Rough handling detection and crying-based rejective feedback for a baby-like robot: system design and evaluations by older adult users and caregivers

Caregiving-style physical interaction with baby-like robots can support the engagement of older adults with cognitive impairment in care settings. However, some users handle such robots roughly, which can disrupt the interaction. Without rejective feedback from the robot, the boundary between acceptable and unacceptable handling can remain ambiguous to both users and caregivers; moreover, context-inappropriate feedback can confuse users and hinder caregiver decisions about intervention. We first constructed a behavior recognition model to detect rough handling, defined as tossing and dropping, using onboard inertial sensors. The model was trained on data from 24 older adults and achieves a mean F1-score of 0.808. We then integrated this detection model into a robot system that outputs crying as rejective feedback. We evaluated our approach against non-rejective responses from two stakeholder perspectives. In a lab study with 24 older adults, crying made the perceived robot affect more unpleasant; overall impression and guilt did not differ. In a third-party video study with 199 caregivers, crying increased perceived ease of use, usefulness, and intention to use. Our findings suggest that negative vocal feedback communicates robot discomfort to older adult users and helps caregivers interpret incidents and choose an appropriate response.