Rewind-IL Framework Enhances Robot Imitation Learning with Failure Detection and Recovery

Rewind-IL, a novel online framework that requires no training, tackles deployment issues in imitation learning for robotics. This system integrates a zero-shot failure detection method based on Temporal Inter-chunk Discrepancy Estimate (TIDE) alongside a state-respawning feature that brings robots back to semantically verified safe intermediate states. It is designed for long-horizon action-chunked policies, which can generate locally plausible actions but fail to recover when execution strays from the demonstration manifold. Current runtime monitors need failure data, trigger too easily with minor feature drift, or only detect failures without offering recovery solutions. A vision-language model identifies safe intermediate states offline, and TIDE is fine-tuned using split conformal prediction for enhanced reliability. This framework was published on arXiv under identifier 2604.16683v1 as a cross-type abstract. While imitation learning has allowed robots to learn intricate visuomotor skills from demonstrations, deployment failures continue to pose significant challenges. This research offers a thorough solution that detects failures without the need for training data and includes a recovery mechanism rather than mere detection.