Imitation learning sidesteps reinforcement learning’s hardest requirements (a reward function, safe exploration, environment resets) by learning directly from examples of the desired behavior, usually collected by a human teleoperating the robot or performing the task themselves. The simplest instance is behavior cloning, plain supervised learning on state-action pairs.
Its fundamental limitation is the ceiling built into the objective: a policy trained only on demonstrations can at best match its demonstrator and has never seen its own mistakes, so it cannot learn to recover from them. Modern robot learning treats imitation as the middle of a pipeline rather than the whole of it: pretrain on demonstrations for competence, then improve with reinforcement learning on the policy’s own experience.