ARReSVG: Intelligent Multi-AAV Navigation in Partially Observable Spaces Using Adaptive Deep Reinforcement Learning Approach

Unmanned Aerial Vehicles (UAVs) have brought about a revolution in various applications worldwide. Applications such as aerial surveillance, search and rescue, and operations beyond the line of sight rely on the autonomous navigation capabilities of multi-UAVs. This paper introduces a novel framework, Adaptive Rapid Recurrent Stochastic Valued Gradient (ARReSVG), empowered by Deep Reinforcement Learning (DRL) to address navigational challenges faced by UAVs in partially observable spaces. The proposed ARReSVG framework incorporates an adaptive learning approach to address issues related to partial observability-based autonomous navigation effectively. Compared with traditional obstacle avoidance techniques, the proposed navigation framework utilizes a distributed Multi-UAV Consensus Path Planning (MUCP2) mechanism to perform collaborative path planning for multiple UAVs to ensure collision avoidance trajectories. Additionally, the learned value-action policy implemented in the UAV alleviates the computational demands of real-time map-building. The ARReSVG framework leverages a Gated Recurrent Unit (GRU) to reduce computing complexity and accurately approximate the action-value function. The simulation results show that the ARReSVG framework achieves an impressive 94%