Equalized Coverage Model Predicts Self-Driving Car Control Under Failures
A lightweight regression model with statistical guarantees monitors actuator degradation in real time.
Automated driving systems must continuously monitor their own capabilities, especially when components degrade or fail. Engineers from TU Braunschweig tackle this with a new prediction model for motion control performance. They focus on highly over-actuated vehicles where multiple actuators can fail in varied patterns. Their lightweight model, based on conformalized quantile regression, predicts whether an automated vehicle can stay within a safe lateral deviation from its planned trajectory under nominal, degraded, and failed actuator conditions.
A key innovation is the use of equalized coverage methods to ensure statistical guarantees not just on average (marginal coverage) but across different regimes (conditional coverage). This makes the prediction reliable even in rare but critical failure scenarios. During runtime, the predictor serves as a heuristic to define the vehicle's admissible action space—what maneuvers are still safe. The paper, accepted for the 2026 IEEE International Conference on Intelligent Transportation Systems, demonstrates the method's application and discusses its limitations, advancing real-time capability monitoring for autonomous driving.
- Lightweight prediction model uses conformalized quantile regression to handle varied actuator failures.
- Equalized coverage ensures statistical guarantees across all operating regimes, not just on average.
- Model helps define safe action space in real time for self-adaptive road vehicles.
Why It Matters
Enables safer autonomous driving by predicting vehicle control limits during actuator degradations or failures.