Comparison of Different Controller Architectures for Autonomous Driving and Recommendations for Robust and Safe Implementations

This comprehensive review examines various controller architectures for autonomous driving systems, from rule-based approaches to advanced deep learning methods. Research trends reveal a significant shift toward deep learning approaches (65.6%) compared to rule-based methods (34.4%), reflecting the growing dominance of data-driven techniques in autonomous vehicle research. Performance analysis of transformer-based models demonstrates exceptional accuracy, with ViT-SAC achieving 100% success rate in low-density traffic scenarios and DRLNDT reaching 99.9% success rate in navigation tasks. Temporal reasoning capabilities assessment shows BEVWorld excelling in context maintenance and historical data integration (both 95/100), while Holistic Transformer demonstrates superior noise robustness (95/100). Computational efficiency varies significantly, with VCNN (38.50 FPS) and DSCNN Transformer (34.07 FPS) exceeding real-time thresholds, while complex BEV architectures like BEVSegformer (3.97 FPS) require further optimization. Simulation platform comparison identifies CARLA as the most comprehensive environment, supporting five of seven key testing features, though no single platform provides complete coverage of all requirements. Technical challenges assessment quantifies real-time processing requirements as the most critical challenge (90/100), followed by generalization limitations (85/100). These suggest that while rule-based approaches offer computational efficiency and interpretability, deep learning methods demonstrate superior perception and decision-making capabilities. A balanced combination of learning-based, rule-based, and simulation-based validation approaches, with particular emphasis on addressing real-time performance and generalization capabilities, will likely be necessary to achieve reliable autonomous driving systems capable of navigating complex and dynamic environments.