Obstacle avoidance parking based on improved deep reinforcement learning SAC algorithm

초록

Recent research uses neural-network-based approaches to generate time-optimized parking trajectories in linear time. However, the generalization of these neural networks in different parking lot scenarios is not fully considered and relies on high-quality data. To address these issues, this paper proposes an improved soft actor critic (SAC)based parking trajectory planning method to achieve fast convergence and high success rate. A trajectory data-based reward function and an experience replay sampling strategy are designed simultaneously improve the learning efficiency. Meanwhile, the Flatten-T Swish Plus (FTSPlus) activation function is firstly introduced into the SAC's neural network structure to enhance the convergence capabilities. The simulation results show that compared with traditional SAC methods, the success rate has been greatly improved and convergence rate has also been improved by more than 60%. Furthermore, it has the good model generalization ability while maintaining collision-free trajectories. © 2024 IEEE.

키워드