MuJoCo-Drones-Gym: GPU-Accelerated Simulator for Multi-Drone RL Training

Manan Tayal introduces MuJoCo-Drones-Gym, an open-source, GPU-accelerated multi-drone simulator designed to overcome the trade-off between physical fidelity, multi-agent support, and training throughput in aerial robotics. Built on the MuJoCo physics engine, it supports an arbitrary number of Bitcraze Crazyflie 2.x nano-quadcopters and offers a modular API for selecting physics models (rigid-body MuJoCo, explicit Python dynamics, or combinations including ground effect, blade drag, and inter-drone downwash), action interfaces (per-motor RPMs, collective normalized thrust, velocity setpoints, or PID waypoint commands), and observation spaces (kinematic state vectors, RGB/depth/segmentation cameras, or neighborhood adjacency information). A PettingZoo ParallelEnv wrapper enables drop-in multi-agent reinforcement learning, and the simulator includes seven task environments: single-drone hover, velocity tracking, multi-drone hover, waypoint navigation, formation flight, gate racing, and a generic multi-agent template.

The simulator directly mirrors the capabilities of gym-pybullet-drones but takes advantage of MuJoCo's superior contact handling, rendering quality, and parallelizability—critical for modern deep RL pipelines that demand high throughput. The paper details environment design, underlying quadcopter dynamics, and includes control and learning examples. By enabling GPU-accelerated simulation of hundreds of drones with realistic aerodynamic effects, MuJoCo-Drones-Gym provides a unified platform for developing both classical controllers and RL policies for multi-drone systems. This open-source release (arXiv:2606.08039) is expected to accelerate research in swarm robotics, autonomous navigation, and aerial manipulation.