crazyMARL

crazyMARL is a framework for multi-agent reinforcement learning experiments with Crazyflie quadcopters, built on top of JaxMARL. It provides research-grade implementations of popular algorithms, experiment management, and visualization utilities for simulating and training cooperative multi-quadrotor systems in a performant and flexible way.

Features

• Multi-agent RL algorithms: IPPO, MAPPO (compatible with JaxMARL)
• JAX and MuJoCo backends for fast simulations
• Batched and single-environment rollouts, video rendering and plotting tools
• Experiment tracking with ASDF format and Weights & Biases integration
• Docker-based development environment for reproducible setups

Trained Policies

We provide several trained IPPO policies for different scenarios:

• single_quad_no_payload: Single quadcopter trained to recover from harsh conditions and track position setpoint
• single_quad_payload: Single quadcopter trained to carry a payload and track position setpoint with payload
• two_quad: Two quadcopters trained to cooperatively track a position setpoint with payload
• three_quad: Three quadcopters trained to cooperatively track a position setpoint with payload

The results can be seen in these videos:

Installation

Clone and install:

git clone https://github.com/viktorlorentz/crazyMARL.git
cd crazyMARL
pip install -e .[algs,dev]

Docker (optional):

./build.sh    # Build Docker image with or without CUDA support
./run.sh      # Run interactive shell in the container

Quick Start

Training

Launch training with a predefined configuration:

python ./crazymarl/train/train.py --config your_config_name

Or specify a custom YAML file:

python ./crazymarl/train/train.py --config-file path/to/config.yaml

• Available configs live in crazymarl/train/configs:
  • single_quad
  • two_quad
  • three_quad
  • six_quad
  • test

Flight Experiments & Rollouts

Run flight experiments and render videos:

python ./crazymarl/experiments/fly.py --config figure_eight --model-path path/to/model.tflite

Experiment YAML files are in crazymarl/experiments/configs.