Confidence-driven weighted retraining for predicting safety-critical failures in autonomous driving systems

Area of Science:

• Cyber-physical systems
• Autonomous driving technology
• Machine learning for safety-critical applications

Background:

• Safe handling of hazardous driving situations is crucial for trustworthy autonomous driving systems.
• Accurate vehicle confidence prediction is needed to prevent failures during unpredictable conditions.
• Adapting misbehavior predictors with real-time data presents significant challenges.

Purpose of the Study:

• To present a framework for the continual learning of misbehavior predictors in autonomous driving systems.
• To address the challenge of adapting misbehavior predictors with in-field knowledge.
• To improve the reliability and trustworthiness of autonomous driving systems.

Main Methods:

• A framework for continual learning of misbehavior predictors is proposed.
• In-field behavioral data is recorded to identify appropriate data for adaptation.
• Adaptive retraining is guided by in-field confidence metric selection and reconstruction error-based weighing.
• The framework was evaluated on the Udacity self-driving car simulator.

Main Results:

• The proposed framework significantly reduces the false positive rate of misbehavior prediction.
• The system demonstrates adaptability to nominal behavior drifts.
• The original capability to predict failures several seconds in advance is maintained.
• Improved performance of a literature misbehavior predictor was achieved.

Conclusions:

• The developed framework enables effective continual learning for autonomous driving misbehavior predictors.
• The approach enhances system safety by reducing false alarms and improving early failure detection.
• This method contributes to building more reliable and trustworthy cyber-physical systems for autonomous driving.