Continuous Reasoning for Vision-Language-Action

Continuous Reasoning for VLA: a shared Gaussian latent interface trained with a self-verification objective, so that reasoning is reusable, shareable, and aligned with chunk-level control. +40.4% on TX-G2 and +26.3% on HSR real-robot subtask success over pi_0.5.

关键词：vision-language-action, VLA, reasoning, continuous thoughts, self-verification, EMA teacher, latent reasoning, AIRoA