Gradient Descent-Based Collision Avoidance for Drone Swarms: A Comparative Analysis of Efficacy and Scalability

This paper presents a new methodology for real-time control of an autonomous drone swarm under constantly changing environments. It examines swarm formation, path planning, task allocation, energy management and collision avoidance. Built on computational planning, decision making, and action selection, new coordination methodologies make it easy for the drones to autonomously work in collaboration. In a decentralized control structure, instead of being controlled by one or two operators, drones fly to designated areas autonomously and can therefore cooperate faster to carry out tasks and efficiently respond to changes in environments. Recent advances in deep learning enabled replacing Dijkstra’s algorithm with a form of real time predictive path planning and avoiding potential risk associated with unintended collisions. In terms of market-based approaches in task allocation of helicopters, the helicopters are assigned self-managed tasks depending on their capabilities, hence making the system very scalable. This method also includes energy friendly algorithms which take into account moving of the drones and their assigned tasks with respect to remaining battery. An integrated approach has been considered for collision avoidance. Preplanned missions are complemented with tactical decisions in real time using collision avoidance tactics. Reinforcement learning validates the established decisions.