ML-Based Perception and Control for Semi-Autonomous Electric Buggy in Smart Campus Environments

This paper presents the development of an electric buggy integrated with machine learning capabilities to enable semi-autonomous navigation within a university campus environment. Using a combination of NVIDIA Jetson Nano, a monocular RGB camera, and a LiDAR sensor, the system enables real-time visual perception, depth estimation, and motion control. A lightweight duo of neural models—multi-task perception using YOLOP and depth estimation using MiDaS—are TensorRT-optimized for embedded inference so that the vehicle can identify lanes, drivable zones, and obstacles in motion. A customized university Campus Driving Dataset has been complied under adverse weather and lighting conditions such as fog and night time and varied to ensure strong model generalizability. The complete control system runs using power supplied by the onboard battery of the buggy via a buck converter and creates a fully self-contained system. System level evaluations were performed using a 1.88 km campus way-point route and showed stable path tracking for average lateral error of 0.15 m, end-to-end system latency of about 90 ms, and stable obstacle avoidance. The results show that affordable, embedded AI can be used to create reliable, efficient, and sustainable campus-scale transportation solutions.