好奇心驱动的深度强化学习机器人路径规划算法

针对采用深度强化学习算法实现机器人路径规划任务中,训练前期随机性高导致奖励难获取问题,提出内在好奇心驱动的深度确定性策略梯度算法对连续型动作输出的端到端机器人路径规划进行研究。将环境获取的感知信息作为输入状态,输出机器人动作(线速度、角速度)的连续型控制量,在Gazebo仿真平台进行训练并验证。实验结果表明,基于内在好奇心驱动的深度确定性策略梯度路径规划算法可以较好地实现端到端的机器人路径规划,并且有利于解决训练前期奖励难获取问题,与离散型动作输出的深度Q学习网络模型进行了对比分析,结果表明本文算法决策控制效果更优越。在真实环境中进行了验证,在静态障碍和动态障碍的场景下,所提出算法可成功到达目标点。

A Robot Path Planning Algorithm Based on Curiosity-driven Deep Reinforcement Learning

In early training phase of robot path planning, deep reinforcement learning will cause reward difficult to obtain. To reduce training time, an intrinsic curiosity deep deterministic strategy gradient (ICDDPG) algorithm is proposed on end-to-end robot path planning of continuous action output. Environment information of perception as input, the output is robot motion (linear velocity and angular velocity) continuous control. Train and validate in the Gazebo simulation platform. The simulation results show ICDDPG is helpful to solve the problem of reward difficult to obtain, and the proposed algorithm has better control strategy compared with deep Q-learning networks. It is verified in a real environment, and the proposed algorithm can successfully reach the target points under static and dynamic obstacles.