Evaluate perceived naturalness and open-vocabulary target-selection accuracy

Evaluate whether the Realistic Gaze Transformer and its asynchronous Vision-Language Model targeting pipeline produce gaze that people perceive as natural, and determine the accuracy of open-vocabulary target selection under controlled evaluation.

Background

The paper reports robot-space kinematic fidelity and real-time inference for the integrated gaze-control system, but its benchmark supplies all controllers with the same ground-truth eye target. Consequently, the benchmark does not evaluate whether the Vision-LLM selects the intended referent or whether the generated gaze is perceived as natural by human observers.

The authors explicitly identify both perceived naturalness and open-vocabulary target-selection accuracy as unresolved questions requiring controlled evaluation rather than treating the reported motion-error metrics as sufficient evidence.

References

The evidence establishes robot-space kinematic fidelity and real-time RGT inference; perceived naturalness and open-vocabulary target-selection accuracy remain questions for controlled evaluation.

— Eye, Robot: Learning Expressive Gaze Dynamics via a VLM-Guided Transformer-VAE  (2610.19635 - Hu et al., 14 Oct 2026) in Abstract

These findings motivate a next HRI evaluation that tests whether people can infer the robot's attention during handover or collaborative search, as well as whether they prefer its gaze to alternative controllers.

— Eye, Robot: Learning Expressive Gaze Dynamics via a VLM-Guided Transformer-VAE  (2610.19635 - Hu et al., 14 Oct 2026) in Section 5, Conclusion