Where WiFi Sensing Fails:A Reproducibility Study of CSI-IMU Substitution for Human Activity Recognition

We investigate whether ambient WiFi Channel State Information (CSI) can substitute for a body-worn inertial measurement unit (IMU) in human activity recognition (HAR), using the Multimodal Activity Sensing Dataset (MASD; 27 activities, 20 participants). We find that the dataset's released machine-learning-ready files expose a degraded signed CSI representation rather than the documented amplitude, invalidating conclusions drawn from the files as shipped. We rebuild proper amplitude and antenna-ratio Doppler from the raw complex CSI, recover omitted participant identifiers, and re-evaluate subject-independently. Under a controlled, multi-seed evaluation, WiFi is systematically weak: every backbone we try reaches 6-11% weighted accuracy on the 27-class set, against 73-80% for IMU, with usable accuracy on 0 of 27 activities. The corrected representation mattered (+14 percentage points on the easy subset), yet WiFi still cannot substitute, nor does fusing it with IMU beat IMU alone. A well-regularised gated fusion model degrades gracefully instead, closing its gate and becoming invariant to WiFi. A 27K-parameter feature model matches the deep network at 11x fewer parameters but fails on the same activities WiFi cannot recover. We provide the first at-scale account of where and why substitution fails, and release a corrected, subject-labelled CSI pipeline for MASD at https://github.com/Joost080/masd-wifi-imu-substitutability.