Back to Ecosystem Pulse
EvaluationProduction Ready

gym-pybullet-drones

by utiasDSL

PyBullet Gymnasium drone environments for single- and multi-agent RL

Python
Updated Feb 6, 2026
Share:
1.8k
Stars
514
Forks

View on GitHub

Summary

Provides PyBullet + Gymnasium environments for single- and multi-agent quadcopter control. Exposes configurable physics, sensors, and reward setups so you can train and evaluate RL agents in realistic simulated flight scenarios, enabling reproducible experiments via the Planning Pattern. Includes multi-agent scenarios and baselines that make it straightforward to reproduce control experiments and failure cases, leveraging the Event-Driven Agent Pattern for coordination.

Key Benefits

As agents move from toy tasks to physical-world action, high-fidelity simulation is essential for assessing safety, failure modes, and coordination between agents. These drone environments let teams test agent behaviors and interaction patterns before real-world deployment, supporting reproducible A2A evaluation and pre-production agent testing. Having a common simulated ground truth helps quantify agent track record and compare control strategies under identical conditions, aligned with Defense in Depth Pattern to improve safety and resilience.

Ideal For

Researchers and engineers who need reproducible drone simulation for training, testing, and evaluating single- or multi-agent control policies, supported by the Agent Protocol for consistent tooling and integration.

Applications

  • Validate multi-agent coordination and collision-avoidance strategies in simulation before deployment
  • Benchmark and compare reinforcement-learning flight controllers using shared environments
  • Reproduce and analyze agent failure modes and robustness under different physics and sensor settings
Works With
gymnasiumpybulletstable-baselines3crazyflie
Topics
betaflightcontrolcrazyfliegymgymnasiummulti-agentpybulletquadcopterquadrotorreinforcement-learning+4 more
Similar Tools
gym-pybulletairsim
Keywords
multi-agentsimulationpre-production agent testingagent failure modes