Autonomous Sidewalk Navigation Featuring End-to-End RGB-D Dual-ConvNet Steering

Abstract

This paper proposes an innovative navigation framework for Autonomous Mobile Robots (AMRs) operating on pedestrian side walks. The proposed method breaks down the navigation mission into discrete sidewalk-following tasks between intermediate waypoints, ensuring a seamless journey from start to finish. Two integral modules collaborate to facilitate the entire process. The global sense module (GloS) integrates GNSS (Global Navigation Satellite System), compass, and odometry to plan the path and track the robot’s location. Concurrently, an end-to-end automatic steering module (ASM) le verages a dual-EfficientNetV2 architecture to in tegrate RGB and depth visions to classify diverse sidewalk scenarios and generate specific local maneuvers to follow the sidewalk and avoid obstacles. Our empirical evaluation demonstrated an accuracy of $98 \%$ of the ASM, with a real-time performance on the robot’s embedded system. By deploying the proposed system on our two-wheeled, self-balancing mobile robot for field tests in city streets, the feasibility of employing a geometrical-model-free frame work for sidewalk navigation was attested.