Vision-language-action (VLA) policies need to connect language instructions to the right objects, contact points, and target locations, but existing robot datasets rarely include human gaze alongside the multi-camera observations these policies use. The GazeAct dataset addresses this gap by recording gaze directly on the robot camera feed during screen-based teleoperation, yielding 12,000 demonstrations across 40 tasks and five domains. Gaze is captured at 300 Hz and synchronized with images, proprioception, and actions.

The authors transfer gaze from the operator-selected view to the other synchronized cameras, creating spatially consistent attention targets across all policy inputs. These targets are then used to regularize language-conditioned visual attention during training, while inference remains gaze-free. Ablations show that soft supervision works best and that real human gaze outperforms generic attention-shaping priors.

Across 18 real-world tasks, the method raises average success from 48.1% to 58.6%, with individual gains up to 25 percentage points. The improvements also transfer to the LIBERO-Spatial benchmark (57.9% to 69.6%) and to a different VLA backbone. The results indicate that gaze collected in robot observation space can serve as an effective training-time signal for improving visual grounding in manipulation policies.