On-device training (ODT) enables deep learning models to be trained directly on resource-constrained edge devices, facilitating adaptation to new data post-deployment. Despite its emergence as a promising technology, practical implementation strategies for ODT remain underexplored. This work addresses this gap by experimentally investigating several promising ODT solutions. Through direct experimentation, the implementation effort, capabilities, and performance of these solutions are evaluated and compared within the context of a realistic on-device training task.