Indoor Obstacle Avoidance System Design and Evaluation Using Deep Learning and SLAM-Based Approaches

We aim in this paper to design a robust and cost-efficient system specifically targeted at indoor mobile robot obstacle avoidance. The system incorporates the fusion of 2D LiDAR-based Simultaneous Localization and Mapping (SLAM) with a Rapidly Exploring Random Trees (RRT) algorithm for effective path planning. Furthermore, we propose an innovative and pioneering approach for obstacle avoidance based on deep learning. The deep learning model is trained using data collected from a simulated environment utilizing a 2D LiDAR sensor, which serves both for SLAM and data acquisition purposes. This paper includes a comparative analysis of odometry-based and SLAM-based pose computation methods, providing valuable insights into the data collection and training procedures essential for successful deep learning-based obstacle avoidance. The implementation is carried out within the Robot Operating System 2 (ROS2) interface.