No One at the Wheel? Advancing Reliability and Trust for the Autonomous Driving Era

Level 4+ autonomous vehicles (AVs) can deliver unique capacities for safer roads, reduced congestion, and greater mobility if designed well with regulatory compliance. Even with the recent great progress in artificial intelligence (AI), deep learning, and sensor fusion technologies, public hesitation and regulatory caution persist mainly because of concerns over reliability and trust issues. This paper introduces a new TRUSTEDrive framework to deal with these challenges by combining Deep Imitation Learning (DIL), Explainable AI (XAI), and human-in-the-loop mechanisms for end-to-end autonomous navigation. Compared to traditional modular systems, TRUSTEDrive learns from the behavior of expert drivers, along with internal states, and Visio-Psycho-Physiological (VPP) information (e.g., eye-gaze movements, muscle tensions, heart rate variability, and galvanic skin response) to understand the how and why behind human decisions. In addition to this, it supports the concerned autonomous driving agent (ADA) to be human-like reactive and explainable. Moreover, TRUSTEDrive can explain its priorities online based on XAI methods to offset the black-box nature of end-to-end AI and also to help passengers and pedestrians understand its underlying logic. Finally, the approach in which trust can be continuously quantified and adjusted via user feedback cycles and context-dependent explanations is presented, which helps adapt driving styles to individual desired levels.