REPAIR-Bench: A Benchmark for Robot Error Perception and Interaction Recovery

A benchmark built on 214 interaction trials from 41 participants, spanning four induced failure types with synchronized multimodal signals: facial action units (OpenFace), head pose, speech transcripts, and post-interaction affect and recovery reports. We formalize three novel evaluation tasks capturing the full failure lifecycle: longitudinal failure detection across interdependent sessions, visual failure-type classification, and user-centered recovery-preference prediction. Hierarchical recurrent modeling improves failure detection over single-session baselines (strict F1: 0.80 vs. 0.68), with failure-localization median absolute error of 2.97s. We fine-tune Mistral-7B with QLoRA for recovery prediction (Hit@5 = 0.76, F1@5 = 0.32).