A Reinforcement Learning-Based Soft Actor-Critic Approach for Autonomous Vehicle Collision Avoidance in Dynamic Environment

This study explores how reinforcement learning, specifically the Soft Actor-Critic (SAC) algorithm, can help strike a balance between speed and safety. Using the CARLA driving simulator, we tested how well SAC performs at two different speeds: a moderate 14 m/s and a higher 25 m/s. To improve learning stability, the SAC approach is enhanced with Prioritized Experience Replay (PER) and entropy tuning. The setup involves combining data from multiple sensors, including radar, collision detectors, and lane invasion alerts, into a 12dimensional state input for the agent. A carefully designed reward function guided the vehicle to maintain appropriate speed, avoid collisions, and follow traffic rules. Results showed that SAC worked very effectively at a moderate speed, delivering high rewards and keeping collision rates low. However, at higher speeds, performance dropped because the sensors could not provide enough reaction time, which frequently resulted in safety problems. This research encourages the development of smarter, more adaptive reinforcement learning systems for autonomous driving. The insights gained here may help direct future real-world vehicle deployments as well as simulation-based research.