This study focuses on dialogue breakdowns that can occur in robot-assisted language learning (RALL). Our aim is to analyse how children use gaze to resolve these breakdowns, that is, interruptions in the interaction caused by the robot's inability to understand the children and react appropriately. Our corpus consists of 18 video filmed L2 learning situations where 36 primary school children talk pairwise with an educational robot for the first time. Our participants are 10-13 years old mono- and bilingual children from Swedish speaking schools in Finland learning L2 English. After detecting the breakdowns from the data, we use a multimodal analysis to identify most typical gaze patterns during these sequences. Our results show that when breakdowns occur, children's first gaze is directed most frequently towards the robot, but after that, they shift their gaze most often towards the teacher. This result suggests that the children try first to resolve the breakdowns with the robot, but when the problem persists, they also use gaze to seek assistance from other participants present in the learning situation. We interpret this finding to show that children attempt to treat social robots as human-like conversational partners in RALL, but that they turn to other human participants if the robot does not follow the expected interactional behavior.