This paper tackles the fundamental challenge of balancing safety and driving efficiency when autonomous vehicles interact with human-driven vehicles exhibiting uncertain, time-varying behaviors.
The Challenge: Human drivers have diverse, unpredictable behaviors. Worst-case assumptions lead to overly conservative driving, while ignoring uncertainty risks collisions. Traditional methods struggle to adapt online to evolving uncertainty estimates.
Our Approach: We developed a real-time contingency trajectory optimization framework that:
- Uses event-triggered online learning to dynamically refine estimates of human driver behavior (control-intent sets)
- Propagates Forward Reachable Sets (FRS) as ellipsoidal over-approximations of where humans could be
- Enforces safety through FRS-based barrier constraints that guarantee collision avoidance without relying on accurate trajectory predictions
- Jointly optimizes a nominal performance trajectory and a safety contingency trajectory via consensus ADMM
Key Results:
- Zero collisions in tested scenarios versus 18-27% collision rates for baselines
- 31% lower peak longitudinal jerk, 18.3% lower lateral jerk than ST-RHC
- 14-20% higher average speed than conservative baselines
- 64.4% reduction in lateral jerk versus uncertainty-aware baseline
- Real-time performance (34 ms average solve time) validated on 1:10 scale robots with aggressive human driver behaviors
The framework adapts safety margins automatically based on observed driver behavior — maintaining efficiency with predictable drivers while expanding safety buffers for erratic ones — all while preserving formal safety guarantees.