Soft Actor-Critic Combining Potential Field for Global Path Planning of Autonomous Mobile Robot

Global path planning is a critical technology in the field of autonomous mobile robot navigation. Serving as the upper-layer component of path planning, it provides the global reference path for the local trajectory planning module. However, the majority of conventional methods focus solely on optimizing path length, which can lead to redundant obstacle avoidance maneuvers by the lower-layer local planner or even planning failure. Furthermore, graph-searching methods commonly suffer from prolonged computation times and low efficiency. To address these challenges, this paper proposed a global path planning method based on deep reinforcement learning that integrates artificial potential fields. The method expanded the network structure of Soft Actor-Critic (SAC) by employing the constructed potential field to conduct supervised learning on two additional critic networks. Subsequently, the predicted values from the critic network were integrated into the actor network to guide agents in choosing states with smaller potential field values. Additionally, to mitigate the time cost of retraining due to changes in the global environment, a risk assessment module employing Monte Carlo random sampling was incorporated. The computed risk value was subsequently integrated into the network as the new state. Experimental results show that our method reduces computation time by 38.64% compared to conventional methods. The convergence is 40.48% faster and the path potential value is 95.72% lower than other reinforcement learning methods.