Consequence Learning for Trajectory Planning in Autonomous Driving

Abstract

Imitation learning (IL) teaches autonomous driving models what an expert does, but not why that action is safe or optimal. This critical gap arises because a single expert trajectory cannot illuminate the broader solution space: a complex performance landscape with multiple, distinct solutions and sharp “performance cliffs”, where minor deviations can trigger catastrophic failure. Without a map of this landscape, the model fails to learn the link between actions and their consequences. We introduce CODA (COnsequence-Driven Autonomy), a framework that enriches IL by providing an explicit map of action consequences. First, for each scene, we construct a dense performance landscape by evaluating a wide spectrum of potential behaviors. This landscape serves as a rich supervisory signal, training the model to directly predict the closed-loop quality of any given plan from its context. This process compels the model to develop a deeper understanding of driving dynamics, moving beyond simple imitation. Second, the learned consequence predictor serves as an internal critic for self-evaluation during inference. Third, CODA facilitates a low-cost, expert-free DAgger variant. Finally, it addresses the systemic gap between planned and executed trajectories. Experiments on the large-scale real-world nuPlan dataset show that CODA significantly improves closed-loop driving performance, outperforming state-of-the-art methods, especially in complex and challenging scenarios.