Human Activity Recognition (HAR) systems are increasingly used in portable and resource-constrained devices. However, most existing approaches rely on models that are trained externally and deployed to devices only for inference. This limits the system’s ability to adapt to individual users and creates potential privacy concerns, as personal data must be shared for training. In this study, we aim to overcome these limitations by enabling learning directly on the device. We explore three learning configurations: using a general model without adaptation, updating only selected parts of the model with local data, and fully retraining the model for complete personalization. Additionally, we implement a collaborative approach where multiple devices contribute to building a shared model without exchanging raw data. Our goal is to support flexible, adaptive, and privacy-preserving HAR without relying on external servers or centralized training.