COMPARISON OF RL ALGORITHM ADAPTIVITY FOR UAV COLLISION AVOIDANCE

Abstract

This paper investigates the adaptivity of reinforcement learning (RL) algorithms, specifically DQN, PPO, SAC, TD3, DDPG, and A2C, in a simulated three-dimensional environment that mirrors unmanned-aerial-vehicle (UAV) collision-avoidance scenarios. The aim of the study is to compare each algorithm’s ability to generalize acquired knowledge and apply it in new, unpredictable flight conditions.The experimental section consists of two stages. In the first stage, drones are trained in conditions where their trajectories are directed toward one another, with fixed initial speed and position parameters. In the second stage, each algorithm is validated in scenarios that lie outside the training domain, featuring random changes in flight direction and velocity characteristics. This approach makes it possible to assess the stability of model behavior in newly encountered, unforeseen conditions. The study compares the rates of successful collision avoidance, reaction speed, and maneuver quality for each algorithm. The data show that some algorithms, such as SAC and TD3, exhibit higher stability under strong trajectory fluctuations, whereas DQN and PPO may be less robust during unpredictable test phases. A2C and DDPG proved intermediate in both success rate and decision-making time.The proposed research provides valuable information for the practical implementation of autonomous UAV control systems in dynamic and uncertain environments. The findings can be used to develop adaptive controllers capable of rapidly correcting drone behavior when unforeseen situations arise. Future studies should incorporate sensor systems that closely replicate real flight conditions and assess how they affect algorithm adaptivity. This will further enhance the reliability and safety of unmanned systems in real flight missions.