<p>Human Activity Recognition (HAR) is a critical area in mobile health and context-aware computing, providing foundational technology for applications such as nurse-care monitoring, rehabilitation assessment, and athletic performance evaluation. Existing HAR methods predominantly rely on manual feature extraction or employ Convolutional Neural Networks (CNNs) for automated feature learning, yet they remain sensitive to real-world challenges like sensor noise, variability in sensor placement, and inter-subject variations. This study introduces a novel Transformer-based architecture inspired by the Kolmogorov-Arnold representation theorem, designed to robustly model complex spatiotemporal dependencies from heterogeneous sensor data. Comprehensive evaluations on standard HAR datasets demonstrate that our Kolmogorov-Arnold Transformer architecture outperforms traditional machine learning techniques and contemporary deep learning methods, delivering superior accuracy and enhanced robustness to noise and sensor perturbations. This approach thus offers significant potential for advancing practical HAR applications in diverse, real-world environments.</p>