Lesson video
Narration is AI-generated. Video chapters, transcript and sources
Transcript
Transcript
UNIT-01 asks
AI can write essays. So why can't it tidy up a table? This episode, we make it clear: what physical AI is, and how it really differs from the AI that chats.
From talking to doing
First, the AI that chats. It lives inside a screen: text goes in, text comes out. If it gets something wrong, just generate it again. Physical AI has to step out of the screen. Its input is the world, as cameras and sensors see it. Its output is real force, from motors. In the real world, when the blocks fall, they've fallen. There's no undo, and no starting over. So physical AI has to go around this loop in the real world, again and again: see, think, act. Harder still: if it goes wrong halfway, it still has to finish the job. One 2026 paper built two thousand recovery scenarios like this, to train and test robots. So chatting isn't the hard part. Doing things is!
Why doing is so hard
Imagine a table with a cup and a water bottle on it. The robot first has to tell them apart, then plan a path for its hand. As the hand reaches over, it has to grip the cup firmly and steer clear of the bottle. Every step has to adjust to what's in front of it. That's what makes doing hard: see what changed, decide the next step, have the motors carry it out, and keep repeating. No wonder! Tidying a table looks simple, but it takes quite a few steps. So how far can today's machines go?
What physical AI looks like
Physical AI doesn't have to look human. Robot arms in factories, carts in warehouses, self-driving cars on the road, and the humanoid robots drawing so much attention: any of them can be its body. How big is it? In 2025, industrial robots at work in factories worldwide reached about five million, more than double the number seven years before. These figures come from the International Federation of Robotics. What's more, of the industrial robots newly installed worldwide in 2025, nearly six in ten went to China. But if a part moves and the path is fixed, the robot arm may not reach it. Physical AI wants machines that can see, think, and cope with change. Let's recap, three points. One: physical AI steps out of the screen and acts in the real world. Two: acting takes seeing the scene, planning moves, and controlling motors. Three: robots come in many forms, and are trying to cope with change. So the AI that writes essays still struggles to tidy a table. Not because it isn't smart enough: tidying a table really is harder! So why now, of all times, are robots suddenly learning to use their hands? See you next time!
Evidence and sources
Practice and self-check
- Interactive quiz5 questions, explained as you answer
- Browser labno install: drag and tune it yourself
- One-page summaryprintable; opens in a new tab
The quiz is public and explains each answer as you go; answers to the guide's exercises are for signed-in members.
Why can an AI that writes poetry not tidy up a table? That is the question that opens episode 0.1 of Physical AI Lab. This article is the text version of the episode: it explains how physical AI differs from chat AI, why acting is harder than chatting, and how large today’s robot fleet is, and it gives the source and a way to check each number used in the episode.
Note: this lesson now has Chinese, English and Japanese videos, each with a transcript and sources. This article can be read on its own, and self-check questions are at the end.
1. From chatting to doing
Chat AI lives inside a screen: text goes in, text comes out. If it gets an answer wrong, you regenerate it, at a cost of a few seconds.
Physical AI has to leave the screen. Its input is the real world as seen by cameras and other sensors, and its output is real force from motors. The episode defines it as AI that can perceive the real world, make decisions, and carry them out physically.
| Chat AI | Physical AI | |
|---|---|---|
| Input | Text | The world it sees (cameras, depth, force, and other sensors) |
| Output | Text | Motor force (joint torques, velocities) |
| When it goes wrong | Regenerate | Something has already fallen over |
The last row is the key. The real world has no undo button: once the blocks fall, they have fallen; once the cup is knocked over, the water is spilled. So physical AI has to run the same loop in the real world again and again: see, think, act. See the current state, decide on the next step, move, then look at the result. That is a closed loop.
Sign in to keep reading: the rest of this lesson, its self-check questions and the answers.
Limitations
- This episode is an L0 introduction and deliberately leaves out technical detail. “See, think, act” is a simplified description; in real systems, perception, planning, and control are often intertwined and run at different frequencies. The planned episodes 0.3 (see, think, act) and 1.2 (control and control frequency) will cover this in detail.
- Moravec’s paradox is an empirical observation, not a theorem. “A billion years” is Moravec’s own estimate, and this article cites it as R-level supplementary reading.
- The IFR statistics cover industrial robots and do not include service robots, self-driving cars, or humanoid robots.
Links
- RoboRecover: Benchmarking Robot Policy Recovery under Execution Deviations (arXiv 2609.28952)
- Hans Moravec, Mind Children (1988): bibliographic information
- IFR: Five Million Robots Now Operate in Factories Globally (World Robotics 2026 press release)
- Physical AI Radar (uses the same O/R/M evidence levels)
