imitation learning
/ im-ih-TAY-shun LER-ning /
The fastest way to teach someone to drive isn't a rulebook — it's to sit them beside an experienced driver and have them copy what the expert does in each situation. Imitation learning takes the same shortcut for machines: rather than letting a system discover good behavior through endless trial and error, you give it recordings of an expert doing the task and train it to mimic them.
Each demonstration is a stream of situations and the actions the expert took in them. The simplest form, called behavioral cloning, treats this as ordinary supervised learning: the situation is the input, the expert's action is the correct label, and the model learns to map one to the other. This sidesteps the hardest part of reinforcement learning — designing a reward signal and exploring safely — which is why imitation learning is popular in robotics and self-driving, where random trial and error in the real world would be dangerous or impossibly slow.
Its great weakness is a snowball problem. The model only ever saw situations the expert visited; the expert, being good, rarely got into trouble. So the first time the model makes a small mistake and drifts into an unfamiliar state, it has no demonstration to copy, makes a worse choice, drifts further, and errors compound. Fixes exist — letting the expert correct the model in the states it actually reaches, or blending in reinforcement learning — but plain imitation can also only ever be as good as the demonstrator, and it quietly copies the expert's bad habits along with the good ones.
To teach a robot arm to pour coffee, a person guides the arm through the motion a few dozen times while the system records every joint position and movement. The robot then learns to reproduce the pour. But if it slightly overshoots into a pose no demonstration covered, it can flail — because it only knows how to copy, not how to recover.
Copy the expert — but it has no script for states the expert never reached.
Plain imitation can only be as good as the demonstrator, and copies their bad habits too. Its classic failure is compounding error: one small mistake leads to an unfamiliar state with no demonstration to copy, and things spiral.