Robots and world modelsMiddleLesson 2586 min read

Why robots lag behind chatbots

Chatbots learned from a huge amount of text. Robots have far less data, a messy world, real safety risks and costly hardware. Here is why the gap exists.

✦ Ask the tutor 🧪 Lab · 20 min 📝 Lesson test

Speed

Lesson in motion

In 60 seconds

Why robots lag behind chatbots

Chatbots learned from a huge amount of text. Robots have far less data, a messy world, real safety risks and costly hardware. Here is why the gap exists.

1/4

In simple words

A chatbot can write a poem about making chai in one second. But ask a robot to actually make the chai: find the pan, light the stove, pour milk without spilling, stop before it boils over. That is much, much harder for a machine.

In recent years, language models became good at writing, summarising and answering questions. Robots that work in homes and fields have improved too, but far more slowly. This lesson explains why. The reasons are not only about clever algorithms. They are about data, the physical world, safety and money.

Moravec's paradox: the easy things are hard

In the 1980s, robotics researchers noticed something strange. Things that feel hard to people, like chess or algebra, turned out to be fairly easy for computers. Things that feel easy to people, like walking on a crowded street or picking up a soft fruit, turned out to be very hard. This idea is called Moravec's paradox, after the robotics researcher Hans Moravec.

One common explanation: animals and people have had a very long time of evolution to get good at seeing, balancing and grasping. We do these things without thinking, so we do not notice how complex they are. Abstract reasoning is newer for us, so it feels hard, even though the rules are fewer and cleaner.

Feels hard to a person	Feels easy to a person
Winning at chess	Picking up an egg without breaking it
Solving a long maths sum	Walking across a wet, uneven field
Writing a summary of a report	Folding a crumpled shirt
Translating a paragraph	Tying a knot in a thread

Note

The right column is where robots still struggle most. Many of the "easy" tasks need touch, balance and quick correction, not just seeing.

Reason 1: very little robot data

Language models learned from a huge amount of text that people had already written and put online. There is no internet of robot actions. Nobody has been quietly recording, for decades, the joint angles and grip forces of robots doing chores.

Robot data must usually be collected on purpose, often by a person guiding the robot, one task at a time.
Data from one robot may not fit another robot with a different arm, gripper or camera.
Videos of people doing tasks help a little, but a video does not show the forces, and a human hand is not a robot gripper.

Reason 2: the physical world is messy

Text is clean: a word is a word. The physical world is not. Light changes. Objects slide, bend, tear and spill. Cameras get dust on them. A cup can be full or empty, and look the same from above. Every home and every farm is laid out differently.

A chatbot's mistake is often a wrong sentence you can ignore. A robot's mistake happens in the world, right now, and it cannot press "undo". This is the lesson from Grounding: finding out you were wrong: the world tells a robot when it was wrong, but sometimes only after something breaks.

Reason 3: safety

A robot arm is heavy, strong and fast. A wrong move can hurt a person, a pet or itself. So robots must be tested much more carefully than a chat app, and they often move slowly on purpose. It is also hard to let a robot "try things and learn from mistakes" in a real home.

Reason 4: hardware costs and breaks

Motors, sensors, batteries and hands cost real money, and they wear out.
Collecting data needs real robots running for many hours, with people watching.
Copying a software chatbot to a million users is cheap. Building a million robots is not.

Chatbot

Learns from text people already wrote
A mistake is a wrong answer on a screen
Can be copied to millions of users almost for free
Can try again instantly

Robot

Needs new data collected on purpose
A mistake can break things or hurt someone
Every copy is a real machine that costs money
Each try takes real time in the real world

Honest status

Robots already do useful work in places that are controlled and repeatable, such as some factory lines and warehouses, where the same part arrives in the same place again and again. The hard frontier is general-purpose robots that can handle new objects, new rooms and new tasks. Researchers are making progress, but nobody knows how fast the gap will close.

Watch out

Robot demo videos can mislead. A clip may be sped up, remote-controlled by a person, or the best of many tries. When you watch one, ask: was it autonomous? At normal speed? How often did it fail?

Try this. Pick one chore you do at home. List every small step and every thing that could go differently each time (where objects are, light, spills). How many surprises would a robot need to handle?

Watch and read more

Lab

A paper breakdown of one household chore, showing where a robot would struggle.

~20 min

The problem

Choose one chore: making tea, folding clothes, washing a steel plate, or watering plants in pots.

Write every step as a list, as if instructing someone who has never seen a kitchen or garden.
For each step, write what could be different each time (position, size, wetness, light, a person nearby).
Mark each step: S if it mostly needs seeing, T if it needs touch or force, P if it needs planning.
Circle the two steps you think are hardest for a robot and explain why in one sentence each.
Write one way to make the task easier for a robot by changing the setting (for example, always keeping the cup in the same holder).

You are done when

The list has at least eight small, concrete steps
Each step names real sources of variation, not just "it might be hard"
The hardest steps are explained with touch, force, messy objects or safety, not only with seeing
The "change the setting" idea shows why controlled places like factories are easier

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1A robot does a task perfectly 95 times out of 100 in a lab. Why might that still not be good enough for a home?Reveal

Because it would do the task many times every day, so failures would come often. It also matters how it fails: dropping a towel is fine, dropping a hot pan is not. The home is also less controlled than the lab, so the real success rate may be lower. What matters is the rate of serious failures, in real conditions.

Q2Why might a robot that is excellent at one task, in one factory, not help much with general-purpose robots?Reveal

Because it may succeed by relying on the setting being fixed: the same part, same place, same light. That removes exactly the variation a general robot must handle. Its data and skills may not transfer to new objects, rooms or bodies. Specialised success and general ability are different problems.

Please sign in to continue.

Questions people ask

If language models are so smart, why not just put one inside a robot?

A language model can help a robot understand an instruction and plan steps. But knowing that you should lift the cup is different from knowing how: exactly where to put the fingers, how hard to squeeze, and how to react when it slips. That low-level control needs different data and fast feedback from sensors.

Is Moravec's paradox a law of nature?

No. It is an observation that researchers made, and it has held up well so far. It does not say these tasks are impossible for machines, only that they turned out to be much harder than people first expected.

Can't robots learn from YouTube videos of people?

Researchers are trying this, and videos can teach something about what objects are and how tasks go. But videos do not show forces or touch, the camera angle is different from the robot's, and human hands are not robot grippers. So video is usually used together with real robot data, not instead of it.

Are robots useless then?

Not at all. Robots do valuable, repeated work in many factories and warehouses. What is still hard is a robot that can walk into any home or field and do many different tasks well.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.