Walter Iran Amador Valenzuela

Autonomous driving remains a challenging problem in mobile robotics due to the need for reliable perception, decision making, and control under limited computational resources. This paper presents a lightweight hybrid vision framework for autonomous driving in the FIRA Urban Challenge using the AGILEX LIMO platform. The proposed system combines classical computer vision techniques for lane and obstacle detection with a YOLOv8n-based traffic sign recognition module. A hierarchical finite state machine architecture coordinates line following, junction handling, traffic sign interpretation, and obstacle avoidance, while a proportional-derivative controller generates real-time steering commands. To evaluate the suitability of learning-based perception for embedded robotic platforms, a comparative study between AprilTag and YOLO-based traffic sign recognition was conducted. Experimental results demonstrate that although AprilTags provide superior geometric accuracy, the YOLO-based approach offers greater robustness under partial occlusion, varying illumination conditions, and non-ideal viewing angles. Operating entirely on onboard CPU resources using only the RGB stream from an RGB-D camera, the proposed framework achieved an average processing rate of 14.5 FPS while successfully performing autonomous navigation tasks in FIRA-style environments. The results demonstrate that combining lightweight deep learning with classical vision methods provides an effective solution for embedded autonomous driving applications.

Walter Iran Amador Valenzuela, Bo Zhou, Meng Cheng Lau Verified by AVIS CertificateFIRA World Summit 2026Submitted 17 Sept 2026