Personalized Adaptive Cruise Control (P-ACC)

Multi-phase approach combining offline inverse reinforcement learning with online Gaussian process regression for real-time driver preference adaptation within a cloud-vehicle digital twin framework.

• Offline Phase: IRL algorithm recovers reward functions from naturalistic driving data, classified by driver type and weather conditions
• Online Phase: GPR-based adaptation updates driving gap preference table (DGPT) in real-time using driver feedback
• Cloud Architecture: Digital Twin framework stores personalized models, enables incremental learning and federated model sharing
• Maximum cumulative reward criterion for model selection and real-time implementation

P-ACC Digital Twin framework

P-ACC IRL (offline) + GPR (online) learning

Demonstrated improvements in personalization accuracy and driver satisfaction across multiple validation studies using both naturalistic driving data and human-in-the-loop simulation.

• Numerical Simulation: 30.1% improvement in speed reproduction, 36.5% improvement in distance gap accuracy vs IDM
• Human-in-the-Loop: Up to 93.4% reduction in driver takeover frequency compared to IDM-based ACC
• Real-time Learning: 62.8% reduction in driver intervention (PoI) and 62.2% reduction in intervention frequency (NIM)
• Weather Adaptation: Successful personalization across clear sky, night, and foggy weather conditions
• Driver Classification: Effective model selection using cumulative reward criterion for untrained drivers