Good Fences Make Great Robots

When manufacturers first solved automation half a century ago, they accomplished it by removing uncertainty from the environment rather than adding intelligence to the machine. Toyota's famous production system mastered this. Tens of billions of dollars of installed cutters, welders, drillers and assemblers are running right now at repeatability measured in hundredths of a millimeter — precise movements, on precise inputs, in tightly controlled settings. It is one of the great engineering achievements of the last century, and almost none of it involves anything we would today call AI.

Put a good fence around a robot's task and you get a great robot. That has been true since 1975.

Artificial intelligence doesn't take away fences for robots. Instead, it lets us move the fence out to cover much more area.

The humanoid hype isn't my primary bet, though I do think it's a big opportunity. The next wave of industrial robotics will not be generalist machines executing any decision task in any environment. It will be purpose-built robots that loosen only the constraints they have to — heavy judgment inside a tightly bounded environment, like the pick-and-pack cell or the operating room, or narrow repetitive tasks in environments that are messier but still limited, like public roads, farm rows, a mine site, a shipping lane.

Loosen some variables, but keep others clamped down. Purpose-built robots are the recipe, and it's where I expect the next five years of returns in physical AI to be made.

Waymo Launched by Doing “Way-Less”

I rode in a self-driving car for the first time in 2013, as a Google employee. A retrofitted Lexus with a spinning bucket on the roof drove the backroads of Mountain View while I sat there deciding whether to be amazed or terrified. I've stayed close to the story since — the original director of business development for what became Waymo is a friend, and I've watched the arc from the inside of the conversation rather than the outside of the headlines.

Here is what people get wrong about that arc. The popular telling is that self-driving took fifteen years because the problem was hard and the models finally got good enough. That's half right at best. What actually happened is that the winner stopped trying to solve the general problem.

Waymo successfully launched by picking a small set of metros with good weather, and by only doing passenger pick-up and drop-off. Human teleoperators stand by for the situations the system was explicitly designed not to handle.

Waymo didn't win by learning to drive everywhere. It won by deciding, in writing, everywhere it would not go.

The industry even built formal vocabulary for the fence. It's called the Operational Design Domain, and it's written into international standards — ISO 34503, BSI PAS 1883. Read the standards and the philosophy is unmistakable: before a regulator will let an autonomous system carry a human being, the builder has to specify, in writing, everything the system will not do. The first question is never "what can it handle." It's "where does it stop."

Failing to build a general robot model isn't a consolation prize. That's the shape of the win.

The Fence Map

Once you see it this way, two questions predict almost everything about whether a robotics company ships or stalls.

How structured is the environment? Is the world bolted down, or does it show up different every time? And how much judgment does the task require? Is it the same motion forever, or does the machine have to decide something?

Cross those two and you get a map. I've come to think it's the most useful single frame in physical AI, because it sorts a decade of winners and casualties without needing to know anything about the underlying model architecture.

Two Kinds of Specialist, Both Getting Paid

Look at the map and the investable claim gets embarrassingly simple: robots make money today in the two quadrants that leave some variables loose and keep the rest clamped down. They are not opposites. They are the same move, executed from opposite directions.

The Fenced Specialist draws its fence in space. The world is nailed down and the inputs are not. A bin-picking cell lives in the same three cubic meters forever, but no two bins are alike. A weld cell sees the same station every time, but the seam wanders. An operating room has the same walls every night and a completely different mess in it. Fix the room, and the model only has to be excellent at what shows up inside it — which is a problem today's models are genuinely good at, and one you can test to exhaustion before you ship.

The Disguised Specialist draws its fence around the task instead. The world is wide open — public roads, farm fields, open ocean, an active mine — but the job has been narrowed until it's almost boring. Waymo does one thing. Zipline flies defined corridors with one payload class. Laser weeders make one visual decision, weed or crop, millions of times an hour. Autonomous haul trucks drive private roads mapped to the inch. Every autonomy story the public files under "AI miracle" sits in this quadrant, and every one of them got there by shrinking the task until the open world stopped mattering.

The Unfenced Generalist refuses both moves. The environment is loose, the task is loose, and in the humanoid case the body is loose too — high degrees of freedom, contact-rich manipulation, unmapped rooms, open-ended instructions. Three fences down at once. The total addressable market is spectacular, which is exactly why the capital shows up. However, for the foreseeable future, generalists will still be limited to fences where they can prove safety and cost-effectiveness against a purpose-built robot. That leaves most of the market to the purpose-built version.

Why the Fence Pays

The strongest argument for narrowness isn't technical. In the physical world, executing a wrong decision is an incident report.

When a chatbot generates a bad citation, you roll your eyes and ask again. When an autonomous machine generates a bad motion plan, it does so with metal arms or spinning wheels posing real physical risk. It closes a gripper on a technician's hand. It sprays herbicide on the crop instead of the weed across forty acres. It performs the wrong surgical procedure. There is no “undo” button on physical AI mistakes. No reputable customer will buy on an impressive benchmark score if there are material risks.

Which is why the real gatekeepers of autonomy are not researchers. They're underwriters, safety officers, and regulators — and all three ask the same question the ODD standards ask: where does it stop? You can price the risk of a bounded system. You can test it to death, log a million reps inside the envelope, and produce an actuarial record. You cannot price "it can do anything." The scope of a general system is precisely the scope of its liability, and nobody will write that policy.

Generality is unfalsifiable. Narrowness is certifiable. Certifiable is what actually gets bolted to a factory floor.

One clarification is worth making here, because it's the part people get lazy about. The claim is not that small models win. It's that purpose-built models win. Sometimes purpose-built is small — a few billion parameters distilled for one grasp, quantized onto a modest edge module. Sometimes it's large, because the task set genuinely demands it, and you eat the cost of the silicon to get it. The model that loses was built for tasks far unlike your tasks, and then pointed at your process and told to figure it out. Fit for the task is key.

Why Not Just Build a Robot That Drives the Car?

Let me steelman the other side, because the humanoid bulls are more right than my map makes them look.

In several years, when humanoid robots clear their technical and safety milestones, they will have a very large market. Human environments are built for human bodies. A machine with our proportions inherits doorways, stairs, tool handles, vehicle cabs, and a few centuries of workflow designed around arms and legs. That generality has genuine economic value anywhere the work is too varied or too low-volume to justify custom automation. I expect that market to be enormous, and I'd like to own a piece of it.

I also expect purpose-built robots to be better, faster and cheaper at 99% of the tasks inside it.

Here's the cleanest way to see it. We have spent fifteen years and something north of a hundred billion dollars building self-driving cars. Why bother? Why not just build one humanoid that sits in the driver's seat of any car and drives it — solving not only driving but every other human task on the same platform?

Ask the question and the answer is immediate.

Better, faster, cheaper — and not narrowly. That's the general law, and it applies far beyond cars: when a task is worth doing at volume, build the task into the machine rather than building a machine that operates the tools we made for hands. Human-shaped tools exist because humans were the only available general-purpose actuator. That constraint is an accident of biology, not a design requirement, and copying it into steel inherits every compromise it forced on us.

Humanoids win the long tail, where variety is high and volume is low. Purpose-built machines win the head of the distribution — which is where the revenue is.

Both of those can be true at once, and confusing them is precisely how capital gets destroyed. The humanoid bet is a bet on the tail. It is not a bet that the head goes away.

What This Looks Like With Real Money Behind It

This isn't a theory I hold at a distance. Several of our best-performing positions at 1Flourish are companies that drew a fence and got very good inside it.

The pattern repeats across the specialized robotics companies we look at. The good ones can tell you in one sentence what their machine does, and in three sentences what it refuses to do. The ones that worry me open with total addressable market and never get to the fence at all.

Two Ways to Own This

If purpose-built wins, there are two shapes of company worth underwriting, and they are close to opposites.

The historical rhyme is semiconductors. Chip designers got the headlines, but the most durable compounding went to the horizontal layers everyone had to buy — EDA software like Synopsys and Cadence, process equipment like ASML and Applied Materials — while a handful of vertically integrated players who owned an entire product category did extraordinarily well on the other side. Both shapes work.

What almost never works is the middle: a company selling a general capability with no deployment of its own and no platform position, hoping one of the other two eventually buys it. That is not a strategy. It's a wish with a burn rate.

The Five-Year Window — and the Part That Compounds

Here's my actual forecast, and I'd like to be held to it. For the next five years, the returns in physical AI go to purpose-built models that do specific things extremely well. Not because general physical intelligence is impossible, but because narrowness is the only configuration that clears the bars that matter between now and 2031: it can be verified, it can be insured, it can run on the hardware that exists, and it can be sold to a buyer with a budget this fiscal year.

Now the part most people miss, and the reason this isn't merely a defensive bet.

First: purpose-built and cheap keeps winning even after the generalists arrive. We already ran this experiment in software. Frontier models kept getting bigger, and simultaneously the fastest-growing part of the ecosystem became open-weight models tuned for a specific job — cheaper to run, easier to control, deployable where you actually need them. A better frontier model did not kill that market; it expanded it, by producing better teachers to distill from. Physical AI will do the same, only more so, because in robotics the deployment constraints are made of watts and milliseconds rather than preferences.

Second, and more important: the data compounds in the specialist's favor. Large language models could scale because the internet already existed — trillions of tokens of human behavior, free for the taking. There is no equivalent corpus for physical manipulation. There is no internet of hands. The training data for general-purpose robotics does not exist yet, and it cannot be scraped. It has to be manufactured, one deployment at a time, by machines doing real work in the real world under real consequences.

So ask who is generating it. Not the humanoid doing a backflip in a promotional video. The picking cell running ten million grasps a year. The weeding rig making a billion classifications a season. The disinfection system that has now read tens of thousands of real operating rooms, including all the ugly ones. Every one of those reps is labeled by physics — it worked or it didn't — which makes it the highest-quality reinforcement signal in the entire field.

The specialists aren't waiting for the general model. They're mining the fuel it will need to exist.

That's the compounding structure, and it's what makes this a growth thesis instead of a hedge. A company that starts with a model built for one task, gets it into production, and accumulates years of real-world reinforcement data holds an asset that appreciates as the field advances. When broader models do arrive, that company doesn't get disrupted by them — it fine-tunes on top of them, from a proprietary data position nobody can replicate without deploying hardware for a decade. Narrow first, then broaden, on the strength of a dataset you own outright.

Don't read this as defeatism about general robots. General robots will eventually come. However, the first wave of winners — and the lion's share of the money — will be made by purpose-built robots that continuously move the fence out.