Human gaze is more than a pointer; it coordinates attention through the coupled motion of the eyes, head, and eyelids. For humanoid robots, deciding where to look is only half the problem—the robot must also generate a movement that feels natural to a human partner. Fixed tracking rules, which simply follow a target, do not capture the expressive quality of real gaze behavior.

The paper proposes a VLM-guided Transformer-VAE to address this. A vision-language model provides high-level contextual guidance, while a Transformer-based variational autoencoder learns the dynamics of gaze over time. This lets the system generate coordinated eye, head, and eyelid movements that are both target-directed and expressive, rather than relying on rigid, rule-based tracking.

The work highlights a shift toward learning-based, context-aware gaze generation for humanoid robots. By modeling the full gaze apparatus instead of just the direction of the eyes, the approach aims to make robot interactions feel more socially fluent and less mechanical. The abstract does not include quantitative results, so the reported contribution is the architecture and its motivation rather than a benchmarked performance claim.