A Taxonomy of Mental Model Mismatches for Human-Robot Interaction

Abstract

Mental model mismatches (MMM) in human-robot interaction occur when humans’ internal representations of robots diverge from robots’ actual properties or behaviors, leading to communication breakdowns, task failures, and reduced trust. These mismatches are subtle and difficult to detect, yet the field lacks shared terminology to identify specific types. This makes it difficult to diagnose problems, design targeted interventions, or compare findings across studies. We present a theoretically grounded taxonomy of sixteen MMM types organized by the Sense-Plan-Act (SPA) architecture into sensing (7 types), planning (5 types), and acting (4 types). Drawing from cognitive neuroscience, psychology, communication studies, and human factors, each type addresses a distinct cognitive divergence that arises when users apply mechanisms evolved for understanding humans to robotic systems. The taxonomy's exhaustiveness is established through a dimensional decomposition: three representational dimensions grounded in the neurocognitive architecture of human perception (What, Where, How) are systematically crossed with the three SPA processing stages, and every cell in the resulting matrix is accounted for by a dedicated type or an explicit justification. Each type is equipped with a formal definition and diagnostic examples. We demonstrate the taxonomy's applicability across two contrasting datasets, TEACh (task-oriented simulation) and SHREC (social therapeutic dialogue), revealing distinct mismatch profiles that reflect the structural differences between the two settings. This taxonomy provides precise terminology enabling researchers to identify specific cognitive mismatches, and a foundation for developing targeted interventions, rather than treating all problems as generic ”communication failures.”