A Wearable Non-Visual Guidance System with an Inspectable Spatial-Cue Core

Abstract Wearable assistance for blind and low-vision users must convert egocentric perception into interpretable action cues. This paper presents a three-node guidance system comprising smart glasses, an RK3588 edge unit, and a four-motor wristband. RGB-D/stereo perception, YOLO11s detection, metric depth estimation, and 3D back-projection produce task-specific spatial relations. An upstream cue core validates task intent, entity, mode, relation, and state before transmitting vector packets over BLE; the wristband applies roll compensation and renders directional haptic feedback. Placement-level relative-vector errors are 1.13 ± 0.38, 1.66 ± 0.47, and 1.76 ± 0.30 cm at nominal 30, 50, and 70 cm placements (n = 15 per distance). Grasping succeeds in 37/40 trials under a 60 s timeout. A supplementary study report compares smart glasses and a white cane in 100 participants, including 50 blind and 50 low-vision participants, using a counterbalanced mixed design and collision counts as the primary outcome. In the blind group, mean collision counts are 4.46 with glasses and 12.54 with the cane; the report gives t(49) = 30.92, p < 0.001, and Cohen’s d = 4.37. In the low-vision group, the mean paired difference is 0.14 collisions, with a reported 90% confidence interval of [−0.01, 0.29], meeting the specified ±1.5-collision equivalence criterion. The reported age-adjusted group-by-device interaction is significant (F (1, 97) = 412.5, p < 0.001), supporting visual-status-dependent benefit. By integrating performance indicators and behavioral outcomes, this multidimensional framework provides a practical tool for evaluating wearable assistive systems, supports safer and more effective autonomous assistance, and contributes to broader human–machine interaction research.