Treating imitation as supervised learning
Behavioral cloning is the most direct way to turn demonstrations into a policy: collect a dataset of (state, action) pairs from expert demonstrations, then train a policy with ordinary supervised learning to predict the demonstrated action given the observed state. There's no simulation rollout, no reward signal, and no notion of long-term consequences during training, just a regression or classification problem mapping states to actions. This simplicity is exactly what makes it attractive: it reuses standard supervised learning machinery and infrastructure, and it can be trained entirely offline on a fixed dataset, without ever running the policy on the real robot during training.
For short, well-covered tasks with plentiful demonstrations, behavioral cloning can work surprisingly well. But its simplicity hides an assumption that turns out to matter a great deal: it assumes the states the policy will encounter at execution time look like the states in the training data. That assumption holds during training, where the states are literally the ones an expert visited. It does not reliably hold once the trained policy is actually running the task itself.