What is Imitation Learning? — one analogy

1The analogy

First time helping out on canning day, you stand at your neighbor's counter and just watch her hands. How tightly she packs the cucumbers into the jar, how many sprigs of dill she drops in before sealing it, how far she leaves the brine below the rim. There's barely any explanation for why she does it that way. You just copy the motion.

After a few jars, your own packing starts to look a lot like hers, and the finished jars look about the same too — even though you've never once tasted the result, since the jars won't be ready to open for weeks. That's because you carried over a skilled pair of hands instead of experiencing the outcome yourself. This is exactly the scene playing out when an AI learns by watching a person demonstrate.

2In detail

Copying the hands without tasting the result

Learning through trial and error means acting, getting a score, and correcting based on it. For canning, that would mean sealing a jar, waiting for it to cure, tasting it, and adjusting the salt next time. One round takes a long time, and you have to live with the batches that come out wrong.

Imitation learning skips that whole sequence. It gathers a pile of demonstrations, turns them into pairs of "when the cucumbers looked like this, the hands moved like that," and trains toward producing the same motion in the same situation. On the surface, this looks almost identical to training on photos with tags attached — the situation stands in for the photo, and the demonstrator's action stands in for the tag.

That's why it suits tasks where failure is costly. You can't let a car crash thousands of times to learn how to drive, but you can gather records of people driving. It's just as useful when it's hard to write down a scoring rule for what counts as doing well. Putting the conditions for a good batch of pickles into numbers is hard, but a skilled pair of hands can still be shown.

When you meet a cucumber the demo never showed

The weakness shows up in unfamiliar situations. Say every cucumber in the demonstrations happened to be perfectly firm. If an oddly soft one shows up, the person copying the motion has no idea how to adjust — it's a situation they've never seen.

What's worse is that the drift compounds. Pack the first layer a little sloppily, and the next layer doesn't resemble any scene from the demonstrations at all. That produces an even stranger set of movements, and the one after that strays further still. A small slip at the start snowballs as it rolls forward.

Filling the gaps with more demonstrations

The fix is straightforward: fold the drifted situations into the demonstrations too. Let the person copying the motion try it on their own, and the moment their hands go astray, step in and show "here's how you handle this." Add those moments to the data and retrain, and the unfamiliar territory shrinks.

Another approach is deliberately staging unfamiliar situations and demonstrating a recovery. Set out a too-soft cucumber, an over-salted brine, or a torn leaf on purpose, and show the hand that fixes it. Data mixed with recovery moments turns out far sturdier than data made only of clean, successful examples.

Getting better than the demonstrator

Imitation alone rarely surpasses the demonstrator. The good hands get copied, but so do the habits and mistakes. When demonstrations come from several different people, the mix of different styles can even average out into an awkward middle ground.

That's why real-world use runs in two stages. First, imitation gets a system to a usable level quickly. Then, trial and error with actual scoring refines it from there. Starting from a blank slate with pure trial and error means wandering for a long time, but starting from hands that already do reasonably well leaves only the fixing to do. There's also a method that works backward from the demonstrations — inferring not the actions themselves, but what the demonstrator seemed to value — and uses that as the scoring rule instead.

3More precisely

Imitation learning is a method that learns the mapping between state and action from an expert's demonstration data. Its simplest form is called behavior cloning, and structurally it's essentially supervised learning, with state as input and action as output. The problem of error compounding as the system drifts into states the demonstration never covered is called distribution shift, and it's eased by letting the learned policy act on its own, then having the expert label the resulting trajectory with the correct action afterward. The branch that instead infers the underlying reward function from a demonstration is called inverse reinforcement learning.

The analogy breaks down in places. On canning day you can ask the person next to you why they did something; imitation learning only sees the action, never the reasoning behind it. The same hand motion might be an essential step in one case and just a personal habit in another, but the data holds no distinction between the two. There's another gap: ruin one batch of pickles and you simply start the next one fresh, but in tasks with continuous flow, like driving or robotics, there's no chance to undo a drift once it happens — which is exactly where this method's weakness shows up hardest.

4Try it yourself

Tetris AI ailearn.space Watch where the blocks land and decide for yourself, first, where you'd have placed each one. Imitation learning works exactly this way — whatever move gets made in each situation becomes the training data itself

5Common misconceptions

It's easy to think imitation learning is a type of reinforcement learning, but actually it's much closer to supervised learning, pairing state with action with no reward involved at all.
It's easy to think gathering more demonstrations solves everything, but actually piling up more clean, successful examples alone doesn't shrink the unfamiliar territory — recovery moments need to be mixed in too.
It's easy to think enough copying eventually surpasses the demonstrator, but actually without a separate stage of refining through scored trial and error, it plateaus at the demonstrator's level.

7One-line summary

In shortImitation learning carries over a skilled demonstrator's action for each situation instead of experiencing the outcome directly, and how well it fills in situations the demonstration never covered decides whether it succeeds.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02