Safe and Nonconservative Contingency Planning for Autonomous Vehicles via Online Learning-Based Reachable Set Barriers

Abstract

Autonomous vehicles must navigate dynamicallyuncertain environments while balancing safety and efficiency.This challenge is exacerbated by unpredictable human-drivenvehicle (HV) behaviors and perception inaccuracies, necessitatingplanners who adapt to evolving uncertainties while maintainingsafe trajectories. Overly conservative planning degrades drivingefficiency, while deterministic methods risk failure in unexpectedscenarios. To address these issues, we propose a real-timecontingency trajectory optimization framework. Our methodemploys event-triggered online learning of HV control-intent setsto dynamically quantify multimodal HV uncertainties and incre-mentally refine their forward reachable sets (FRSs). Crucially, weenforce invariant safety through FRS-based barrier constraintsthat ensure safety without reliance on accurate trajectory predic-tion. These constraints are seamlessly embedded in contingencytrajectory optimization and solved efficiently through consensusalternating direction method of multipliers (ADMM). The systemcontinuously adapts to HV behavioral uncertainties, preservingfeasibility and safety without excessive conservatism. High-fidelitysimulations on highway and urban scenarios, along with a seriesof real-world experiments, demonstrate significant improvementsin driving efficiency and passenger comfort while maintainingsafety under uncertainty.