercial or closed-source projects would require explicit license confirmation.

Limitations & Caveats

The project requires specific, older versions of OpenAI Gym (0.10.8) and Mujoco (2.1), which may pose installation challenges. The README points to Gymnasium Robotics as a more actively maintained and compatible alternative.