Toward Trust-calibrated Customized Vehicle Automation

Abstract

Vehicle automation has long been an essential component of modern vehicular technologies. From the Advanced Driver Assistance System (ADAS) to fully autonomous vehicles, these technologies are trying to enhance road safety and improve the driving experience. However, their acceptance does not seem to keep up with the fast-growing market penetration. There have been many attempts to improve the acceptance of vehicle automation. The most widely adopted methods are customization and trust calibration. However, to the best of our knowledge, none of the current studies provides a holistic framework to design a customized vehicle automation system that involves humans in the loop. This dissertation aims to propose a systematic paradigm to design a human-aware and trust-calibrated vehicle automation system. We start from the system side to build customized vehicle automation using driving demonstrations from naturalistic driving data. The procedure consists of identifying driving style using Multivariate Functional Principal Component Analysis (MFPCA) and clustering analysis. Then, using the clustering result, we apply the Maximum Entropy Inverse Reinforcement Learning (MaxEnt IRL) algorithm to train the personalized vehicle automation system from demonstrations. The result shows the effectiveness of our proposed method. Next, from the human side, we develop a customized real-time trust prediction model based on trust dynamics. ``Confident'' and ``Skeptical'' trust dynamics are identified from the trust levels during interaction with the vehicle automation.Furthermore, the State Space models for real-time trust prediction improve trust prediction performance for further trust calibration considerations. Finally, we integrate customized vehicle automation with the trust models using the human-in-the-loop method to achieve trust-calibrated customized vehicle automation. The proposed algorithm takes a step toward our goal of building human-aware customized automated systems. The findings have implications for future human-vehicle interaction design and trust calibration in vehicle automation.