behavior cloning

Behavior cloning is the simplest way to teach a robot a skill by example: you record an expert doing the task, and then train the robot to copy what the expert did. Each recorded moment is turned into a single flashcard — on the front is what the robot was seeing and feeling at that instant (the camera image, the joint positions, the gripper state), and on the back is the exact action the expert took next (move this way, close the gripper, turn the wheel). The robot studies thousands of these flashcards and learns a rule of thumb that maps 'this is what I see' straight to 'this is what I should do.' Because it is just matching answers to questions, this is ordinary supervised learning, the same kind of pattern-fitting used to label photos — only here the labels are actions instead of names.

The appeal is that it needs no reward, no trial and error, and no simulator: a few human demonstrations can be enough to get a robot mimicking a pick-and-place or a steering motion. The catch is a quiet but serious flaw called compounding distribution shift, or compounding error. The robot only ever saw the situations the expert visited, which are tidy and on-track. The moment its own small mistake nudges it slightly off that track — a cup gripped a centimeter too far left — it faces a scene the expert never demonstrated, so its guess gets worse, which pushes it further off track, which makes the next scene even stranger. The errors snowball, and a robot that looked perfect in testing can drift into a situation it has no idea how to handle.

Practitioners fight this snowballing by collecting many more demonstrations, including deliberately messy ones that show how to recover from mistakes, or by letting the expert correct the robot live as it practices so it learns to find its way back onto the track. Plain behavior cloning is still the workhorse first attempt, though, precisely because it is so easy to set up.