Dashing for the Golden Snitch: Multi-Drone RL

0. Introduction

• A multi-agent environment for time-optimal motion planning. This repository presents a decentralized policy network for time-optimal multi-drone flight using multi-agent reinforcement learning.
• This project is a reimplementation of gym-pybullet-drones, optimized for multi-agent scenarios. We have adjusted the code to make it more suitable for handling a large number of agents simultaneously.
• We customize PPO in a centralized training, decentralized execution (CTDE) fashion, based on stable-baselines3 and inspired by the on-policy(MAPPO) repository.

Demonstration Video

Real-world experiments with two quadrotors using the same network achieve a maximum speed of 13.65 m/s and a maximum body rate of 13.4 rad/s in a 5.5 m x 5.5 m x 2.0 m space across various tracks, relying entirely on onboard computation.

Related Papers

• Dashing for the Golden Snitch: Multi-Drone Time-Optimal Motion Planning with Multi-Agent Reinforcement Learning, Wang, X., Zhou, J., Feng, Y., Mei, J., Chen, J., & Li, S. (2024), arXiv preprint arXiv:2409.16720.

The public and release version is coming soon...