An EEG-based non-linear state-space model for trust inference in a space-relevant human-autonomy teaming task.

Abstract

Effective collaboration in human-autonomy teaming (HAT) depends on maintaining appropriately calibrated trust in the autonomous partner. Trust evolves as operators observe system behavior and task outcomes, requiring continuous monitoring rather than overall surveys or post-task self-reports. Electroencephalogram (EEG) offers high temporal resolution for tracking cognitive states; however, most existing EEG-based trust models are static or rely on linear mappings that fail to capture the non-linear and time-dependent nature of trust dynamics. This paper presents an EEG-based non-linear state-space model for continuous trust inference during a space-relevant supervisory task. The model integrates a low-dimensional linear latent process to represent the temporal evolution of trust with a non-linear manifold that maps EEG spectral features to the latent state, enabling trust prediction from both latent dynamics and neural representations. Model performance was evaluated using EEG data collected from participants monitoring autonomous systems with varying reliability and transparency, and compared against a linear dynamic model and a static model. After correcting the unit of statistical inference to participant-level averages, the proposed model achieved significantly higher correlation with self-reported trust than both the linear dynamic model (p = 0.0021) and the static model (p < 0.0001), and significantly lower prediction error than the static model (p = 0.0020). The RMSE reduction relative to the linear dynamic model was numerically favorable but not statistically significant after participant-level aggregation. Model interpretation revealed that trust inference relied on distributed neural features, with contributions from frontal and prefrontal regions in delta and alpha bands and additional involvement of higher-frequency activity (low- and mid-gamma bands), and that high- and low-trust states occupy distinct regions of the inferred latent space. These results demonstrate that incorporating non-linear neural structure within a dynamic modeling framework improves EEG-based trust inference and supports the development of cognition-aware adaptive autonomy.