Goal Decision-making in Active SLAM Using 2D Lidar Based on DRL Driven by Intrinsic Rewards

摘要

Utilizing LiDAR as a sensor and collecting point cloud data for simultaneous localization and mapping (SLAM) remains the predominant mapping approach, widely applied in various industrial and service scenarios. In laser SLAM, the algorithm estimates the front-end odometer through raw LiDAR data, facilitating the construction of a map environment. Active SLAM is an essential technique in mobile robotics that enables autonomous robot operation, with its primary objective being to autonomously construct a map of an unknown environment without human intervention. Given the absence of human guidance, active SLAM relies on actively perceiving environmental information through LiDAR sensors. After sensing environmental information through LiDAR sensors, active SLAM also needs to implement SLAM, decision-making, and navigation, where the decision-making module typically determines the next goal to facilitate map construction using deep reinforcement learning(DRL). Despite offering greater flexibility than traditional methods, most DRL-based decision-making approaches suffer from low efficiency or slow convergence due to inadequate consideration of action space or sparse rewards. Hence, to address these challenges, we propose employing 2D LiDAR as the environmental perception sensor and introducing a goal decision-making model based on DRL driven by intrinsic reward. Specifically, we consider all cells on the occupancy grid map as potential goals to overcome limitations caused by solely relying on frontiers in previous studies. Additionally, we utilize Shannon entropy reduction within a fixed range centered around each goal as an intrinsic reward mechanism to motivate the robot’s navigation towards areas with higher information entropy. We validate our proposed method across different environments and demonstrate superior performance in terms of achieving higher rewards within shorter distances and time consumption.