基于碰撞概率与速度障碍的深度强化学习安全导航研究

Research on Deep Reinforcement Learning Navigation with Safety Shield

摘要:

针对移动机器人在复杂多动态障碍物环境中导航效率低的问题，提出了一种基于碰撞概率和速度障碍的深度强化学习导航算法；为了保证导航策略的安全性，基于控制障碍函数设计了一个安全屏障用来调整动作；定义碰撞概率估计函数，用来评估障碍物发生碰撞的风险，将高风险的关键障碍物信息纳入深度强化学习算法的状态空间，减少特征提取时间；引入速度障碍理论设计了一种引导机器人针对关键障碍物进行主动避障的奖励函数，降低了机器人寻找最优航向角的时间；训练所得策略在不同环境下的测试结果验证该算法实现了安全快速导航。

Abstract:

To address the problem of low navigation efficiency of mobile robot in a complex environment with multiple dynamic obstacles, a deep reinforcement learning navigation algorithm based on collision probability and velocity obstacles is proposed; A safety shield is designed based on the control barrier function to adjust the action in order to ensure the safety of the navigation policy. The collision probability estimation function is defined to evaluate the risk of collision between obstacles, and the critical obstacles information with high risk is incorporated into the state space of the deep reinforcement learning algorithm to reduce the time of feature extraction; The velocity obstacles theory is introduced to design a reward function to guide the robot to actively avoid the critical obstacles; This reduces the time for the robot to find the optimal heading angle. The test results of the training policy in different environments verify that the algorithm realizes safe and fast navigation.