Researchers introduced GazeAct, a real-robot dataset and training method that pairs 12,000 demonstrations across 40 tasks with human gaze recorded at 300 Hz. The gaze helps train vision-language-action policies to focus on relevant objects and locations; the authors say the robots do not need gaze input when running.
Across 18 real-world tasks, the authors report average success rising from 48.1% to 58.6%, with gains of up to 25 percentage points. They also report an increase on the LIBERO-Spatial benchmark, from 57.9% to 69.6%.
Why it matters
For robot researchers, GazeAct offers a way to use operator eye movements as training guidance without requiring gaze sensors during deployment; whether the gains hold beyond the reported tests remains unclear.
