A Review on Multi-Scenario Decision-Making for Urban Autonomous Driving

Autonomous vehicles depend on decision-making systems capable of interpreting environmental data and vehicle states to operate safely and efficiently. Single-scenario methods perform well in limited conditions but lack adaptability for diverse scenarios. Generalized approaches address these challenges but face issues with efficiency and safety. This systematic review compares single-scenario and multi-scenario decisionmaking techniques, categorized as rule-based, optimization, learning, and hybrid methods. It highlights gaps and advances needed to enhance flexibility and reliability for handling complex driving scenarios. Analyses show that hybrid methods tend to offer greater adaptability and security by combining explicit rules and constraints with the flexibility of learning. System evaluation uses databases, simulation, and closed-track and open-road testing, but the focus varies depending on the approach: in single-scenario, the aim is to diagnose and correct specific behaviors under controlled conditions; in multiple scenarios, the objective is to measure the system’s robustness and generalizability within the operational domain. Despite advances, gaps persist in the ability to deal with unknown and critical situations, demanding future investigations into generalization strategies, incremental learning and integration of contextual representations of the environment.