Tri-Info: Generalizable, Interpretable Failure Prediction for VLA Models via Information Theory

The article introduces Tri-Info, a novel method for failure prediction in Vision-Language-Action (VLA) models, leveraging information theory to analyze the differences in information-theoretic signatures between successful and failed rollouts. Tri-Info operates as a closed-loop information pipeline, achieving 83% accuracy in real-world tasks across six VLA models and three benchmark environments, while demonstrating strong generalization without the need for retraining. This approach enhances the interpretability of failure detection, providing actionable insights into the underlying causes of failures, which is crucial for practitioners developing robust AI systems.