Dual-Process Framework Improves Safety in Open-World Autonomous Driving

Summary

The paper proposes a dual-process framework for making autonomous-driving systems safer in unfamiliar open-world situations. A neural network handles intuitive, learning-based planning during routine driving, while model predictive control provides reasoning-based planning when situations fall outside the model’s experience. A meta-cognitive component decides when to switch between the two, using a knowledge graph to reason about contextual risk from explicit perceptions, traffic rules, and social norms. Risk fields represent that contextual risk and guide both switching and the reasoning-based planner’s compliance with safety requirements. The authors evaluate the architecture in CARLA on variations of an out-of-distribution scenario involving emergency vehicles running a red light. Compared with a neural-network-only planner, the proposed system reduces collisions by 89% and improves compliance with special right-of-way rules.