【摘要】 针对城市泊车问题，泊车机器人应运而生，其路径规划是重要的研究方向。由于A*算法的局限性，本文引入深度强化学习思想，并对由此发展起来的D3QN算法进行改进，将残差网络取代卷积网络，引入注意力机制，从而提出SERD3QN算法，以改善网络退化现象和提高收敛速度，并提升模型的精准率。在算法训练过程中，改进奖惩机制，以实现最优方案的快速收敛。通过与D3QN算法、增加残差层的RD3QN算法的对比实验，结果表明本文提出的SE-RD3QN算法在模型训练时可实现更快的收敛速度。与目前常用的A*+TEB算法的对比实验，结果表明本文算法在路径规划时可获得更短的路径长度与规划时间。最后通过模拟小车的实物实验，进一步验证了算法的有效性。这为停车路径规划提供了新的解决方案。更多还原

【Abstract】 The parking robot emerges as a solution to the urban parking problem,and its path planning is an important research direction. Due to the limitations of the A* algorithm,the deep reinforcement learning idea is introduced in this article,and improves the D3QN algorithm. Through replacing the convolutional network with a residual network and introducing attention mechanisms,the SE-RD3QN algorithm is proposed to improve network degradation and convergence speed,and enhance model accuracy. During the algorithm training process,the reward and punishment mechanism is improved to achieve rapid convergence of the optimal solution. Through comparing the experimental results of the D3QN algorithm and the RD3QN algorithm with added residual layers,it shows that the SE-RD3QN algorithm achieves faster convergence during model training. Compared with the currently used A*+TEB algorithm,SE-RD3QN can obtain shorter path length and planning time in path planning. Finally,the effectiveness of the algorithm is further verified through physical experiments simulating a car,which provides a new solution for parking path planning.更多还原