Mini Autonomous Car Driving Based on 3D Convolutional Neural Networks
Resumo
Autonomous driving applications have become increasingly relevant in the automotive industry due to their potential to enhance safety, efficiency, and user experience. Nevertheless, building reliable systems remains challenging because of the complexity of neural models, long training times, and difficulties in generalization. Mini Autonomous Cars (MACs) provide a practical and cost-effective testbed, enabling rapid prototyping and evaluation of learning-based driving algorithms in scaled-down environments. In this work, we present a methodology for MAC autonomous driving based on RGB-D data and Three-Dimensional Convolutional Neural Networks (3D-CNNs), evaluated against Recurrent Neural Networks (RNNs) in two simulated tracks with distinct levels of complexity. Each model was trained and tested over 30 autonomous laps and assessed by lap time, variability, and track deviations. The results indicate that a reduced 3D-CNN architecture delivers the best performance on the visually complex Mini Monaco track, whereas a GRU-based RNN achieved the fastest times on the simpler Generated track. To further analyze robustness, we conducted an ablation study on the 3D-CNN. Interestingly, the version with fewer convolutional layers outperformed deeper variants, highlighting that reduced complexity can improve stability and generalization in challenging environments. These findings provide practical insights for selecting neural architectures in embedded implementations, balancing performance, robustness, and computational cost. Video link: https://youtu.be/6S2mfHWn3E8
