On device training with TensorFlow Lite (Micro)

I would like to understand the concept and the use of “on-device training” for new data acquired from sensor, camera, etc.
How to prepare new (unseen) data in terms of proper labeling? I assume that these data are unknown or unseen.
Do I need to prepare a batch of data, or can I just use a single sample?
Is this “On device training” applicable for supervised or unsupervised learning?
What are the limitations on On-Device Training with TensorFlow Lite
Is that concept somewhere described and explained (no reading or exploring code)?

Thank you for helping me to understand it!