Reinforcement Learning at the Forefront of Robot Path Planning

Abstract-Robots are used in many different domains to accomplish tasks. Irrespective of the type of task assigned, a robot needs to move from one place to another to complete its task. While moving, it needs to avoid obstacles in its path. This means that path planning for the robot is required. There are different methods available for robot path planning. This paper proposes a path planning method that uses reinforcement learning to predict the turning angle of a robot in a customised environment. A turning angle is used by a robot to deviate from its current path to avoid an obstacle. A deep deterministic policy gradient agent is trained in the custom environment that consists of the proposed path learning method. The agent predicts the turning angle, which is used by the environment to update a robot’s position. The environment generates observation and reward, which are given back to the agent for angle prediction and policy update, respectively. The proposed method has been tested in three unknown environments. The results show that a proper obstacle free path was planned. Also, the proposed method was applied to tractor-trailer robotic system. The method was used to generates points and the tractor-trailer robot moved from previous point to the generated point using kinematic equations. The proposed method was evaluated for its performance in 200 different scenarios obtaining a success rate of $79 \%$.