The first thing I learned on Tesla Autopilot was that the car got better every week, and nobody touched the hardware. It had the same cameras and the same computer it had the week before. What changed was the data.
Recordings came back from the fleet, the strange moments got labeled, the model retrained, and the next version drove a little better than the last. I was the third software engineer on Tesla's Autopilot AI team and the first to work on fleet learning, the systems that kept that loop running.
Here is what I took from it: the model was never the asset. Every model we shipped was replaced by a better one. The asset was the loop that kept producing them.
A fleet is what makes that loop worth so much. When a person learns a skill, it stays in one head. When a fleet learns one, every machine has it by the next update.
Almost everything about a robot has a price
I think robotics is heading to the same place. The loop will be the asset, because the other parts of a robot are getting easy to buy.
Capable robot models are now published with their weights (the learned values that make up the model), free to download. The body has a list price too: you can find a humanoid listed from about $5,000. Compute rents by the hour.
Those are list prices and license terms, not what a deployment costs. The cheapest humanoid is not the version you can develop on. Not every open model allows commercial use. I wrote about how to review what is open before you build on it. Still, all three are on the open market, which means your competitors can get them too.
What nobody sells is what a robot has learned in your building, doing your task. You can't rent that experience, and it decides whether a capable robot becomes a reliable one.
So my bet is this. What separates one robot company from another will be what its machines have lived through, and who owns the loop that turns that experience into the next release. Some companies will own that loop. The rest will rent their intelligence from someone who does, and improve on someone else's schedule.
This argument favors the work Avala does, so weigh it accordingly. Further down, I name three developments that would prove it wrong.
There is no internet of physical work
Language models had an internet of text to learn from. Robots have nothing like it. Bessemer's estimate, which we cite in our buyer's guide to robot data, puts all the robot manipulation data in the world at about 300,000 hours, against roughly 1 billion hours of internet video. That is an estimate, not a census. A robot produces at most 24 hours of experience a day, and it has to show up to produce any of it.
That changes which input is scarce. In language, data was abundant, so the contest moved to the weights and then to the compute behind them. For language, I think that reading is right. In the physical world, weights are getting easier to obtain, and compute is expensive but for sale. Your fleet's experience is neither.
Tenants and owners
I think of the two kinds of company as tenants and owners.
A tenant runs someone else's model, sends its recordings upstream, and files a ticket when the robot fails. If those recordings make the model better, the tenant rents the improvement back, and so do its competitors.
An owner holds four things: the recordings, the judgments made about them, the tests a model has to pass, and the history that ties all three to every release. The judgments are the part no fleet produces by itself. A log shows what the robot did, and someone still has to say what doing it right looks like. Holding all four is what lets you fix the failure you find at 2 a.m. instead of filing a ticket about it.
None of this means building everything yourself. Rent the compute, even if it gets more expensive. Rent the model that turns a sentence into a plan. Start from an open base model (a general model you adapt to your own task) where the license allows it. Own the loop, and rent the rest.
There is a harder reason than speed. A robot acts in the physical world on someone's behalf, and when it breaks something, nobody blames the robot. They blame the name on its side. You can't answer for a machine whose learning you can't trace.
The law is moving the same way. The EU's revised Product Liability Directive applies to products placed on the market after December 9, 2026, and it covers software, including AI systems. Under the directive's evidence rules, a court can presume a product defective when a defendant fails to disclose evidence it was ordered to produce. Whether and how that applies to a particular system is a question for counsel, not for me.
So the tenant makes a poor trade: it answers for the machine and gives away what the machine learns.
What the CoRL 2026 papers showed
The Conference on Robot Learning (CoRL) ended in Austin on November 12, 2026. In my reading, its strongest papers were about data, not architecture, and we covered them in our notes on the CoRL 2026 papers. Three of their numbers, read together, look like a ladder. Each rung needs less data than the one above it and sits closer to a single machine.
The three figures come from different systems and are not a scaling law. Read the shape, not the ratios.
- At the top: 20,854 hours. EgoScale (NVIDIA, UC Berkeley, and the University of Maryland) pretrains on human video converted into robot actions. That gives a model its prior, the general knowledge it starts from, and I expect priors like this to be bought, shared, or downloaded.
- One rung down: about 54 hours. The same paper uses roughly 50 hours of matched human data and 4 hours of robot data to connect that prior to one robot hand.
- At the bottom: 30 minutes. DreamZero (NVIDIA) reports adapting to a new robot body with about 30 minutes of what the authors call play data. That is one system and one transfer setting, not proof that 30 minutes is generally enough.
The top of the ladder is becoming a commodity, and the bottom is shrinking. It has not reached zero. What is left is specific to your robot, in your building, doing your task, and nobody can rent you that.
Two more results point the same way. π0.7 (Physical Intelligence) trains on mixed-quality data, failures included, by attaching to each attempt a record of how it went. ATHENA (Shanghai Innovation Institute and collaborators) scores each real-robot demonstration for how much it helped, and reports that two thirds of them were enough to match training on the full set. My reading of both: knowing what your experience is worth beats having more of it.
Three things would prove me wrong
Here is what would change my mind.
- Simulation closes the gap for hands. Simulation already carries much of legged locomotion, and GRAIL (NVIDIA and UCLA) trained humanoid policies, the models that control the robot, on generated data alone. Soft and slippery things are still hard: cloth, cables, anything wet.
- Pooling beats owning. A provider's model trained on 100 fleets outperforms what any owner gets from its own fleet, even on the owner's own task. That already holds for general priors, at the top of the ladder. I have not seen it shown at the bottom.
- Experience transfers across bodies. If someone else's fleet can teach your robot, your own matters less. The 30-minute result points that way.
What still holds if I am wrong
Suppose all three happen, and teaching a robot gets cheap. Proving that it works does not get cheap with it. That is my forecast, and it is a judgment, not a measurement: by 2030, I expect teaching a robot a new task to cost far less than proving it can do that task reliably in production.
The reason is physical. Software can be tested in 1,000 copies for almost nothing. A physical test takes a robot, a floor, and real time. A model can arrive already trained and still have to prove itself on your packaging, under your lighting, at your cycle time.
A benchmark score does not settle it. The LIBERO-Plus authors report models that scored about 95% on a standard benchmark falling below 30% under modest changes, including camera viewpoint and the robot's starting position. That was in simulation, so it shows how fragile a benchmark score can be, not how any model performs in a real deployment.
A test also outlives the model it was written for. Turn a failure into a test, and it checks every model you adopt after that, whoever built it.
Automation speeds the loop up without removing it. ENPIRE (NVIDIA, Carnegie Mellon University, and UC Berkeley) had coding agents run the whole cycle on real robots: reset, execute, verify, refine. Every round still depended on a check of whether the attempt worked.
Five questions that tell you which one you are
These are the questions I would ask about any fleet.
- Can you say what changed? For any release, an owner can list the recordings, labels, tests, and decisions that separate it from the last one. A tenant gets a version number.
- Do you keep your failures? I mean the interventions, the near misses, and the recoveries, each with a record of what happened. Failures mark where a model's knowledge ends. If your provider sees a failure and you don't, it teaches their model before it teaches yours.
- How fast does a lesson travel? When one machine fails on Tuesday, how long is it until every machine has learned from it? On Autopilot the loop turned every week. If yours turns on someone else's release schedule, you are a tenant.
- Can you say no? An owner can reject data that is not good enough and hold back a release that fails its own tests. A tenant ships what arrives.
- Can you leave? If you changed base models tomorrow, would your experience and your tests come with you? An owner keeps them in open formats and takes them along. A tenant starts over.
Ownership gets decided early
Until now the question was whether a robot could do the task once, on video. The next question is whether it does the task better in March than it did in January, in your building, and who owns the reason why.
I hold the timing loosely. Robots are arriving more slowly than the headlines suggest, because atoms are slow. Ownership does not wait for them. Data rights are written into the first contract, long before the fleet is large: who may record, who may train on the recordings, and whether they can be pooled with a competitor's.
There is a precedent for settling ownership on purpose. In 1862 the asset that mattered was land, and the Homestead Act transferred 270 million acres, a tenth of the country, into private ownership through more than 1.6 million homestead applications. In 1944 it was education and a home, and the GI Bill did it again. Both were unjust in who they left out. Both worked for the people they reached, and for the same reason: they made owners. I made the longer version of this argument in The Fourth Founding.
This era's asset is the AI stack, and I expect most of it to concentrate, because chips and data centers cost what they cost. Experience is the exception. It is produced wherever a machine works, in every warehouse, on every farm, and on every road, so it starts out distributed. It only concentrates if the people producing it hand it over.
The same goes for the people who teach the machines. Their judgment is capital, and they should own a share of what it builds.
Why I started Avala
A demo is not a moat, and neither is a dataset or a set of weights. Each one is a snapshot. What compounds is the experience of doing your work, the tests that prove it, and the loop that connects them.
That is why I started Avala. Tesla built that loop for its own fleet. Avala is Fleet Learning Infrastructure for Physical AI: we build the loop as infrastructure, so a small robotics team does not have to become a data company first. The customer owns the model. We own the infrastructure that makes it improve. A customer's recordings, labels, evaluation results, and models stay inside its own program, on open formats it can leave with, never pooled with another company's and never rented back.
Those are commitments about ownership, not a list of features. What a given program covers is set in its scope.
I want every fleet to learn from its own experience, and whoever owns that experience to own what it teaches. The machines are going to learn either way. The question is who they are learning for.
Building robots? Bring us a production workflow worth improving.
Emal Alwis is the founder and CEO of Avala.



