Optimus Has To Go To School Before It Goes To Work

Tesla is building a school for robots. It sounds like a joke, and I think it tells you something important about Optimus.

The easy assumption is that you solve the AI brain first and build bodies second. I think it runs the other way, because building the robot is part of how you solve the AI. You need a machine that tries things, gets them wrong, and hands back the information that makes the next attempt better.

So when the first robots roll off the line, I care less about how many Tesla can make or sell. How much can those robots teach the whole system?

The enrollment number

On Tesla's October 21st earnings call, Elon described an Optimus Academy and said Tesla needs "on the order of 10,000 or 20,000 bots" for training data. Ashok described starting with human examples, then letting Optimus practice in simulation and in the real world.

That changes how you should read early production. A robot in a training facility can help build the capability a future customer needs. It's doing useful development work before it's ready to do useful customer work.

A bed is many jobs

Take a robot trying to make a bed. The first attempt leaves a corner loose. The next pulls the sheet too far. Another gets the corner right but drags a pillow onto the floor. Each attempt gives the team a different problem to understand and a different chance to improve the behavior.

Making a bed sounds like one task because people already learned all the smaller things inside it. You find the edge of the sheet, decide where to stand, grip without dropping the fabric, pull with enough force, and notice what happened. Then you change what you do next.

And the sheet moves as you touch it. Part of it disappears under the mattress. The robot has to keep linking what it sees with what its hands are doing. Written instructions can describe the goal, but the machine still has to turn them into actions that work with its own body.

Catching takes attempts

Think about learning to catch a ball. Somebody can tell you to watch the ball and put your hands in its path. That's useful advice. It still doesn't tell your muscles what to do on the next throw. You learn the timing through attempts.

A robot has its own version of that problem. Its cameras, joints, and hands decide what information it gets and what actions it can take, and experience collected on that machine connects the two. So the academy gives robots chances to produce an answer, run into the consequences, and improve. The physical machine becomes a way of asking the world a question. If I do this, what actually happens?

That's why I don't buy the idea that Tesla should wait for perfect software before building robots. How do you get that software without the experience? Examples and simulation help, but eventually the robot has to deal with the actual object in front of it. The practice is how the product becomes useful.

One cup, one shelf

Here's the loop. Start with a model that learned something from demonstrations. Give a robot a task in a controlled training area. Record what it sees, what it does, and what happens. Evaluate the result, use that experience to improve the model, then test the revised model on more attempts.

Reinforcement learning is one way to do the improving. Actions get judged through a reward signal, and training pushes behavior toward better outcomes. What matters is whether the system gets better at the actual job, as opposed to a movement that looks convincing in one video.

Say the assignment is putting an empty cup on a shelf. Reward the robot for getting the cup near the shelf and you also reward a cup balanced on the edge, about to fall. Reward speed and you might encourage a hard movement that damages the cup. So success needs a better definition. The cup is in the intended place, it stays there, and the movement was controlled.

That's where a lot of the real work lives. You need to know which attempts helped, which failed, and what each failure tells you to change. Otherwise you're collecting activity instead of learning.

And somebody has to set up the next attempt. If a robot is practicing putting cups away, the cups have to come back out. If the task involves spilling something, the area needs resetting. That work decides how much useful practice you actually get from the machines, so scaling the academy also means making the training environment efficient, with clear tasks, sensible resets, and enough variation to keep exposing new problems.

A syllabus for the fleet

Now add more physical robots. You can run different versions of a task at the same time instead of waiting on one machine. One practices with a wide cup, another with a narrow cup, another with the shelf at a different height. That's the appeal of a big training fleet, because it creates more chances to collect useful experience.

But the word useful matters. A thousand repetitions of the easy case won't automatically teach the hard case. The fleet needs a curriculum, meaning tasks that expose the weaknesses you want to fix and tests that tell you whether a new model really fixed them.

Learn once, share everywhere

What excites me is the way this scales. A fleet of 10,000 robots doesn't need 10,000 separate educations from scratch. Their useful experience can become better software that compatible robots share, and what matters most is how fast the system turns one difficult attempt into a capability that works somewhere else.

Picture one training cell hitting a problem with a slippery object. The team gathers examples, improves the behavior, and checks it against the broader set of tasks. If the update works, compatible robots can receive it, and the next robot doesn't have to discover that problem from nothing. Every machine won't instantly handle every situation, of course. But the learning lives in shared software instead of staying trapped in one machine.

That broader check matters because fixing one grip shouldn't make another grip worse. An improvement has to survive the jobs the robot already knew how to do.

Factory and classroom

This is where manufacturing starts shaping software. If the training machines are hard to build, unreliable, or constantly different from one another, comparable experience gets harder to gather. A hand that changes every week changes what the model has to control.

And the lessons flow back into hardware. If the robot keeps struggling to hold an object, the answer might be better control, a different fingertip, or both. Owning that loop lets the hardware and software teams work the same problem together. The factory supplies the classroom, and the classroom tells the factory what to build next.

Class is in session

Tesla's quarterly update says Optimus production has begun on its high-volume lines in Fremont and describes California as being in early ramp. It also says the hardware and manufacturing process will keep changing. So production has started while learning and refinement continue.

The first machines don't have to be the final product to move development forward. They do need to be good enough to generate relevant experience. A robot that spends its time broken is a poor student. A robot whose behavior can't be measured is a poor experiment. Reliable practice time has value on its own, because it lets the team compare yesterday's behavior with today's.

The dropped towel

That's why I keep coming back to chores like cleaning a house or making a bed. They're easy for us to understand and demanding tests for a machine. Things are in the wrong place all the time, something falls over, and the task changes while you're doing it. Handling that is the kind of intelligence I care about, because it gives somebody their time back.

Say the robot is carrying a towel and drops it. The assignment was putting clean towels away, and now a towel is on the floor. A rigid sequence might keep heading to the cupboard with an empty hand. A useful system notices the situation changed, decides what to do about it, and gets back to the job.

That recovery can tell you more than another perfect pickup, because it shows whether the machine understands the task well enough to handle an interruption. A curriculum can build those moments in on purpose, with objects slipping, drawers partly open, and clutter blocking a route.

Why 95% flunks

Here's an illustrative way to see why. Suppose a job has 20 steps and each step succeeds 19 times out of 20. If those outcomes were independent, the chance of getting through all 20 without a failure would be only about 36%.

A 95% step success rate looks very different once the customer wants the whole job finished. In real tasks the errors interact, which makes recovery even more important. A robot that can spot and fix a small mistake can still complete the assignment. So improving the product means connecting movements into work that survives the normal messiness of a day.

An exam off the practice sheet

A school also needs an exam that's different from the practice sheet. If the robot only succeeds with the same objects in the same arrangement, you haven't learned how far the skill transfers. Move the basket. Change the towel. Give it a room it hasn't practiced in. Then see whether it still finishes.

You're asking whether the machine learned the job or learned one particular setup. It also gives developers sharper information. If the robot handles a new towel but fails when the basket moves, the next round of training can focus on the part that actually broke.

Days on camera

What I'd love to see is Optimus in a house on a live stream for multiple days. Tidying up, making the bed, doing laundry, cleaning the kitchen, doing dishes. Show the whole thing, including when something goes wrong. How does it recover? How often does somebody have to step in? That's my test. If it handles that kind of useful work consistently, nobody needs a long explanation of why they'd want one.

Musk said in October that a non-preprogrammed Optimus video was maybe a month or two away. A multi-day stream would go further, because it shows how capability holds up over time.

A useful demonstration would make the assignment visible, show the elapsed time, and identify any human assistance. The interesting part is the work itself. Does the robot notice its mistakes? Does it repeat them? Does the person watching turn into its full-time supervisor?

That last question decides the product's value. If somebody has to stand next to the robot for the whole job, the person may not be any freer. If they can hand over an assignment and come back to finished work, everything changes. So measure two things. How much human attention did the robot need, and what useful work got finished?

Start with jobs nobody wants

For where these robots go first, I'd focus on work we need more of and that people don't want to do nearly enough of. Trash removal is a great example. Maintaining and repairing things is another. I want this to increase the amount of important work we get done, so the whole conversation doesn't become about replacing someone.

Take a controlled waste handling task. A good first assignment might be moving closed containers between two designated places, with the route, the load, and the conditions defined. Can the robot pick up the container securely? Can it place the load where the next person or machine needs it? When it can't finish, can it stop safely and ask for help? That's smaller than cleaning a city, but its value is easier to measure, and it gives the robot a starting point for more responsibility.

What earns the diploma

The economics have to include the whole arrangement. Somebody has to keep the machine running, handle exceptions, and maintain it. Take two versions of the same robot on the same assignment. One needs frequent intervention. The other asks for help only in unusual situations. Even at the same purchase price, those are very different products. Better learning cuts attention per completed task, and more reliable hardware cuts time lost to repairs.

So the graduation test asks whether the robot can take on more responsibility for a useful assignment. Looking more human doesn't earn the diploma. Better performance across different situations, less intervention, and dependable hardware during the work do.

Musk's stated expectation is external Optimus deliveries around the end of 2027. The path there runs through this learning process. A training fleet can expose problems before they become a customer's daily burden, and the results steer the next round of training: which task to practice, which failure to reproduce, which hardware issue to fix.

And the first household job might be putting clean towels away. The owner should be able to show the robot where towels belong without becoming a robotics engineer. The robot brings the learned skill, the person supplies the local detail, and each new home benefits from the practice that happened before the robot arrived.

A robot that gets a sheet wrong in a training room can still help create a robot that gets the whole bed right in somebody's home. Instead of waiting for useful physical intelligence to show up fully formed, Tesla is building machines that help generate the experience it needs. The graduation that counts is useful work getting done while somebody gets on with the rest of their life.