This article presents the development and validation of a polymer optical fiber (POF) sensor system for quantifying human–robot interaction (HRI) torque in knee exoskeletons. The device consists of a four-channel POF array embedded in a shank cuff that provides online torque signals during gait tasks. The channels were calibrated against a reference load cell, yielding dynamic force RMSE in the range 3.93–5.14 N and correlation between 0.88 and 0.94. In quasi-static interface tests with six participants, the reconstructed POF torque tracked full torque transfer through the cuff, with RMSE of 1.4–1.8 N ⋅ m for sinusoidal and triangular actuator profiles. In treadmill trials, the POF-based torque was used as a reference to assess disturbance observer (DOB) techniques. All observers reproduced the main interaction-torque patterns but showed attenuation and temporal shifts with respect to the POF reference, with RMSE between 0.8 and 1.3 N ⋅ m. These results indicate that the POF cuff delivers a consistent interaction-torque signal at the interface and can serve as a practical reference when analyzing bandwidth and robustness limitations of observer-based estimators.