You didn't learn to tie your shoes from a manual. Neither will robots.

Think back to how you learned to tie your shoes. Almost certainly not from written instructions. Someone sat beside you, did it slowly, maybe a dozen times, and you copied them until your fingers found the motion on their own. That is the oldest learning method on the planet, older than language itself: watch someone who can, then imitate. It is now the dominant way we teach robots, and hiding inside it is a puzzle that turns out to be surprisingly deep.

The method has a name, learning from demonstration, and it has quietly taken over robotics. This is a guide to what it is, the puzzle at its center, and why the whole approach now lives or dies on a single thing: not the cleverness of the algorithm, but the quality and quantity of the demonstrations.

Show, don't program

Learning from demonstration means teaching a robot a task by showing it examples instead of writing out the instructions. The older way to make a robot do something was to program every motion by hand: move to these exact coordinates, close the gripper this far, rotate by that many degrees. That works on a factory line where nothing ever changes. It falls apart the moment the world gets messy, because no one can hand-write a rule for every way a sock might be crumpled or a cup might be turned.

Demonstration flips that around. You show the robot a task being done, many times, and it learns the mapping from what it sees to what it should do. The simplest version is called behavioral cloning: for every situation in the demonstration, copy the action the demonstrator took. See this, do that. It is the same instinct as a toddler mirroring a parent, turned into math.

This is not a fringe technique. Essentially every major robot foundation model of the last few years was taught this way, including Google DeepMind's RT-2 and the models from Physical Intelligence, all trained on large piles of demonstrations rather than hand-written rules. Programming does not scale to the real world. Showing does.

The puzzle nobody mentions: whose body?

Here is the deep part. When you copied the shoe-tying, you were watching someone else's hands, not your own. Your brain quietly solved a translation problem: their fingers are not your fingers, their angle is not your angle, yet somehow you mapped their motion onto your own body without thinking about it. Roboticists call this the correspondence problem, and for a machine it is far harder than it ever was for you.

A robot watching a human faces the widest version of the gap. A human hand has five fingers, dozens of joints, and proportions nothing like a typical robot gripper with two or three stiff digits. So when a robot sees a person pinch and twist a bottle cap, it cannot simply replay those finger movements, because it does not have those fingers. It has to work out what its own, very different body should do to get the same result. Watch a left-handed person tie a knot and you have to mentally mirror them. A robot has to do that too, except the mirror is its entire anatomy.

There are two ways to deal with this, and they turn out to be the two great schools of teaching robots.

Strategy one: dodge the problem with teleoperation

The cleanest way around the correspondence problem is to never have it. In teleoperation, a person does not demonstrate with their own hands at all. They puppeteer the robot's body directly, through controllers or a matching pair of arms, while the robot records its own joints moving. There is no human-to-robot translation to do, because the demonstration was already in the robot's body from the start. The data comes out pristine.

The price is scale. Every demonstration needs the physical robot, in real time, with a skilled human driving it. You feel that cost fast. The DROID dataset, a serious teleoperation effort, took 50 people across 18 labs working 12 months to gather 76,000 demonstrations, and much of that time went to resetting the scene rather than teaching. Teleoperation buys you perfect data one robot-hour at a time, and it can never be scraped or downloaded. You earn every second of it.

Strategy two: solve the problem with human video

The other school accepts the correspondence problem and does the hard translation work, because the prize on the far side is enormous. Instead of puppeteering a robot, you record ordinary people doing tasks, ideally from their own point of view with a camera on the head or in the hand, and then you translate that human motion into something a robot can use: retargeting the tracked hand onto the robot's gripper, grounding everything in real 3D space, and recovering the forces a camera cannot see.

It is harder, but it scales to the entire human race instead of one robot at a time, and the evidence now says it works, often better than the expensive alternative. Training on 20,854 hours of first-person human video raised robot dexterity 54% along a clean scaling curve. Head to head against the same volume of teleoperation data, human video won, by roughly 52% on familiar tasks and 90% on tasks the robot had never seen. One system even taught a robot seven tasks from smart-glasses footage with no robot used in the collection at all. We went deeper on this in why human data beats more robots.

The real bottleneck

For years the assumption was that better imitation needed a smarter algorithm, that the secret was some clever new learning trick. The frontier has stopped believing that. The algorithms are now good enough that the thing holding robots back is no longer how they learn, it is what they have to learn from. The bottleneck is the data, not the algorithm.

And not just any data. A million sloppy demonstrations of the same easy grab teach a robot less than a few thousand diverse, well-captured ones that include the moments things went wrong and a hand had to recover. Good demonstrations are grounded in real 3D, carry the forces vision misses, cover a genuine range of objects and situations, and keep the failures most datasets throw away. The craft of producing them is the actual work, and we broke it down in what turns a video into training data.

The part that will age well

There is a final dimension that the demonstration era forces into the open. A demonstration is not scraped off the web; it is performed by a real person who chose to do it. That makes the old questions unavoidable, in a good way: did they agree to it, were they paid fairly, can you trace exactly who taught the robot what. The data that trained today's chatbots was largely taken without consent. One audit found that about 60% of a giant dataset came from sites that forbid scraping, and no major labeling platform cleared even a basic minimum-wage bar. Robot demonstrations are being created from scratch right now, which means this time it can be done with consent, fair pay, and full provenance from the very first frame.

The bottom line

Learning from demonstration is just the oldest idea in teaching, applied to machines: do not explain, show. The hard part was never the showing. It is the translation from a human body to a robot one, and the unglamorous work of turning a demonstration into something grounded, force-aware, diverse, and honest about its failures.

The teams that win will not be the ones with the cleverest imitation algorithm, because those are close to a commodity now. They will be the ones who can produce the best human demonstrations at scale: real, grounded, recovery-rich, and collected with consent. You learned by being shown. So will every robot worth trusting. The only question left is whose hands do the showing, and whether they agreed to it.