Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning

Milan Ganai ♯ \sharp , Katie Luo ♯ \sharp , Jonas Frey ♯ \sharp ♭ \flat , Clark Barrett ♯ \sharp , and Marco Pavone ♯ \sharp ♮ \natural · 2026

Read Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning by Milan Ganai ♯ \sharp , Katie Luo ♯ \sharp , Jonas Frey ♯ \sharp ♭ \flat , Clark Barrett ♯ \sharp , and Marco Pavone ♯ \sharp ♮ \natural (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-08167.txt.

Description

Vision-language-action models reason in chains of thought, but rigid templates for thought primitives can force models to process irrelevant details. Milan Ganai and four co-authors offer R&B-EnCoRe, which bootstraps embodied thought from internet-scale knowledge through self-supervised refinement.

Treating thought as a latent variable in importance-weighted variational inference, the model generates and distills a dataset of embodiment-specific thought strategies without external rewards, verifiers, or human labels. Validation spans simulated and hardware manipulation and legged navigation. Robotics teams gain a route past the chicken-and-egg problem of thought quality: better thought no longer waits for successful policies to verify it, nor vice versa.

Catalog record

CREATORMilan Ganai ♯ \sharp , Katie Luo ♯ \sharp , Jonas Frey ♯ \sharp ♭ \flat , Clark Barrett ♯ \sharp , and Marco Pavone ♯ \sharp ♮ \natural
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE97 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-08167

Browse the full library · About the Tower of Babel