Face to FACE® – Investigating horses' perception of facial expressions performed by an android robot

The ability to perceive and respond to emotional cues across species plays a central role in human–animal interactions and has implications for animal welfare, cognition, and social communication. Horses are believed to be able to understand the valence of human facial emotions presented in photographs, videos, and live interactions. However, the extent to which horses rely on purely visual facial cues, independently of human presence and other sensory modalities, remains underrated. To address this issue, the present study investigated whether horses can distinguish human-like emotional facial expressions displayed by an android robot under highly controlled conditions. Twelve adult mares were individually exposed to four facial expressions (neutral, happy, angry, surprise) performed by the robot FACE (Facial Automaton for Conveying Emotions®), with no humans present during testing. Heart rate variability (HRV), behavioural responses, and facial movements coded using the Equine Facial Action Coding System (EquiFACS) were recorded. The experimental design included baseline periods before and after exposure to the robot, allowing assessment of autonomic recovery and potential carry-over effects. Exposure to the robot induced significant autonomic changes compared to baseline, as reflected by shifts in HRV parameters across all conditions. Emotional expressions elicited concurrent sympathetic and parasympathetic modulations, indicating increased arousal. However, differences among emotional expressions were limited and inconsistent. Behavioural responses were non-specific, with increased locomotor activity observed across both neutral and emotional conditions. EquiFACS analysis revealed no significant variation in Action Units’ recruitment depending on the expressions, suggesting that horses displayed a common facial response pattern regardless of the robot’s display. To further explore multivariate response patterns, a machine-learning approach combining physiological and behavioural features was applied. Classification accuracy was moderate, with the neutral condition being identified more reliably than emotional expressions, which showed substantial overlap. Together, these findings indicate that while the android robot elicited measurable autonomic and behavioural activation, horses did not show evidence of discriminating the emotional valence of human-like facial expressions based on visual cues alone. Overall, the results suggest that horses’ responses reflected a general affective arousal rather than emotion-specific recognition. These findings challenge assumptions about equine sensitivity to isolated human facial expressions and highlight the importance of multimodal cues (i.e. vocal, postural, or contextual information) in interspecies emotional communication. The use of humanoid robots provides a valuable methodological tool for isolating visual sensory channels and controlling for unintentional human influence, offering new perspectives for research on animal emotions, cognition, and welfare.