Autonomous Building Entry Integrating ConvNet Object Detection and Cascaded Navigation Control

Abstract

As Autonomous Mobile Robots (AMRs) play an increasingly crucial role in logistics and service industries, their ability to navigate indoor and outdoor environments becomes paramount. A significant challenge lies in ensuring the effective entry of AMRs into buildings, mainly via the accessible ramps. This paper proposes an innovative method that empowers our two-wheeled self-balancing AMR to localize and navigate accessible ramps using solely RGB inputs for environmental assessment. Our model-free, behavior-reflex navigation approach harnesses deep learning-based object detection to identify and assess critical environmental elements for precise navigation control. A three-stage cascaded navigation strategy is proposed to ensure effective navigation for AMRs approaching the ramp from various directions and poses. Fuzzy mechanisms are de vised to eliminate detection outliers and enhance accuracy, while a data history control scheme manages occasions when features are out of sight. Field experiments validated the effective ness of our approach, demonstrating an average mission accomplishment rate of over 90%.