A Gaze-Driven Robotic System for Post-Stroke Active Ankle Rehabilitation Training.

Gaze-driven quadrant selection achieved near-perfect accuracy in healthy adults across 1,600 trials.

• Study enrolled 20 healthy adults with a total of 1,600 trials.
• Selection accuracy reached 99.94%.
• Median end-to-end delay (gaze onset to motor onset) was 157.6 ms.
• A dwell/occupancy rule was used to trigger commands, reducing false activations.

The system demonstrated robustness to head tremor, producing no wrong-quadrant decisions during a head-tremor stress test.

• A head-tremor stress test was performed at multiple severity levels.
• No wrong-quadrant decisions were recorded at any tremor severity.
• At the highest tremor severity, decisions were withheld because the occupancy criterion was not met.
• The conservative dwell/occupancy rule was credited with maintaining decision integrity under perturbation.

Passive tracking of ankle motion under robot-driven trajectories was achieved with sub-degree angular error.

• Tracking RMSE was ≤0.224° under passive drives.
• Motion profiles were described as smooth.
• Torques remained within software-defined limits throughout passive operation.
• The robot provided two degrees of freedom: dorsiflexion/plantarflexion and internal/external axial rotation.

Human factors outcomes including usability, workload, and visual fatigue were favorable in both healthy adults and a pilot post-stroke cohort.

• Visual fatigue change was minimal: ΔVAS 0.14 (healthy) and 0.21 (post-stroke pilot cohort).
• Outcomes included high usability and low workload ratings.
• Both healthy participants and post-stroke participants were assessed.
• Results were characterized as 'favorable' human factors outcomes supporting clinical workflow integration.

The gaze tracking system used pupil center detection with near-infrared tracking, mapped to a unit-normalized screen plane via low-order regression with ArUco-guided homography and rapid affine correction.

• Pupil centers were detected using a near-infrared tracker in a binocular configuration.
• Mapping to the screen plane used low-order regression combined with ArUco-guided homography.
• Rapid affine correction was applied to maintain calibration accuracy.
• The system was described as 'clinic-friendly' and designed for seated ankle training workflows.

The ankle robot employed a cascaded position-velocity-current control architecture with safety features including software rate/torque limits and watchdog supervision.

• Control architecture was cascaded position-velocity-current control.
• Jerk-limited trajectories were executed upon command trigger.
• Software rate and torque limits were implemented as safety bounds.
• Watchdog supervision was included as an additional safety mechanism.
• Inversion/eversion could be left compliant or mechanically constrained as clinically needed.