【摘要】 为解决未知环境中移动机器人的自主路径规划问题，提出了一种基于深度确定性策略梯度(DDPG)的改进算法。该算法通过激光雷达数据、机器人位姿以及速度信息建立策略映射，在连续动作域中输出线、角速度直接控制机器人底盘运动。设计了新的连续奖惩函数，缓解了奖励稀疏问题；融合了优先经验回放机制、人工演示数据机制以及多控制器引导机制，提高了算法训练效率。通过ROS+Gazebo平台进行模型训练，训练结果表明，改进算法仅需原始算法训练步数的47%,就获得了相同水平的奖励；设计对比实验，结果表明，相较于原始算法和传统的局部路径规划动态窗口法，改进算法在无碰撞的基础上运动轨迹更加平滑且耗时更短，验证了改进算法的有效性。最后搭建轮式差速机器人平台，设计未知环境导航实验，证明了算法的实用性。更多还原

【Abstract】 To solve the problem of autonomous path planning for mobile robots in unknown environments, an improved algorithm based on deep deterministic policy gradient(DDPG) was proposed. Based on the lidar data, robot position and speed information, the strategy mapping is established, and the line and angular velocity are output in the continuous action domain to directly control the motion of the robot chassis. A new continuous reward and punishment function is designed to alleviate the problem of reward sparsity. The priority experience replay mechanism, manual demonstration data mechanism and multi-controller guidance mechanism are introduced to improve the training efficiency of the algorithm. The model was trained by ROS+Gazebo platform, and the training results showed that the improved algorithm only needed 47% of the training steps of the original algorithm to obtain the same level of reward. Compared with the original algorithm and the traditional dynamic window approach, the results show that the improved algorithm has smoother trajectory and shorter time on the basis of non-collision, which verifies the effectiveness of the improved algorithm. Finally, the wheeled differential robot platform is built, and the navigation experiment in unknown environment is designed, which proves the practicability of the algorithm.更多还原