Artificial Potential Field-Enhanced Deep Reinforcement Learning for Motion Planning
Tengyue Wang, Xianhao Chen, Chen Lei, L. Yang · 2024
To address the significant challenge of local motion planning for ground mobile agents in unknown, dense, and dynamic environments, this paper presents a hybrid motion planning approach that merges artificial potential field (APF) and deep reinforcement learning (DRL). The utilization of potential fields enables the characterization of surrounding obstacle distributions and the construction of a reward function. By training with an APF -enhanced reinforcement learning algorithm across three distinct scenarios, the agent acquires the ability to tackle both stationary and moving obstacles, as well as track moving targets. Results demonstrate the practicality and adaptability of the proposed approach in simulations and real-world implementation. Notably, the approach successfully overcomes the inherent limitations commonly associated with APF, further enhancing its effectiveness and applicability.