VLA-FAIL: Efficient Task Failure Detection for Finetuned Vision-Language-Action Models
VLA-FAIL is a lightweight framework for detecting failures in Vision-Language-Action Models (VLAs) during deployment without needing failure data. It combines two methods: Last-Layer Mahalanobis Distance (LLMD) to spot internal feature shifts and Action Chunk Consistency (ACC) to find erratic behavior in actions. This combination offers reliable, early detection with minimal computational cost.