Researchers have a new way to make robots doubt themselves before they mess up.
Vision-Language-Action models, or VLAs, are the AI systems robots use to turn a camera feed and a spoken instruction into a physical action, like "pick up the red cup." A team of researchers built a framework called MAE (Markov Attention Entropy) that lets these models score their own confidence before they even act. Instead of relying on human-labeled examples or judging only the final output, MAE looks at how the model's internal attention shifts while it processes what it sees. The team found this internal "entropy" - a measure of how scattered or focused a model's attention is - looks noticeably different when a task is about to succeed versus fail, and that pattern holds across VLAs built on different underlying architectures. They tested the approach on a new 4,000-episode benchmark called LIBERO-Reflect, split evenly between standard tasks and deliberately hard ones, and MAE outperformed existing self-evaluation methods on every metric reported.
This matters because robot failures are usually silent until they get expensive: a warehouse arm drops a box, a home robot knocks over a glass. A model that can flag "I'm not confident about this one" before acting opens a path to catching mistakes without a human watching every move, and without paying to hand-label thousands of failure cases. The researchers also showed MAE can guide a robot to sample several possible actions and pick the most reliable one, improving robustness for only a small runtime cost.
Worth noting: this is simulation work, run on the LIBERO benchmark suite, not a warehouse floor. Confidence scores that track failure in a simulator do not automatically survive contact with real lighting, real clutter, and real gravity.