FMCW radar has emerged as a promising technology for human activity recognition (HAR) due to its high resolution and rich signal fidelity. However, the strong performances reported in ideal settings rarely hold when models face new environments, subjects, or hardware parameters. These variations induce data distribution shift, commonly referred to as domain generalization (DG) or out-of-distribution (OOD) generalization problems. Despite rapid advances in DG research, these methods have not yet been applied to FMCW radar HAR. To address this limitation, this paper applies multiple DG algorithms to the public Glasgow dataset with two different data splitting criteria, providing a comprehensive assessment of robust cross-domain radar HAR. The results show that causality-inspired algorithm consistently improve traditional algorithm in both splits. On the other hand, aligning features across domain and loss-based regularizations have a weaker impact on improving the baseline.

Domain generalization, out-of-distribution, FMCW, human activity recognition