Deep reinforcement learning-based local path planning in dynamic environments for mobile robot☆

Path planning for robots in dynamic environments is a challenging task, as it requires balancing obstacle avoidance, trajectory smoothness, and path length during real-time planning.This paper proposes an algorithm called Adaptive Soft Actor-Critic (ASAC), which combines the Soft Actor-Critic (SAC) algorithm, tile coding, and the Dynamic Window Approach (DWA) to enhance path planning capabilities. ASAC leverages SAC with an automatic entropy adjustment mechanism to balance exploration and exploitation, integrates tile coding for improved feature representation, and utilizes DWA to define the action space through parameters such as target heading, obstacle distance, and velocity In this framework, the action space is defined by DWA's three weighting parameters: target heading deviation, distance to the nearest obstacle, and velocity. To facilitate the learning process, a non-sparse reward function is designed, incorporating factors such as Time- to-Collision (TTC), heading, and velocity. To validate the effectiveness of the algorithm, experiments were conducted in four different environments, and the algorithm was evaluated based on metrics such as trajectory deviation, smoothness, and time to reach the end point. The results demonstrate that ASAC outperforms existing algorithms in terms of trajectory smoothness, arrival time, and overall adaptability across various scenarios, effectively enabling path planning in dynamic environments.