Does the VLA Know It Will Fail? Interpretable Success and Failure Patterns in Physical AI Systems

The researchers found that sparse autoencoder (SAE) representations activated in OpenVLA contain outcome-prediction information that can predict task failure before the robot enters an irrecoverable state. They identified individual SAE features whose strong activations are associated with specific stages of success or failure, and by localizing the time intervals in which these feature activations change significantly, they induced recurring behavioral patterns. The team also developed a VLM-based interpretation pipeline to generate behavioral descriptions for these patterns, and validated on held-out intervals that the descriptions correspond to feature activations.