Multi-Robot RL

A research codebase for multi-robot reinforcement learning in MuJoCo, built on mjlab. It defines three tasks where a variable number of robots (masspoints or UR10e arms) must jointly solve a shared objective, trained with PPO or FastSAC.

Tasks

Task	Description
reach	Robots must move their end-effectors to a set of unassigned 3D goal positions.
push	Robots must push cuboids along a surface to unassigned target poses (XY position + yaw).
type	Robots must press a sequence of highlighted keys on a physical keyboard.

Installation

This project requires uv, which can be installed with:

curl -LsSf https://astral.sh/uv/install.sh | sh

The following commands clone the repo, install the dependencies and visualize a task with a random policy. Robot counts and other parameters can be changed in the configs directory.

git clone git@github.com:MarvStein/multi-robot-rl.git
cd multi-robot-rl
uv sync
uv run play <reach/push/type> --agent random

For more details on training, rollouts, recordings and more checkout the documentation below.

Documentation

• Getting Started
• API Reference