<p>Human activity recognition (HAR) has achieved remarkable performance in the task of semi-supervised learning (SSL) paradigm. However, most existing SSL for HAR models assume that unlabeled data contains only classes previously encountered in the labeled training data. Their performance can be seriously degraded when this assumption is violated in practice. To tackle this challenging problem, we propose a novel unseen-aware semi-supervised model for robust HAR. Specifically, we first deploy the open-set semi-supervised learning technique, that identify examples of novel/unseen categories as outliers (known categories as inliers), to make SSL more realistic and practical. Then, we try to improve the robustness of model via designed weighting functions, leading to alleviate the performance drop caused by class distribution mismatch. Finally, we formulate our problem into the bi-level optimization framework and demonstrate the efficiency of its optimization algorithm. Theoretically, the performance of our proposed model has been guaranteed with efficient generalization in compared to the learned model merely with labeled data. We conduct extensive experiments on benchmark HAR datasets to demonstrate the superiority of our proposed model.</p>