Blind Robot Policies Shine in Demos But Fail in Production Due to Vision Overfitting

chris_j_paxton · x · 2026-08-05

A researcher points out that many current robot locomotion policies look impressive in demos because they overfit to proprioceptive states rather than genuinely utilizing RGB visual perception.

While these blind policies operate robustly in controlled environments, they fail when deployed in actual production settings. Consequently, relying on blind policies creates great demos but offers zero value for real-world commercialization.

Original post →

More from Embodied

Embodied channel →