A Comprehensive Review of Machine Learning Approaches for Decision-Making in Autonomous Vehicles Collision Avoidance Systems

Abstract

Autonomous vehicles (AVs) hold the potential to reduce road accidents statistics through intelligent collision avoidance systems. This paper presents a comprehensive review of machine learning (ML) approaches for decision-making in AV collision avoidance. The comprehensive review is systematically categorized into four categories which are supervised learning (SL), hybrid learning (HL), deep learning (DL), and reinforcement learning (RL). Each category is evaluated in terms of its applicability in decision-making tasks, which include perception for situational awareness task, behavioral-level decision-making task, motion planning task, and multi-agent coordination task. The review highlight that SL methods offer reliable performance in structured scenarios but have constraints under real-world uncertainties. For HL, the methods utilized enable for feature selection and risk prediction improvement but still face challenge of computational overhead. Meanwhile, DL models advance in perception and control enhancement through spatial-temporal reasoning but encounter large-scale data and high processing power demand issue. As for RL, the methods demonstrate advantages of strong adaptability and scalability, supporting multi-agent coordination and decision-making under uncertainty. However, they still face challenges in training efficiency and real-time responsiveness. Based on the findings, the results offer a thorough guide for creating ML-enabled collision avoidance frameworks that are secure, scalable, and reliable for autonomous driving applications. This review also emphasizes existing research challenges such as computational inefficiency, restricted generalization, sensor blockage, and ethical uncertainty in unavoidable collision situations. The review concludes by stating the future research directions toward hybrid architectures, lightweight and edge-deployable models, uncertainty-aware reasoning deployment, and ethically integrated decision-making frameworks. This review provides a structured classification and comparative analysis to guide researchers and practitioners in advancing safer, more robust and socially acceptable autonomous driving systems.