Overview
This research addresses gaze-based assistive manipulation, specifically focusing on the challenge of enabling robot placement at arbitrary spatial positions. While gaze-based systems typically facilitate object selection, achieving precise placement at non-predefined locations necessitates accurate spatial synchronization between a head-mounted display and a robotic system. The study highlights that conventional metrics of pose accuracy do not consistently correlate with task accuracy in gaze-based manipulation, as translational and rotational errors can interact, potentially compensating for each other's effects.
Research Context
Gaze-based assistive manipulation conventionally supports object selection tasks. However, the expansion to arbitrary-position placement requires rigorous spatial alignment between the user's headset and the robotic manipulator. The research identifies a critical distinction: the accuracy of pose (position and orientation) metrics does not inherently translate into equivalent task accuracy. This is due to the complex interplay of translational and rotational errors, which, when jointly affecting the transformed gaze ray, might either exacerbate or mitigate overall targeting inaccuracies.
Approach
To investigate cross-device alignment from a task-oriented perspective, the researchers developed a markerless interaction framework. This framework was complemented by a dedicated cross-device dataset designed to support the evaluation of gaze-robot alignment. Within this framework, a method termed Graph-based Reference Selection was proposed to address scenarios characterized by sparse robot reference points.
The study further involved the development and benchmarking of multiple task-specific alignment pipelines. These pipelines were evaluated under a unified protocol to ensure consistency. A key component of the evaluation methodology was the introduction of Gaze–Surface Intersection Error (GSIE). GSIE is defined as a direct measurement of the spatial error associated with a gaze-specified target, providing a task-specific metric for performance assessment.
Findings
Experiments conducted using the established framework and GSIE metric revealed that alignment methods which achieved high rankings according to conventional pose metrics did not invariably perform optimally when evaluated using GSIE. This observation directly indicated the importance of assessing gaze-based manipulation performance at the task level, rather than solely relying on device-level pose accuracy metrics.