imitation learning

Imitation learning is teaching a robot by showing it the right thing to do, the way a child learns to whisk eggs by watching a parent rather than by guessing wildly until the kitchen is a mess. Instead of letting the robot blunder through endless trial and error chasing a reward, you give it recordings of an expert doing the task well — each moment paired with the action the expert chose — and the robot trains its policy to reproduce those choices. Show it enough good examples and it learns to act like the demonstrator.

The big win is that it sidesteps two of the hardest parts of reward-based learning: you do not have to handcraft a tricky reward function, and the robot does not have to crash thousands of times to discover what works, because the expert already knows. The classic catch is that a pure imitator only knows the situations it was shown. Let it drift somewhere the expert never went — a slightly odd angle, a spot it never quite reached — and it has no demonstration to copy, so small mistakes can snowball. Modern methods fight this by mixing in a little correction or extra practice, but at heart imitation learning is the simple, powerful idea of learning by example.