Efficient multi-agent deep reinforcement learning algorithm for multi UAV collision avoidance

The rapid expansion of uncrewed aerial vehicles (UAVs) across industries has led to increased airspace congestion. The increasing use of drones across many fields and locations has caused serious problems, especially in avoiding collisions. In the rapidly developing field of drone technology, ensuring UAV flight safety and reducing the risk of UAV collisions have therefore become urgent concerns. There are many artificial intelligence (AI) algorithms designed to solve this problem, but most work only in situations with a single agent. Multi-agent reinforcement learning is a promising way to solve these problems. It enables drones to operate with greater intelligence and flexibility, even in challenging situations, alongside other agents. This work presents a Multi-Agent Deep Reinforcement Learning Algorithm based on an efficient graph attention network for collision avoidance in a dense, complex multi-agent drone environment. We propose both curriculum learning and transfer learning by adding more agents over time and subsequently employing learning models. This makes the system more scalable and more coordinated. The training process is significantly advanced by the suggested method, which outperforms the current benchmark in continuous settings. Our findings indicate that the proposed approach achieves 17% higher cumulative reward, up to 10% fewer loss-of-separation time steps, and about 44% fewer active interaction edges than the benchmark. Furthermore, the proposed method reduces action-selection bias, improving decision-making stability in dense multi-UAV settings. • Suggest an Efficient Multi-Agent Deep Reinforcement Learning Algorithm for UAV collision avoidance in dynamic environments. • Propose an efficient graph attention network architecture to model UAV interactions efficiently. • Demonstrate improved scalability with increasing numbers of UAV agents based on the suggested architecture and curriculum learning.