Where you are. This is the first page of the course. You need no robotics background and no maths yet. By the end of it you will have a single mental model that every later lesson hangs off.
Reach for something without looking
Put your hand out, off to the side, and pick up whatever is there. A mug, your phone, the edge of the desk. Do it without turning your head.
Notice what happened. You did not compute the whole motion in advance. You moved roughly toward where you believed the thing was, and then you kept adjusting: your arm reported back where it actually was, your fingers reported contact, and you corrected as you went. If the mug had been two inches further away than you expected, you would have found it anyway.
Now notice the harder thing. You cannot un-knock-over the mug. There is no undo. Every correction had to happen before the mistake finished.
That is the whole subject. Everything in this course is a refinement of what your arm just did.
The idea in one paragraph
A robot is a machine that runs a loop: it senses the world, decides what to do, and acts to change the world, then senses again to find out what actually happened. That is it. Everything else in robotics is a refinement of that loop. The hard part is not the loop; it is that the physical world is noisy, delayed and unforgiving compared to software. Sensors lie a little. Motors do not do exactly what you asked. And nothing can be rolled back.
Wider than the screen; scroll it sideways.
This loop runs fast. A typical arm controller runs somewhere between 50 and 1000 times per second, which means the “decide” step often has a budget of a few milliseconds. Hold that number. It is the reason robot architectures split into a fast part and a slow part, and it is why you cannot simply put a large neural network in the loop and call it done.
The parts, named once
Here is the vocabulary. You will use these words for the next twelve months, so it is worth twenty seconds each.
Wider than the screen; scroll it sideways.
Why six is the magic number
Placing an object anywhere in space takes six numbers: three for position and three for orientation (how it is rotated). So six degrees of freedom is the minimum needed to reach any nearby position in any orientation.
Fewer than six and some poses are simply unreachable, no matter how good your software is. More than six, like your own arm’s seven, means extra flexibility: several different joint configurations reach the same pose, and you get to pick the one that avoids the obstacle.
What the robot knows about itself
Two words you will see constantly, worth separating now.
Almost every hard problem in robotics lives on the exteroception side. A robot always knows its own joint angles to a fraction of a degree. What it struggles with is knowing that the thing in front of it is a mug, where exactly the mug is, and whether it is slippery.
The robot you will actually build
The machine at the centre of this course is a pair of small arms with six servos each. You move one of them with your hand, and the other copies it. Three names come with that sentence.
That mirroring is how you will record demonstrations in Module 4, which a neural network then learns to imitate. The same concepts scale unchanged to a hundred-thousand-dollar industrial arm or a humanoid: more degrees of freedom, better motors, same loop, same maths.
What makes this hard, honestly
Software people arrive expecting the hard part to be the AI. It is usually these instead.
| The problem | Why it hurts |
|---|---|
| Partial observability | A camera sees a 2D projection of a 3D world, with things hidden behind other things. State must be inferred, never simply read. |
| Noise and drift | Every sensor reading is a little wrong, differently wrong each time, and the wrongness piles up. |
| No undo | A dropped glass stays dropped. Exploration is expensive, which is exactly why simulation matters so much. |
| The real-time budget | Decisions in milliseconds, forever, without pausing. Unlike a chatbot, the world does not wait for you. |
| The rare cases | The demo works on the ten objects you tested. Object eleven is shiny, and everything breaks. |
That last point is worth sitting with. The reason robots are everywhere in factories and almost nowhere in kitchens is not that factories are more important. It is that factories are arranged. The next lesson is the story of what happened every time someone tried to leave the factory.
Review
One loop, forever, against an uncooperative world
A robot is a machine that runs a loop: it senses the world, decides what to do, and acts to change the world, then senses again to find out what actually happened. That is it, and everything else in robotics is a refinement of that loop. If you have written a controller that reconciles desired state against observed state, you already know the shape. What breaks the analogy is everything that makes robotics hard. Your cluster is physics. Observation is noisy and partial rather than an authoritative store you can query. Actions take real wall-clock time and cannot be retried for free. And there is no eventual consistency: the loop has a deadline, every time.
The hard part is not the AI
Software people arrive expecting the hard part to be the intelligence. It is usually these instead. Partial observability: a camera sees a flat projection of a solid world with things hidden behind other things, so state must be inferred and never simply read. Noise and drift: every reading is a little wrong, differently wrong each time, and the wrongness piles up. No undo: a dropped glass stays dropped, which is exactly why simulation matters. The real-time budget: decisions in milliseconds, forever, and unlike a chatbot the world does not wait for you. And the rare cases: the demo works on the ten objects you tested, and object eleven is shiny. Working on the eleventh object is the central obsession of modern robotics.
Check yourself
1. Why does a 4-DOF arm fundamentally limit what you can do, no matter how good your software is?
Placing an object at an arbitrary nearby position and orientation needs six independent numbers: three for position, three for orientation. With four degrees of freedom you can control four of them, so some position-and-orientation combinations have no joint configuration that reaches them. Software cannot create degrees of freedom the mechanism does not have.
2. Your servo accepts “go to this angle” commands. What are you giving up compared to a motor you command with force?
Direct control over force. The servo’s internal controller pushes as hard as it needs to in order to reach the angle you asked for, which is dangerous on contact: it will crush or fight whatever is in the way. Commanding force instead lets you say “push, but with at most this much.” That is what makes gentle, compliant behaviour possible, and it is why contact-rich tasks are hard on hobby-class arms.
3. In the Kubernetes-controller analogy, what plays the role of the controller manager, and where does the analogy break?
The control loop or policy is the reconciler, pushing observed state toward desired state. It breaks in three places: observation is noisy, partial and inferred rather than an authoritative store; actions take real time and cannot be retried for free; and there is a hard deadline every cycle instead of eventual consistency.
4. Why do robot architectures split into a fast part and a slow part?
The control loop needs a decision every few milliseconds, but good decisions such as understanding a scene or planning a sequence take hundreds of milliseconds or more. So a slow deliberate layer sets goals a few times per second, while a fast reactive layer runs at 50 to 1000 Hz to execute them. You will meet this again as “System 2 and System 1” in the foundation-model lesson.
5. A robot knows its joint angles to a fraction of a degree but cannot reliably tell a mug from a cup. Which kind of sensing is failing, and why is that the norm?
Exteroception, the sense of the outside world. Proprioception is easy because the robot is directly instrumented: an encoder measures the joint it is bolted to. Exteroception is hard because the world is not instrumented for you. A camera gives you pixels, and everything you want to know - what that object is, where exactly it sits, whether it will slip - has to be inferred from them.
Do this
No code yet. Two things, ten minutes total.
-
Watch any teleoperation demo video, searching for “SO-101 teleoperation” or “ALOHA robot”. While it plays, narrate the loop out loud: what is being sensed, what is deciding, what is acting. It feels silly. It builds the reflex of seeing every robot as the loop, which is the single most useful habit in this course.
-
Reach for something without looking again, and this time notice the moment you correct. That correction is feedback control, and you will implement it yourself in Module 1.
What you can now do
You can describe any robot as a sense-decide-act loop, name the parts of an arm, explain why six degrees of freedom is the magic number, and say why the physical world makes this loop harder than any reconciliation loop in software.