Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring
Researchers have developed a new framework called Hide-and-Seek to improve the reliability of robots using Vision-Language-Action (VLA) models. This method detects execution failures by identifying specific actions that indicate a problem, without requiring detailed step-by-step annotations. By using contrastive learning on trajectory-level data, Hide-and-Seek can pinpoint failure signals and offers a good balance between accuracy and timeliness for real-world robotic applications. AI
IMPACT Enhances the reliability of embodied AI systems by enabling more robust failure detection during robotic task execution.