Can a robot understand what you're looking at?

Language instructions alone can't tell a robot which fork to pick up or where exactly to cut. Gaze2Act adds human eye-gaze as a dynamic intent signal, mapping what a person is looking at from their first-person view into the robot's perspective via semantic matching. Tested on 16 real tasks with a Unitree G1 humanoid, it outperforms language-only baselines on object disambiguation and precise interactions.