A new study has found that generative AI-powered hybrid collision avoidance system can help automated vehicles detect high-risk “cut-in” collisions earlier than conventional safety systems.
The research, led by Jamal Raiyn, a professor and researcher in computer science, transportation and telecommunication at Al Qasemi Academic College, was published in Transportation Research Interdisciplinary Perspectives. It examined a hybrid model that blends generative AI, which learns patterns from large volumes of driving data to predict outcomes, with a rule-based system, which follows predefined safety rules and thresholds.
The approach is designed to identify cut-in events, in which a vehicle suddenly merges into another lane, a frequent cause of multi-vehicle crashes. It compared the performance of hybrid collision avoidance models compared with traditional time-to-collision models, which estimate how much time remains before a potential impact. According to the findings, the hybrid system identified risks sooner and initiated responses more quickly.
“The simulation results demonstrate that the Intelligent Rules Approach −based collision avoidance system outperforms traditional TTC-based methods. The system accurately predicts potential collisions and initiates timely evasive actions, significantly reducing the risk of collisions in cut-in scenarios.”
The research also examined the role of driver behaviour in collisions, identifying human factors as the primary cause.
“Most road accidents are primarily caused by human factors and driver errors. Young and inexperienced drivers are disproportionately involved in crashes, with inattentive or aggressive driving behaviors further increasing accident risk.”
The system differs from conventional approaches by adapting in real time to driver behaviour and traffic conditions, rather than relying only on fixed thresholds. It continuously updates its predictions using both learned data patterns and preset safety rules.
“The novelty of this research lies in combining rules-based approach with behavioral insights and dynamic TTC computation to create a comprehensive, adaptive collision avoidance framework that addresses limitations in static, rule-based systems.”
Elements of this approach are already in use in commercial automated driving systems, though not as a single, unified feature. Waymo uses machine learning to predict the behaviour of nearby vehicles, including sudden merges, while rule-based safety layers govern braking and collision avoidance. Tesla applies a similar structure in its full self-driving (FSD) system, where neural networks trained on fleet data anticipate lane changes and cut-ins, with rule-based safeguards for critical interventions.
