Beyond ReconVLA: Annotation-Free Visual Grounding via Language-Attention Masked Reconstruction

The article examines ReconVLA, which uses visual reconstruction as an internal supervisory signal to improve robot visual grounding for manipulation tasks, but identifies critical gaps including its reliance on costly gaze annotations. The author proposes an alternative architecture that replaces gaze annotations with language-driven attention masking, achieving annotation-free training and up to 5x faster inference while addressing the fundamental problem of scattered visual attention in robotic manipulation.