Mobile Robot Path Planning Based on Improved Deep Reinforcement Learning Algorithm

To address inefficiency and high collision rates in existing path planning algorithms for mobile robots in complex scenarios, this paper introduces a novel algorithm, TD3-KDHER, which combines Twin Delayed Deep Deterministic Policy Gradient (TD3) with an enhanced version of Hindsight Experience Replay (HER). The core innovation of TD3-KDHER is integrating Kernel Density Estimation (KDE) to utilize the experience buffer for fitting a probability density model, thereby optimizing the selection of pending experiences and new goals. This enhances the likelihood of selecting rare but high-value samples. Additionally, the algorithm incorporates adaptive action noise and action masking strategies to refine the exploration strategy of the agent. Experimental results show that TD3-KDHER outperforms both TD3 and TD3-HER across multiple metrics in various scenarios.