Robustness to severe distribution shift in zero-shot HAR

Improve robustness to severe distribution shift in zero-shot transfer for the HALO IMU foundation model, particularly on datasets such as HARTH and VTT-ConIoT.

Background

HALO is designed to generalize across heterogeneous IMU datasets and support open-vocabulary activity recognition without per-dataset adaptation. However, the experiments show that severe distribution shifts remain problematic: zero-shot performance collapses on HARTH, whose back- and thigh-mounted accelerometer data differ substantially from the training distributions, and on VTT-ConIoT, which contains industrial activities that are largely absent from the training label vocabulary.

The authors therefore identify robustness under severe sensor and activity-distribution shift as an unresolved limitation and future research direction, specifically involving zero-shot transfer to datasets such as HARTH and VTT-ConIoT. Although supervised adaptation partially recovers performance in some settings, the zero-shot challenge remains unresolved.