Path Planning Algorithm for Two-dimensional Raster Maps based on Deep Reinforcement Learning

Abstract

Path planning is one of the main research contents of decision control. The problem of path planning can be described as finding a collision free path from the initial state to the target state in the environment. Raster-based path planning is a basic problem in robotics, games and other fields. Traditional path planning methods need to model the environment in advance, and then carry out path planning on the basis of known map information. However, there is a lot of dynamic information in the environment, such as pedestrians, movable objects. It is difficult for traditional algorithms to carry out path planning in the dynamic unknown environment. With the development of deep learning and reinforcement learning, such methods provide a new way to solve the problem of path planning. Aiming at resolving the shortcomings of traditional path planning algorithms, this paper adopts deep reinforcement learning algorithm to carry out path planning for two-dimensional raster maps, and finally validates and evaluates the performance of the algorithm within an efficient simulation environment.