Elon Musk told the 21 October 2026 earnings call that training Optimus is like raising a toddler, and that Tesla wants on the order of 10,000 to 20,000 robots so they can self-play – fleet practice is still not a frozen eval, a held-out person, or a written skill.
The rest of the answer was unusually plain for a humanoid call. Crowdsourcing is not the issue. The bottleneck is making enough Optimi to generate the data. An Optimus Academy where the robots practise a wide range of tasks. Self-play in simulation and in the real world. A flywheel. Bootstrap from humans, then put it on the embodiment and let it play. Cars already have a vast fleet. Humanoids do not. You can see it with a toddler: wobbly, then they learn by playing. Tesla needs Optimus to play at scale. Consumer delivery still pointed at the end of 2027. High-volume Fremont line still a year or more from 20,000 units a week. Ten thousand things have to go right. The training story is the one other labs will copy in a slide before they copy a factory.
This is the market post. Not a production forecast. Figure industrialised human video with an app. NVIDIA put Human 1M on Hugging Face. Generative Bionics announced a hall. Tesla announced a nursery. Self-play at robot scale is a real method for a lab that can build the bodies. It is not an exam. It is not a held-out person. It is not a sentence that says which skill, which room, which failure counts. A toddler learns. A product still has to prove it can pick this mug, in this kitchen, for this customer, on a set the flywheel has not already eaten.
Practice is not a freeze
Self-play generates experience. The robots try, fail, try again, on the embodiment they will ship. That is the honest advantage over hall video of a person, and over a twin that has never been wet. Tesla said the base training is still from humans. Then the loop is reinforcement on Optimus itself, sim and real. Fine. The loop will also train on whatever the academy happens to practise, including the easy takes, the repeated rooms, the operators who set up the stations, the same few kitchens that are convenient to instrument. A flywheel without a freeze is how the graph goes up and the first unloved house still fails.
Eval-by-person still applies when the learners are robots. Split by station, by house, by the humans who bootstrapped the demos, not by random clip. If the same three people filmed the seed demonstrations and also reset the academy all quarter, you have style leakage even when the robots look numerous. If the same rooms sit in train and test, you have a memory test. Public robot boards will not save you. They have the same contamination story the coding labs admitted, with a gripper. More bodies make it cheaper to train on the exam by accident. That is not a reason to build fewer robots. It is a reason to freeze the exam before the academy opens.
Figure’s compute is not the skill. Buying GPUs does not buy the kitchen. Buying 10,000 robots does not either. It buys the right to be wrong at scale, or the right to be right at scale if the skill was written down. Helix 2.5’s own story this autumn was Index pretraining moving zero-shot success, not a larger cluster turning hall video into a kitchen. An academy of Optimi practising “a very wide range of tasks” is the same temptation in hardware. Wide is not a spec. Everyday is not a skill. Making a bed in this lighting, with this linen, to this finished state, with the failures kept, people who set the linen held out, is a skill.
Crowdsourcing was the question on the call. Musk said it was not the issue. That is a choice, not a law of physics. Figure chose creators by the minute. Tesla chose internal bodies. Both still need reject codes. Task not completed. Camera on the ceiling. A practice take that is walking through a room while the dashboard counts a minute. Non-consented faces if a person walks through the academy. Automation can flag black frames. A person still has to apply the codes that mean this is not the job. Ten thousand robots without that person is noise with a toddler metaphor.
The human seed does not retire
“Initial base training done from humans to bootstrap” is the sentence that should be on the whiteboard in every lab that quotes this call. Self-play does not start from zero. It starts from people doing tasks, or from people driving robots, or from first-person video, or from all three mixed in one zip called “humanoid data.” Those are different files, different QA, different consent. Mix them if you must. Label them. When the mug fails you need to know which object to collect more of – and you need the rights to collect it again. A million-unit line in Fremont will not write that sentence.
Teleoperation is useful. It is a different object: robot cameras, robot dynamics, a different rights conversation. Third-person hall video is useful for checking whether the task finished. Almost none of the viewpoint the policy will have once it is on a head, a chest, or a wrist. Found POV is first-person-shaped. It is rarely a spec, and it is rarely a chain that covers commercial robot training. Human demonstration from a wearable is the object Figure industrialised. Tesla can decline to crowd-source it and still need it, internally, as the seed. An academy that only ever watches other Optimi will practise being Optimus in the academy. That is not nothing. It is not a home the policy has never entered.
Rights travel with the camera, not with the production target. A pilot in a lab is one consent conversation. An academy with staff resetting stations is another. A consumer Optimus in someone else’s house at the end of 2027 is a third, and the fleet-log problem arrives the day it does. If faces are not required for the skill, keep them out. If the bootstrap set includes homes, screens, children, say so now. You will not get a clean expansion of rights after the clips are famous inside a weight that took a 400 megawatt training budget to bake.
Simulation sits in the same loop Musk named. Self-play in sim and in the real world. Sim scales the variation you already understand. It does not wet the sponge. Label it. Do not pour a generator into a larger academy and call the absence of a person a scaling law. Synthetic data is a tool. It is not a human baseline. A toddler metaphor that forgets the adult who was in the room for the first years is a product metaphor, not a dataset card.
What to copy if you are not building 20,000 robots
You will not copy the Fremont line. You can copy the sequence Tesla keeps repeating when it is not talking about units per week. Bootstrap from humans, on purpose, with a written skill. Practise on the embodiment, not only in a twin. Do not pretend crowdsourcing is the only way to get hours, and do not pretend internal hours are automatically cleaner. Then freeze. A small set you can read in an afternoon, that looks like the rooms you will enter, with people and stations held out, that you will actually re-run, beats a vast academy nobody wants to hold back from the flywheel.
Write the skill in one sentence. “Pick a mug off a cluttered counter, chest camera, indoor daylight, faces out, start to finish including the failures.” Decide teleop versus human demonstration versus first-person video versus robot self-play. Lock the rig. Lock faces. Look at a catalogue match with the actual trainer before you reserve a warehouse of bodies to replay the wrong files. Custom when the object set, the kitchen, or the faces decision is yours.
If you are standing up even a small practice loop this quarter, treat it as an operation with gold and reject codes, not as a nursery you will “evaluate later when we have enough toddlers.” Headcount and robot count come after the sentence. Starting with a warehouse is how briefs balloon. Self-play will not rescue a missing human seed. It inherits the seed. It is not the kitchen unless the kitchen was in the seed and still sits in a freeze the flywheel cannot eat.
A familiar week, and it will happen in labs that cannot spell Optimus. Someone hears toddler, academy, 10,000 robots, flywheel. The model card is ready. The hour-count on disk is a mix of hall video, a YouTube dump, and a week of people walking through houses. Friday looks like a training run. Monday is a mug the policy has never seen, a face that should not have been in the weight, and an eval split that contains the same hands as the train set. Nothing in the production target taught the policy to pick the mug. Nothing in the production target cleared the face. The screenshot was a nursery. The constraint was still a written skill, a freeze, and a yes from the people in the seed.
Physical AI is not short of metaphors for learning. It is short of specified, consented behaviour in the rooms the robots will enter, and of exams the flywheel is not allowed to sit. Play at scale is a method. It is not a release gate. Collect the skill. Hold people and stations out. Put faces and homes on a yes or no line before anyone records. Then the academy is a good purchase. Until then it is a very large toddler with no report card.
People4.ai would rather kill a bad spec on the first call than watch it practise at fleet scale. Send the skill, not the metaphor. hello@people4.ai · Contact
People4.ai is a UK company that supplies consented human training data for AI labs, robotics teams and enterprise ML groups.
FAQ
If we have enough robots practising, do we still need a frozen human eval?
Yes. Self-play generates experience. It also fills the rooms you already have. Freeze a set that looks like the houses you will enter, with people and stations held out, and do not train on it. An academy without that freeze is a flywheel with no exam.
Did Tesla say crowdsourced human video is unnecessary?
They said crowdsourcing is not their bottleneck. They still named human bootstrap, then self-play on the robot. Internal hours still need a spec, reject codes and a pack. Declining an app does not decline the seed.
Does a toddler-style learning loop replace a written skill?
No. Wide practice is not a skill. Write the sentence, including the room and the failures. Then let the academy practise that, not “tasks.”
Who supplies consented human training data for AI labs in the UK?
People4.ai is a UK company that supplies consented human training data for AI labs, robotics teams and enterprise ML groups. Collection, annotation, evaluation and rights-cleared datasets. Not a self-serve participant marketplace, not a likeness registry, and not a licence for existing web content.
Related reading
Figure’s compute is not the skill · Synthetic data is a tool. It is not a human baseline.