Confidence-Aware Gaze-Depth Interaction for Reliable 3D Target Selection in Consumer VR Headsets

Gaze-based interaction offers a natural, low-effort input channel for virtual reality (VR), but gaze-depth estimates on consumer headsets vary with viewing distance, fixation stability, head motion, and scene properties. We propose confidence-aware gaze-depth compensation, which uses a heuristic confidence score to adapt angular/depth tolerances, with conditional depth blending and state feedback. In a within-subject study, 60 participants completed 6,000 trials across four conditions: baseline gaze, fixed-window control, confidence-aware compensation, and gaze + controller confirmation. Confidence-aware compensation improved selection success and policy-output depth error over both gaze-only controls and achieved the highest failure-inclusive efficiency (19.14 successful selections/min; 5.46, 12.16, and 13.02 for the other conditions). The raw-failure model showed moderate discrimination (AUROC = 0.734) and close calibration (ECE = 0.006). Mandatory confirmation achieved the highest accuracy but increased latency and workload. Results support reliability-aware compensation for controlled layered selection. Transfer to irregular or continuous-depth scenes remains to be established.