Multi-Agent Reinforcement Learning for Zero-Shot Coverage Path Planning with Dynamic UAV Networks

Recent advancements in autonomous systems have enabled the development of intelligent multi-robot systems for dynamic environments. Unmanned Aerial Vehicles play an important role in multi-robot applications such as precision agriculture, search-and-rescue, and wildfire monitoring, all of which rely on solving the coverage path planning problem. While Multi-Agent Coverage Path Planning approaches have shown potential, many existing methods lack the scalability and adaptability needed for diverse and dynamic scenarios. This paper presents a decentralized Multi-Agent Coverage Path Planning framework based on Multi-Agent Reinforcement Learning with parameter sharing and Centralized Training with Decentralized Execution. The framework incorporates a customized Rainbow Deep-Q Network, a size-invariant reward function, and a robustness and safety filter to ensure completeness and reliability in dynamic environments. Our training pipeline combines curriculum learning, domain randomization, and transfer learning, enabling the model to generalize to unseen scenarios. We demonstrate zero-shot generalization on scenarios with significantly larger maps, an increased number of obstacles, and a varying number of agents compared to what is seen during training. Furthermore, the models can also adapt to more structured maps and handle different tasks, such as search-and-rescue, without the need for retraining.