Noise-Adaptive Driving Assistance Systems with Deep Reinforcement Learning

V.Kochliaridis, N. Chandrinos, G.Parlitsis, I.Vlahavas, Noise-Adaptive Driving Assistance Systems with Deep Reinforcement Learning Control Engineering Practise, Elsevier, 2025

Vasileios Kochliaridis, Nikolaos Chandrinos, Georgios Parlitsis, Ioannis Vlahavas

Advanced Driving Assistance Systems, Deep Reinforcement Learning, State Estimation, Uncertainty, Sensor Failures

Although Advanced Driving Assistance Systems (ADAS) aim to improve the driving experience and the safety of human drivers, their reliability and performance in uncertain environments, coupled with strict computational requirements, make ADAS development a challenging and time-consuming process. Various Deep Reinforcement Learning (DRL) approaches have been proposed to address these issues and accelerate their development. While current DRL approaches demonstrate promising results in several navigation tasks, they are typically evaluated only under ideal conditions in simulations, observing a substantial decline in both performance and generalization when faced with realistic driving conditions. In this paper, we propose NADAS (Noise-Adaptive Driving Assistant System), a controller architecture that integrates DRL with sensor noise modeling, effective state representation, and a fast-performing state estimation technique. These combined techniques reduce the environmental uncertainty for the controller and improve its performance. We evaluate NADAS by using it to implement an Adaptive Cruise Controller and demonstrate its effectiveness in complex urban simulation scenarios. Finally, we evaluate its computational efficiency on embedded computing boards, highlighting its potential for real-time deployment.