Behavior cloning treats control as a supervised learning problem: collect demonstrations of an expert performing the task, then train a network to predict the expert’s action from the state, exactly as you would fit any classifier or regressor. It inherits supervised learning’s simplicity (no reward function, no exploration) and its blind spot: the guarantee only holds on the training distribution.
Because a policy generates its own future inputs, small mistakes push it into states the demonstrations never covered, where it errs worse. This compounding error grows quadratically with the task horizon in the worst case, and taming it (with interactive corrections like DAgger, or architecturally with action chunking and generative action heads) is the central theme of imitation learning for robotics.