This paper introduces the ERR@HRI 2.0 challenge for detecting and resolving errors (e.g., misunderstanding user intent, user interruptions in conversation, and failure to respond) in conversational robots based on large-scale language models (LLMs). The challenge provides a 16-hour human-robot interaction dataset (containing facial, vocal, and head movement features) annotated with the presence of robot errors and the user’s error-correction intentions. Participants will develop machine learning models to detect robot errors using multimodal data, and will be evaluated on metrics such as detection accuracy and false positive rate. This is an important step toward improving error detection in human-robot interaction through social signal analysis.