Graph Attention-Driven Distributional Deep Reinforcement Learning for Collaborative Multi-AAV Navigation
The rapid growth of Autonomous Aerial Vehicles (AAVs) has underscored their importance across commercial and civil applications. Combining multi-AAVs with Deep Reinforcement Learning (DRL) approaches can achieve significant results, particularly in navigation and path-planning operations. Nevertheless, collaborative decision-making among AAV swarms faces major challenges, including handling partial observability and re-path planning under uncertainty. To address these challenges, we propose an Adaptive Distributional deep Q-network Graph ATtention (ADGAT) learning framework with a Partially Observable Markov Decision Process (POMDP) modeling approach to perform effective AAV swarm navigation. The proposed ADGAT framework combines an adaptive Distributional Deep Q-Network (Di-DQN) reinforcement learning algorithm with a multi-head Graph Attention (GAT) mechanism to enable effective decision-making by distributing rewards rather than estimating the expected value. Additionally, the GAT mechanism enhances each AAV's capabilities by aggregating information from neighboring AAVs. We compare the proposed ADGAT framework against state-of-the-art baselines and validate it using three key performance metrics, achieving an average success rate of 98%, a collision rate of 3%, and a path efficiency of 96%. The extensive simulation results prove the adaptability of the ADGAT framework for enhancing AAV swarm navigation in partially observable spaces.