Visual Local Path Planning Based on Deep Reinforcement Learning

Abstract

Traditional local path planning methods for robots are mostly designed with prior maps, which leads to poor results in navigation combined with visual SLAM. Therefore, most of the traditional robot local path planning methods are designed with prior maps, which leads to poor results in navigation combined with visual SLAM. Therefore, this paper proposes a visual local path planning strategy based on deep reinforcement learning. Firstly, A raster map of the surrounding environment was built based on visual simultaneous localization and mapping (SLAM) technology, and A* algorithm was used to plan the global path. Secondly, a local path planning strategy based on deep reinforcement learning is constructed by comprehensively considering the problems of obstacle avoidance, robot walking efficiency and pose tracking, and discrete action spaces with forward, left and right turning into basic elements are designed, as well as state Spaces based on visual observations such as color map, depth map and feature point map. Using Proximal Policy optimization (PPO) algorithm to learn and explore the best state action mapping network. The results of Habitat simulation platform show that the proposed local path planning strategy can plan an optimal or suboptimal path on the real-time map. Compared with the traditional local path planning algorithm, the average success rate is increased by 53.9%, the loss rate of pose tracking is reduced by 66.5%, and the collision rate is reduced by 30.1%.