Autonomous Navigation of Agricultural Robots Utilizing Path Points and Deep Reinforcement Learning

摘要

When using deep reinforcement learning algorithms for reactive navigation, it is common to encounter situations where the agent gets stuck in a local optimum. In autonomous exploration without any prior knowledge, the exploration of environments is often suboptimal. Moreover, the agricultural robots performing tasks in farm fields face even more great challenges. To address this, we propose an Agricultural Robot Autonomous Exploration System (ARAES) that uses deep reinforcement learning algorithms to drive the robot to explore unknown farm environments. The whole system consists of a Global Path Planning (GPP) module and a Local Path Planning (LPP) module. The GPP module utilizes path points acquired along the heading direction in the environment and selects waypoints based on the collected path points to guide the robot towards the global goal, effectively alleviating local optima issues in reactive navigation. The LPP module learns local navigation motion policies in a simulated environment. It incorporates the learned policies into the motion planning stack, enabling the robot to move towards global goals based on the path points and way points. The advantages of the ARAES system are its autonomy and real-time performance, allowing the robot to operate without needing to know detailed prior information about the environment and enabling real-time adjustments of navigation policies according to current environments. Meanwhile, ARAES can construct maps in real-time as the robot moves in the environment. Experiments demonstrate that ARAES is advantageous because it does not rely on complex static and dynamic environment maps or prior information and delivers superior performance compared to other methods.