Parallelized Simulation Platform for Multi-UAV Aerial Confrontation with Reinforcement Learning

To address the issue of the quadratic growth in simulation time with increasing numbers of Unmanned Aerial Vehicles (UAVs) for multi-UAV combat mission simulations on Central Processing Units (CPUs), a lightweight, GPU-parallel accelerated simulation environment for multi-UAV combat is designed and developed. Unlike existing simulations frameworks that are often limited to single-UAV mis sions or only parallelize multi-UAV dynamics, the proposed framework, focusing on multi-UAV adversar ial scenarios, further implements parallel computation for the relative poses and damage relationships of each UAV within the environment. In response to the limitation that traditional Basic Fighter Maneuvers (BFM) are only well-defined in level flight, an improved set of fully-attitude interpretable basic aerial combat maneuvers is designed. The simulation environment includes a variety of scalable multi-UAV con frontation tasks, which can provide heuristic baseline strategies for various tasks and support proximal policy optimization-based baseline strategies for one-on-one confrontation scenarios. Compared with the multithreaded simulations running on an 8-core, the 16-thread CPU, the proposed platform improves sampling speed by two to three orders of magnitude. The reinforcement learning training results in both CPU and GPU parallel environments demonstrate that the GPU-accelerated simulation can reduce training time to 1/50, without significantly compromising sampling efficiency due to trajectory fragmentation. The pro posed parallel-accelerated multi-UAV combat simulation environment significantly enhances sampling and training speeds, thereby facilitating accelerated research progress in this field.