Guide, Think, Act: Interactive Embodied Reasoning in Vision-Language-Action Models

Researchers introduced GTA-VLA, an interactive vision-language-action framework for robot control.

It accepts spatial cues like points, boxes, and traces, then fuses them into a spatial-visual chain-of-thought for planning.

A lightweight action head carries out the execution step, separating reasoning from control.

Experiments show state-of-the-art performance on SimplerEnv WidowX and stronger success under visual shifts and ambiguous scenes.

The method could help robots use human visual guidance to recover better from failures and distribution shifts.