Robots in children’s group activities must coordinate with the local organization of talk, participation, and attention. Yet robots sometimes act in ways that participants treat as misfitting the ongoing interaction. We conceptualize these moments as interactional misalignment. Drawing on conversation analysis of video-recorded robot-mediated group activities across multiple sites, we develop an inductive taxonomy of recurrent misalignments between robot conduct and participants’ orientations: turn-taking, sensorial, epistemic, deontic, and uptake-related. Across cases, children manage these moments through laughter, resistance, correction, and peer coordination, sustaining progress while revealing expectations for accountable robot behavior. We then propose a taxonomy-driven repair challenge protocol for evaluating robots’ sensitivity to participants’ displayed orientations, reframing robustness in group child--robot interaction as interactional adaptation rather than error prevention.