Enhanced Hallway Navigation Featuring End-to-End RGB-D Quadruple-ConvNet Steering

This paper presents a novel end-to-end RGB-D-based navigation system for Autonomous Mobile Robots (AMRs), integrating a quadruple Convolutional Neural Network (ConvNet) architecture to achieve enhanced navigation accuracy, adaptability, and robustness in dynamic and unstructured environments. The proposed system leverages dual RGB-D cameras to expand the horizontal field of view to 167°, significantly improving cornering success rates and obstacle avoidance. Unlike traditional hierarchical task frameworks that require multiple specialized models, our integrated ConvNet architecture enables multi-task learning within a single network, optimizing computational efficiency. Additionally, a novel robot maneuver control scheme, incorporating compass and odometry data, ensures smoother and more precise navigation trajectories. Field experiments conducted on our two-wheeled self-balancing robot, J4.Deliver, demonstrate 100% navigation accuracy and successful loop closures, highlighting the system’s robustness. This work provides a scalable and computationally efficient framework for real-world AMR deployment, requiring only low-complexity topological maps and minimal onboard computation, making it well-suited for complex, real-world applications.