New SAIL Blog post: R&B-EnCoRe: What Should a Robot Actually Think About…

A vision-language-action model takes in a scene and a task and outputs a low-level control action; adding a 'think before you act' reasoning step to that pipeline has typically relied on a fixed, human-authored template for what the model should reason about. Stanford SAIL's post points to R&B-EnCoRe, by Milan Ganai and Katie Luo, which instead has the model generate its own candidate reasoning.

The self-supervised loop measures how much each candidate piece of reasoning actually improves the model's prediction of the correct action, keeps the useful parts, and uses that signal to bootstrap a stronger policy over repeated cycles, all without a reward model, a verifier, or human annotation of what counts as good reasoning.