Abstract
This paper presents the design and implementation of a low-cost autonomous vehicle prototype that employs machine learning (ML) and computer vision techniques. The prototype integrates a ResNet-18 convolutional neural network (CNN), trained using PyTorch and deployed on an NVIDIA Jetson Nano-a compact, GPU-accelerated computing platform based on the JetBot architecture. The objective is to demonstrate the feasibility of implementing ML algorithms on affordable embedded hardware for real-time navigation tasks such as lane detection and path following. The development workflow includes image data acquisition, model training, and validation through physical testing on a $2.44 ~\mathrm{m} \times 1.24 ~\mathrm{m}$ modular track. More than 2,600 labeled images were used to train the model across five $t$ rack $s$ cenarios, $w$ ith $t$ he $b$ est $c$ onfiguration ob tained after 60 epochs. Experimental results show that the trained ResNet-18 model achieved up to 70 % trajectory-tracking accuracy and a continuous navigation distance of 208 cm before deviation. These results confirm t hat the J etson N ano p latform c an effectively execute CNN-based control tasks in real time. This study contributes to scalable and cost-effective autonomous vehicle research by bridging the gap between theoretical ML applications and practical embedded implementations in resource-constrained environments.