The Rehabilitation of Why

We removed purpose once before

There is a historical pattern here. Ancient philosophy treated purpose as essential to explanation: you had not fully explained something until you had explained what it was for. Modern science rejected that, and for physics, rightly so. "Why?" gave way to "What happens?" and "How does it work?" — a trade that built the modern world.

Psychology made a similar move. Early behaviorism tried to explain behavior without inner structure, reasons, or goals. Only observable actions counted. What worked got repeated.

That inheritance is still visible in robot learning today. Imitation learning reproduces demonstrations. Reinforcement learning scales reward-following. Both can be powerful. But neither requires the system to represent reasons.

And when a learner cannot represent reasons, it becomes vulnerable to something very old: superstition.

When a system cannot ask why, it learns the wrong thing

B. F. Skinner's pigeon experiments made this vivid. Feed pigeons on a timer and they start repeating whatever motion happened to coincide with the reward. Coincidence hardens into ritual. He called it plainly: superstition.

A learner that cannot represent why an outcome happened will learn whatever merely correlates with it. In robotics, that can mean copying surface features instead of task structure. A system may learn to flip the blue-looking items, when the real reason was to inspect labels.

This is what learning looks like when why is unrepresentable.

This is what learning looks like when why is unrepresentable.

Children do something different

Developmental psychology offers a very different model. Even before they can speak, infants do not just track movement. They infer goals. In a classic study, twelve-month-olds watched an animated object leap over a barrier to reach another object. Once the barrier was removed, infants were surprised if the object still took the same inefficient leap. They expected it to move directly instead.

The infant was not memorizing a trajectory. The infant had already inferred a goal, a constraint, and a principle of efficiency.

Developmental psychologists call this the teleological stance: understanding actions as means to goals under constraints.

If we want robots to generalize beyond demonstrations, reproducing motion is not enough. They need to infer what the action is for.

The hardest question in imitation is not how to copy, but what to copy

In a well-known experiment, infants watched an adult switch on a light using her forehead. Later, many copied the same action. But when the adult's hands were visibly occupied, infants mostly used their hands instead.

The children were not blindly imitating. They were interpreting. If the demonstrator's hands were occupied, the forehead action was just a workaround. If the hands were free, the method must have mattered.

Before copying, the infant was already reasoning about why the demonstrator acted that way.

Every imitation-learning system faces the same dilemma. Not just: how do I reproduce this movement? But: which parts belong to the task, and which belong to the person who demonstrated it? That distinction is everything.

Some actions are meaningful. Some are incidental. Some are not yet understood.

Factories are full of actions that are easy to see and hard to interpret. A wrist rotation might be essential. It might be ergonomic. It might reflect an invisible constraint like label visibility, grip stability, or a safety habit.

Children show another side of this in what researchers call over-imitation. In studies comparing children and chimpanzees, children often copied even apparently unnecessary steps — especially when those steps may have carried hidden purpose or social meaning. That is not stupidity. Sometimes a step looks unnecessary but is not. Sometimes the visible action carries an invisible rule.

For robotics, this suggests something important. A good learner should not collapse everything into one category. Some steps should be abstracted. Some should be copied faithfully. And some should remain explicitly not yet understood. That is a far more honest — and far more useful — state than false certainty.

Where the why lives

In our view, why should not remain an intuition floating outside the system. It needs a representation. In practical terms, that means purpose must live inside the structure of the learned program itself — in the conditions that govern loops and branches.

"Rotate until the label is visible" is not just a movement description. It is a purpose hypothesis.

Once a reason is represented explicitly, three things become possible.

It becomes testable. If the hypothesis is wrong, the condition never becomes true, and the system can detect that it does not know.

It becomes transferable. A surface rule like "flip the blueish ones" breaks when appearances change. A purpose rule like "find the label" survives variation.

It becomes discussable. The system can ask a human a meaningful question: "You rotate items until you can see the label, correct?"

That moment matters. Because demonstration is not just data collection. It is often teaching. And a robot that can surface a legible purpose hypothesis is no longer treating demonstration as surveillance footage. It is participating in instruction.

"Rotate until the label is visible" is not just a movement description. It is a purpose hypothesis.

Observation is not enough

There is one more lesson children make impossible to ignore: not everything can be learned by watching. No child learns to ride a bicycle from observation alone. Watching helps with the what. Sometimes it helps with the why. But the how comes from repeated interaction with one's own body.

Children practice, vary, and recalibrate. Development divides labor: intent is learned socially and cheaply, by watching others; competence is grounded privately and expensively, on one's own hardware.

Robotics needs the same division: learn intent socially, ground competence physically, connect both through explicit reasons.

The real question is not whether the robot can imitate

The real question is whether it can understand enough to adapt. For engineering purposes, a why has earned its name when it compresses past behavior, predicts what happens under changed conditions, generalizes along the reason rather than the surface pattern, and survives being checked with the human who knows the task. By that standard, why is not philosophical decoration. It is the difference between a system that repeats behavior and a system that can be trusted across variation.

And variation is the real environment of work.

The capability robotics still underestimates

Robot learning has made huge progress on observation. It is becoming more honest about embodiment. But on purpose — on why — it is still early. A century of watching children suggests the shape of the solution is not optional. Learn what from others. Learn how from the body. Learn why by reconstructing reasons — and check those reasons wherever possible by asking.

The most important capability we may be able to give a learning robot is not better imitation, faster optimization, or more data.

It may be something much simpler. Something every three-year-old already does relentlessly.

The ability to ask why — and a representation in which the answer can be checked.