Behavioural Cloning

You know how some people learn to cook by reading recipes, and others just watch their mum do it and go, “Yep, got it”?
Behavioural cloning is the second one. For robots.

The basic trick

Instead of teaching a robot with a giant list of rules, you record a human doing the task. Pick up the cup. Open the door. Fold the T-shirt without creating a fabric crime scene.

Then you train the robot to copy those actions directly from the data.

That’s behavioural cloning. The robot looks at examples of what the demonstrator did, then learns, “When things look like this, do that.”

It’s basically imitation learning with a fancy haircut.

Like copying your driving instructor

Imagine you’re learning to drive by watching hours of dashcam footage plus steering, braking, and turning data.

• Red light appears, foot hits brake
• Parking spot appears, wheel turns
• Toddler on scooter appears, everyone prays

The system learns the pattern between what was seen and what action happened next. Not the grand theory of driving. Just the behavior.

That’s the “cloning” part. You’re cloning actions, not building a tiny philosopher.

Why people like it

It’s simple. Weirdly simple.

If you already have good demonstration data, behavioural cloning can work fast. No need to hand-code every tiny instruction like you’re assembling IKEA furniture with a legal contract.

It’s especially useful for humanoid robots, because a lot of human tasks are easier to show than explain. Try writing down how to tie shoelaces. You’ll age.

The catch

It only learns from what it sees.

So if the robot runs into a situation that wasn’t in the training data, it can get confused fast. Like a bad driver in light rain.

That’s why behavioural cloning matters. It’s one of the fastest ways to turn human know-how into robot behavior, which is great right up until the real world does what it always does, and gets weird.