Predictions and goals modulate use of partner's gaze during joint action with humanoid robots

During joint actions, partners coordinate attention and actions by using each other's gaze to predict upcoming actions and align with shared goals. Although gaze orienting has been extensively studied using screen-based paradigms, it remains unclear whether partner reliability and action goals calibrate the use of the partner's gaze during embodied interactions. We addressed this using a humanoid robot as a co-actor in two experiments. In Experiment 1, participants performed a color-sorting task while the robot provided informative or noninformative gaze cues. Cue reliability was manipulated between participants: one group experienced a robot whose gaze was informative in 80% of trials, and another at chance level. Task difficulty was manipulated within participants. Only the reliable group made fewer errors on valid than invalid trials under high-difficulty. In Experiment 2, participants completed a handover task in which an auditory cue established an action plan, the robot looked toward a goal-congruent or incongruent object, and a target appeared at a valid or invalid location. Participants responded faster at the cued location only when the robot looked toward the goal-matching object, with eye tracking confirming covert attentional shifts. Together, these findings show that the use of the partner's gaze in embodied joint action operates as a flexible, utility-weighting heuristic. Observers weighted gaze according to its value, upweighting gaze from a reliable partner and when it aligned with shared goals and down-weighting gaze that conflicted with predictions. These results advance understanding of gaze use as a predictive, context-dependent mechanism with implications for reliance on social signals during joint action with humans and artificial agents.