Abstract

Human–robot interaction (HRI) trust research has long emphasized cognitive mechanisms while overlooking the interdependent influence of emotional responses, cognitive evaluations, and individual traits in shaping trust. Despite this, traditional statistical approaches have proven insufficient for capturing the nonlinear and complex interactions that underlie HRI trust. This limitation underscores the necessity for an interpretable, data-driven method that can integrate emotional, cognitive, and individual dimensions. This study proposes a novel approach for assessing HRI trust by integrating machine learning with SHAP-based interpretability and clustering analyses. Emotional responses, cognitive evaluations, and individual characteristics are combined to capture their nonlinear relationships, offering deeper insights into how trust evolves and co-varies with affective and cognitive dimensions. SHAP and main effect analyses are employed to identify key factors and their primary effects on HRI trust, while clustering analysis is conducted to explore trust-related user profiles that integrate both emotional and cognitive dimensions. Results indicate that emotional, cognitive, and individual factors jointly influence trust in HRI. Emotional responses and cognitive evaluations exert nonlinear effects, where emotional influences are non-monotonic, and higher cognitive evaluation levels enhance trust but show threshold and saturation patterns. Stronger physiological arousal corresponds to lower trust, while moderate fixation counts indicate trust formation. Individual traits differ in emotional sensitivity and cognitive respon siveness. Prior robot-use experience affects trust in a nonlinear manner, with extraversion enhancing positive emotions and sense of security, whereas high neuroticism amplifies emotional reactivity under uncertainty. Further clustering analysis identified three emotional-cognitive coupling trust-oriented user groups: Rational Evaluators, Confident Adopters, and Pessimists. This finding reveals population heterogeneity in the emotional- cognitive coupling mechanisms underlying HRI trust. This study helps advance the creation of more effective, user-centered trust adaptation strategies in robotic systems.

Fig. 1. Overview of the research process.

Fig. 2. Feature Set Comparison. h denotes the human feature set, c denotes the cognitive evaluation set, p denotes the physiological feature set— in this study, specifically eye-tracking (ET) derived eye movement features, e denotes the emotion feature set, and r denotes the robot feature set. Combinations of letters (e.g., HC, CP, HEP, etc.) represent the corresponding combinations of feature sets.

https://doi.org/10.1016/j.aei.2026.104729