CMU Robotics researcher Gokul Swamy releases SAILOR, an imitation learning framework that helps robots recover from errors using 10x less data

*Very* excited about this new paper -- we introduce a principled pipeline for training a closed-loop "elicitor" policy that robustly steers a VLA towards success. Plus, if there's a Hitchhiker's Guide reference in the title, you know the paper is gonna be good :p. [1/n]

When a VLA fails, it's *so hard* to tell whether it lacks the capability or was just prompted poorly.🤔

Joe showed that you can learn a language feedback policy that elicits the capabilities present in a frozen VLA, while refusing to steer when the VLA is incapable at the task!

We propose a method for doing a tractable "local search" (similar to what we explored in https://gokul.dev/sailor/) in the space of natural language, before using expert iteration to train our "language feedback policy" (LFP). [3/n]

We can perform this conformalization process with limited data as predicting whether improvement is *possible* is often easier than knowing the *precise* language sequence that will elicit the improvement. [5/n]

However, this LFP can generalize poorly OOD. In response, we use techniques from conformal prediction to figure out when to "fall back" to the default user prompt, making sure steering is "mostly harmless." [4/n]

Anyhow, I think this is an *incredibly* impressive first paper of the PhD for @hyunjoej, who continually impressed me with his work ethic and depth of thought. I can't recommend working with him enough :). Check out our website for more: https://hyunjoe.xyz/LanguagePolicy/.