Deep Learning–Based Prediction of Human–Robot Trust Dynamics in Future Construction Using Worker Neuropsychophysiological Responses

Abstract

Because current construction activities are safety-critical and physically demanding, the incorporation of such autonomous technologies as robots and drones via worker–robot teaming has drawn interest from researchers and practitioners alike. However, this teaming relationship may impose additional safety concerns for future jobsites due to workers’ inappropriate trust—overtrust and/or distrust—in robots. The literature has highlighted that trust is a complicated and dynamic concept that fluctuates over time, highlighting the need to continuously understand workers’ trust levels in real-time by collecting and interpreting workers’ psychophysiological signals. Consequently, deep learning (DL) has been deployed in various projects to identify trust-related psychophysiological patterns and to predict trust. However, current implementations suffer from three limitations: (1) focusing only on static settings, (2) manually extracting features, and (3) disregarding the trust continuum. Therefore, this study presents a DL model that automatically extracts important features from multiple psychophysiological signals and predicts workers’ increasing or decreasing trust within such dynamic workplaces as construction sites. The developed model can achieve accuracy, recall, precision, and F1 score all above 70%. This study also provides insights into a cost-effective strategy to prioritize data with high importance to trust prediction. Thus, the primary innovations of this research are (1) the consideration of the dynamic nature of construction sites, variability among workers, and trust continuum during model development; and (2) how pivotal knowledge about workers’ real-time trust can be harnessed to facilitate the development of human-centered robots in the future.