1.1 The agent loop: What does an agent do?

A software agent runs a loop.

perception → planning → control → execution

It reads its context. That is perception. A model decides the next step. That is planning. A harness dispatches the step and checks what came back. That is control. The step itself is a function call. That is execution.

An agent is a loop: perceive, plan, control, execute.

1.2 Execution: Why isn't an action a function call?

A function call is the cleanest kind of action there is. It takes typed arguments. It does the same thing every time. It returns a result or raises an error, so the agent knows exactly what state the world is in afterwards.

A motor command has none of these properties.

command → torque → motion

The motion depends on the load, the friction, how warm the motor is and whatever the hand is touching. The same command produces a slightly different motion each time. And it returns nothing. The robot doesn't get told what happened. It has to look.

In software, execution is a function call. In robotics, execution is a force applied to the world, and the world doesn't return a value.

1.3 State estimation: What is the extra step?

Because execution doesn't return a value, a robot needs a step that a software agent doesn't have. Between perception and planning it has to work out what is probably true: where its own arm ended up, whether the cup moved, whether the grasp held. The sensors are noisy and the last action didn't land exactly where it was commanded, so the robot keeps a belief and updates it. That step is state estimation.

perception → state estimation → planning → control → execution

Control changes too. In software it is dispatch. In a robot it is a feedback loop that measures error = desired − actual and corrects it hundreds of times a second.

State estimation exists because physical actions don't report back.

2.1 Body: Who makes the parts that act and sense?

The body industry is the old one. Industrial robotics has been building arms for fifty years. Factories install more than half a million industrial robots a year, over half of them in China, and more than four million are running. Four companies define the category: FANUC and Yaskawa in Japan, ABB in Switzerland, KUKA in Germany. Around them is a parts industry that makes motors, reducers, screws, bearings, encoders and sensors.

What is new is the shape of the body. Legs and hands are being added to arms. Tactile skin is being added to cameras. The suppliers are mostly the same kind of company, and often the same companies.

Until now the body industry also supplied the behaviour, and it was written by hand. An engineer programs the motion, the arm is bolted inside a cage, and the parts arrive in the same pose every time. Nothing is learned, so nothing generalises. Change the task and you rewrite the program. That business works, and it has stopped growing fast. Teradyne's robotics unit, which owns Universal Robots, the best-known maker of collaborative arms, had revenue of $326 million in 2022 and $308 million in 2025.

The body industry is industrial robotics: arms, legs, hands and the sensors on them. Its robots run a program where a brain should be.

2.2 Bill of materials: Where does the cost go?

Morgan Stanley estimated what it costs to build Tesla's Optimus Gen 2, using quotes from individual suppliers. The answer was $50,000 to $60,000 a unit, before software. The estimate dates from mid-2024, and the report expects it to fall a long way at scale.

The robot has 28 actuators, the motor-and-gear assemblies that drive the joints.

actuators ≈ 95% · compute, cameras, chips ≈ 4% · battery and structure < 1%

The hands alone are 17% of the cost, more than any other body part.

This inverts what I'm used to. In an LLM product, compute is most of the cost. In a humanoid, compute is 4% and the joints are nearly everything else. The catch is that the 4% is hardware. Training the brain doesn't appear on a bill of materials at all.

By cost, a humanoid is almost all body.

2.3 Brain: What is physical AI?

Physical AI means one specific thing: using the transformer models that came out of language to control a robot's actions directly.

LLM: text → text

VLM: image + text → text

VLA: image + text → action

A vision-language-action model takes camera frames and an instruction and outputs motor commands. There is no hand-written motion in between. It replaces the program an industrial robot runs on, and that is why one machine could, in principle, do many tasks.

Three kinds of company build these models. Frontier labs and Nvidia: Google DeepMind has Gemini Robotics, and Nvidia has GR00T. Startups that make only the model: Physical Intelligence and Skild. Robot makers that build their own: Figure with Helix, Tesla, 1X.

The funding is ahead of the volume. Figure was valued at $39 billion in 2025. Skild raised money at more than $14 billion in January 2026. Physical Intelligence raised at $5.6 billion in November 2025.

The body most of this money has picked is the humanoid, and the reason is practical. Doors, stairs, tools and workstations were designed around a human body, so a human-shaped machine fits the world as it is. In China, by one count, more than 150 companies are building one.

Physical AI is a transformer that outputs actions instead of text.

2.4 The layer in between: What makes the motion precise?

Between the model and the motors sits the layer in between. The brain decides what to do. This layer makes the body do it precisely: the hand to the right place, at the right speed, with the right force. That is control, run as a feedback loop on every joint. Most of the companies that sell it describe their product as an operating system.

Traditionally this layer was many small models. Each part of the body had its own controller: one for the arm's trajectory, one for the legs' balance and gait, one for the hand's grasp, each written or trained for that specific hardware. Change the hand and you rewrite its controller. Many humanoids still run this way underneath. A slow model plans a few times a second, a fast policy outputs motions hundreds of times a second, and joint controllers run thousands of times a second.

The claim of a robot OS is that this should be one interface. If the model is capable enough, it should be able to orchestrate any set of actuators, the way one operating system runs on many kinds of hardware.

model output → the layer in between → actuator commands

In agent terms this is the function-calling layer of robotics, and the difficulty is the one from section 1. A function has a signature. A body doesn't have a clean one. Every robot has a different number of joints, different dynamics and different limits, and the call doesn't return. It is hardest in manipulation, where the hand is in contact with something and precision matters most. Section 5.1 comes back to that.

The existing glue is ROS, the open-source middleware whose name stands for Robot Operating System. Nvidia sells a full stack around its Isaac simulator and GR00T. Skild and Physical Intelligence pitch one model for any body. Smaller companies use the word directly, like RoboParty with Party OS.

It is also the layer where it is hardest to get paid. Every robot needs it, so robot makers bundle it and model companies absorb it. On a list of 160 robotics companies I sorted, one sold orchestration as its main product.

The layer in between is the function-calling interface of robotics. It takes care of precision, and its promise is one model for any body.

2.5 Other maps: How do analysts cut it?

The simplest map comes from Morgan Stanley. In February 2025 it published the Humanoid 100, a list of a hundred public companies tied to humanoid robots, sorted three ways.

The counts add to 108 because a few companies sit in two buckets. There is no bucket for the layer in between. And the label is noisy: only 52% of the hundred were reported to be involved in humanoids at all. The rest were on the list because a competitor was, or because the analysts saw potential.

The map I find most useful is Robotics Bottleneck Research, which organises the industry by what breaks and splits it into nine systems named after a body.

The last four are the ones a parts list misses. Blood is the training set. Immune is the eval. Neither appears on a bill of materials, and a robot company can't run without either.

3.1 China: Who builds the body?

By Morgan Stanley's count in early 2025, 73% of the companies confirmed to be working on humanoids were in Asia, and 56% were in China. Over the previous five years China had filed 5,688 patents mentioning humanoids. The US had filed 1,483.

The reason is the electric car.

EV supply chain → humanoid supply chain

A humanoid's motors, batteries, power electronics, castings and sensors are a car's, and China already makes all of them. Sanhua and Tuopu, two of Tesla's car suppliers, are now among the best-known actuator suppliers.

The older robot business shows the same shift. Chinese brands now supply 57% of the industrial robots installed in China, up from 47%. Leaderdrive, a Chinese maker of harmonic reducers, sold more than 425,000 of them in 2025 while running its factories at about two-thirds of capacity. Harmonic Drive Systems, the Japanese company the part is named after, reported its reducer sales in China down 37%.

Read the capacity number again. The constraint in China is demand. The supply is already there.

China has the supply chain for the body, and more capacity than orders.

3.2 The US: Who builds the brain?

Thirteen of the 22 brain companies on the 2025 list were American. Nvidia holds the position it holds in LLMs: it sells the computer and gives away the software. Jetson Thor is the on-board computer in AgiBot and 1X robots and an option on Unitree's. Isaac is the simulator. GR00T is a foundation model, and it is free.

The best-funded companies that make only the models, Physical Intelligence and Skild, are American. So is Google DeepMind, which works with Apptronik and Boston Dynamics.

High-end on-board compute is the one part of a Chinese humanoid that China doesn't make.

The US has the models, the chips and the capital.

3.3 Japan and Europe: Who holds the precision?

The highest-grade components still come from a short list of old firms. Nabtesco's gears are in more than 60% of the world's industrial robots. The best planetary roller screws come from Europe and Japan. Chinese firms compete well in the low and middle range. The gap at the top is in precision, stability and how much load a part can take.

There is one more level down.

grinder → gear → reducer → actuator

The machines that grind precision gears and screws come from a handful of German and Swiss toolmakers. Whoever wants to make a better reducer has to buy the grinder first. It's the least-discussed position in the industry, and one of the hardest to copy.

Japan and Europe make the parts that are hard to make, and the machines that make them.

3.4 Rare earths: What is underneath?

rare earths → magnets → motors → actuators → robot

Every motor needs permanent magnets, and the magnets need rare earths. China does nearly all of the world's heavy rare-earth separation.

The stock market priced this before most people noticed it. Morgan Stanley checked on its hundred-company list in September 2025, seven months in.

The deepest dependency in the chain is a mineral.

3.5 Contract manufacturers: Who assembles it?

In March 2026 Morgan Stanley added five contract manufacturers to its list. These are the companies that assemble phones and AI servers. The argument was that most robot startups have neither the capital nor the supply chain to build at volume. It's the role cloud providers play for software startups: you rent the factory instead of building it.

Robots may be built the way phones are: designed by one company and assembled by another.

4.1 Shipments: How many robots are there?

By Morgan Stanley's count, about 19,000 humanoids shipped in the first half of 2026. That was up 272%, and 97% of them came from China. By one tracker's count, AgiBot and Unitree shipped about three-quarters of the units.

For scale:

19,000 humanoids in half a year against 500,000+ industrial arms a year

Humanoid shipments are growing fast from a very small base.

4.2 Deployment: Where do they go?

shipped → deployed → working without help → paid for

Each arrow loses most of the units. Most humanoids go to universities, research labs, data-collection sites, exhibitions and government-backed projects. The factory deployments that make the news are pilots.

There is still no demonstrated market for general-purpose humanoid labour at scale. The robots that are in factories are slow: UBTECH has said its industrial humanoid works at roughly a third to half the pace of a person. And the published specifications are marketing numbers. Payload, battery life and degrees of freedom are defined differently by every vendor.

Units shipped is a noisy label. The quantity I'd want to measure is hours of paid work done without a person stepping in.

Shipped is not the same as working.

4.3 Outcomes: What do customers pay for?

Customers buy outcomes. They don't buy robots.

The places where robots earn money today are narrow. In a recent quarter Intuitive Surgical made about two and a half times as much from the instruments and accessories used in each operation as from selling the machines.

installed base → procedures → instruments

The robot is the installed base, and the business is what runs through it. Symbotic automates warehouses. Carbon Robotics sells a machine that kills weeds with lasers. None of these is general purpose, and all of them have paying customers.

So the purchase price is the wrong number to optimise.

cost per task = (robot + upkeep + human interventions) / tasks finished

A $20,000 robot that needs a remote human to rescue it every few minutes can cost more than a $100,000 robot that runs two shifts by itself.

The metric is cost per finished task.

4.4 Forecasts: How big, and when?

The big numbers come from wages.

labour market ≈ 3.4 billion workers × $9,000 ≈ $30 trillion a year

Morgan Stanley's model of the US judged that about 75% of occupations, and 40% of workers, could be substituted by a humanoid to some degree. From that it forecast 63 million humanoids in the US by 2050.

The curve is close to exponential, so the mass is in the tail. The 2030 figure is 0.06% of the 2050 figure. The 2040 figure is 13%. The 2050 number gets quoted. The 2030 number is the one about the next five years.

The big numbers are possible, and almost all of them come after 2035.

5.1 Manipulation: Why are hands the hard part?

Walking is solved well enough. Contact is not. This is where the layer in between is hardest. Contact is where execution is least like a function call and where precision matters most. I wrote in Robotics 101 that robotics gets much harder the moment the robot touches the world, and the industry's structure shows it. Hands are the most expensive body part, and sixteen of the 160 companies on my list do nothing but make them. The tasks nobody has cracked are the ones with contact in them: cables, bags, soft things, tight fits, tools.

The sensors are getting cheap fast. One sell-side estimate has a tactile sensor falling from about $10,000 to about $350. Deutsche Bank has LiDAR falling from about RMB 11,000 to about RMB 1,660 a unit. The sensor is no longer the constraint. Using its signal is.

The bottleneck is manipulation: precise control under contact. Locomotion is mostly done.

5.2 Data: Where does the training set come from?

LLMs trained on text that already existed. Nobody wrote down the torques. Robot data has to be generated, and there are three ways to do it: people operating robots remotely, simulation, or robots doing real work.

You might expect selling robot data to be a good business. So far it isn't, because the data keeps being given away. AgiBot open-sourced more than a million robot trajectories. Nvidia gives away its simulator and its foundation model.

What is scarce is a robot already doing paid work. It produces real data as a by-product, and the customer covers the cost. That's why the factories and warehouses that host robots are more than customers. They are where the training set comes from.

The scarce asset is a robot already doing a real job.

5.3 Evaluation: Who says it works?

A demo is one sample. A product is the whole distribution.

Deutsche Bank's version of the gap is that task success has to go from under 50% to over 99% before a robot pays for itself. Stated as failures, that is 1 in 2 down to 1 in 100, a 50× reduction in errors.

Somebody has to measure that. There are public benchmarks, but there is no neutral evaluator that customers, insurers and regulators all trust. On my list of 160 companies, two worked on validation.

There are benchmarks. There is no evaluator everyone trusts.

5.4 Value capture: Which layer keeps the margin?

Back to the question from the start. It turns on the layer in between. Either one model can drive any body, or it can't.

Nobody knows which it will be. China leads on hardware and scale. The US leads on models, capital and customers. The layer that decides the economics, autonomy that works, is unsettled.

The research map I mentioned makes a pointed observation here. The market pays the most for the brain, which it scores as the layer least likely to be solved soon. It pays the least attention to the layers that already earn money: integration, components and on-board compute.

This is the argument the LLM world had about models versus applications, with one difference. Here the application layer weighs 60 kilograms and has to be manufactured.

The body is getting cheap. Whether the brain can be sold separately is open.