What is physical AI?

A chatbot lives in text. A robot lives in a kitchen, a warehouse, a field or a hospital corridor, where things have weight, friction and people standing in the way. Physical AI is the name for artificial intelligence that has to cope with that: it takes in the world through sensors, decides what to do, and changes the world with a body. This course is about how that works, from the maths of a joint to the foundation models that now steer whole robots.

A definition you can use#

Physical AI is any system that closes a loop between perception, decision and action in the physical world, and uses learning or reasoning to do it. Three words you will meet are close cousins, and the differences are small:

Nothing here is a hard boundary. A robot vacuum is robotics, an autonomous system and, when it learns the layout of your home, physical AI. Use the words the way the people around you do, and look at what the machine actually does.

The loop every robot runs#

Strip any robot down and you find the same cycle, repeated many times a second:

1. Sense. Cameras, microphones, force sensors, wheel encoders, an inertial unit and more turn the world into numbers.
2. Think. Software works out what those numbers mean (where am I, what is that object, is the cup full) and decides what to do next.
3. Act. Motors, valves or propellers carry out the decision, which changes the world.
4. Repeat. The next sensor reading shows what the action really did, which is often not what was planned.

The fourth step is the one that makes robotics different. In software you call a function and trust its result. A robot commands a motion and then looks to see whether it happened, because the floor was slippery, the object was heavier than expected, or someone moved it.

Why a robot is not just a chatbot with wheels#

Language models have shown that a single large model trained on huge amounts of data can do an astonishing range of tasks. It is tempting to think a robot only needs the same recipe. It does need the recipe, but the physical world changes the problem in six ways:

• No undo. A wrong sentence can be regenerated. A dropped glass cannot. Mistakes cost time, hardware and sometimes safety.
• Real time. A balancing robot that thinks for a second falls over. The answer has to arrive when it is needed, not eventually.
• Partial, noisy information. A camera sees one side of an object, in this light, at this angle, with sensor noise. The robot must act on a best guess.
• Physics. Contact, friction, flexible objects and gravity are hard to model exactly, so predictions drift away from reality.
• Scarce data. The internet holds trillions of words but very few recordings of a robot arm folding a towel. Collecting robot data costs real time on real machines.
• Safety and liability. A machine that moves near people must be safe when everything works and when it does not.

Each of these returns throughout the course. Control theory exists because of real time and physics. Probability and filtering exist because of noisy, partial information. Simulation exists because of scarce data and no undo. The safety chapters exist because the cost of a mistake is not a bad paragraph.

What this course covers#

The course follows the stack from the ground up:

• the maths and physics that describe motion (chapters 7 to 15),
• the hardware: motors, sensors and computers,
• control, perception and planning, the classical core of robotics,
• manipulation, using hands to do useful work,
• learning and foundation models, the part that has changed fastest,
• simulation and software such as ROS 2,
• a tour of the robots that exist today, from industrial arms to humanoids,
• people, safety and society, and
• projects you can build yourself.

If you already build software agents, the agent course shares ideas with this one (a loop, tools, evaluation), and a later chapter, 81, connects the two.