Cooperative Collision Avoidance of UAV Swarms Using Reinforcement Learning

In recent years, making UAV swarms avoid obstacles in real-time has become a big challenge, especially in complex and changing environments. Traditional methods like the Velocity Obstacle (VO) approach are fast and simple, but they only consider the current step, sacrificing long-term optimality. This may lead to suboptimal trajectories in crowded or dynamic environments. To address this limitation, this project combines VO with Reinforcement Learning (RL). VO handles the quick reactions needed to avoid crashes, while RL helps with long-term planning. To validate the effectiveness of this method, a customized 2D simulation environment is built. Experiment results demonstrate that the VO algorithm gives a basic safe velocity, and then the RL model further tunes this velocity to get superior long-term performances.