This week NVIDIA's Jim Fan announced DreamDojo, and it is one of the more impressive things robotics has shipped this year. It is an open-source, interactive world model: you give it a robot's motor controls and it generates the next moment of the physical world directly in pixels. No physics engine, no hand-built meshes, no hand-authored dynamics. It learned how the world moves from a huge pile of human video. Fan calls it "Simulation 2.0" and says 2026 will be remembered as the first year large world models laid real foundations for robotics. He is probably right.
It is worth being clear about what problem this solves, because the problem is enormous. A robot reaching into a kitchen it has never seen has no idea what its next action will do. There is no label that says "if you push here, the cup tips." So the field is building giant generative models that imagine the consequence before the robot commits to it, a model that can dream a plausible future and let the robot plan against the dream. That is the hard version of prediction, and it costs a staggering amount of compute and data to get even partway there.
A sump pump in a basement has the easy version of that exact same problem. And it gives you the answer for free, every few minutes, forever.
The expensive thing the frontier wants is the free thing a machine already does
Strip the prediction problem down to its core and it is this: what is the next physical state, and was my guess about it right? A world model is expensive because the robot never finds out cleanly. The world is open, the outcomes are unlabeled, and the only way to learn is to imagine a future and hope the imagination was close.
A fixed asset closes that loop for you, mechanically, on a schedule. A sump pump cycles. It fills, it kicks on, it pumps down, it shuts off, and it does it again. Each cycle is the machine showing you its actual next state, measured and timestamped, labeled not by a model's guess but by physics. You do not have to dream what a healthy pump does next. The pump performs its next state in front of your sensor a few thousand times a month, and every one of those performances is a fresh, ground-truth, perfectly-labeled example of what normal looks like.
This is the part that gets missed when people assume serious physical AI requires a world model. The world model is what you build when the asset will not tell you what happened. When the asset does tell you, on a fixed cadence, with a clean label attached, you do not need to simulate it. You need to remember it. The hard, generative, dream-the-future machinery is solving a problem your asset solved for you the moment you bolted a sensor to it.
What our sump pump does instead of dreaming
Our sump pump system in a Watertown basement has been running for a long time, and it has never simulated anything. It does not generate a hypothetical future in pixels. It watches the real one. It learned what a healthy cycle looks like by sitting through thousands of actual cycles, each one a free labeled example the pump volunteered, and it built a baseline of the asset's own measured behavior. Then it does one thing: it compares today's cycle to the asset's history, and when today drifts, a cycle running long, a frequency creeping up, a rhythm that no longer matches the pump's own past, it raises a hand and tells a human.
There is no world model in that loop because there does not need to be one. The pump is its own oracle. It announces its next state every few minutes, and our job is just to notice when the announcements stop sounding like the pump we have known for months. That is not a weaker form of physical AI. It is physical AI that recognized it was handed the answer key and decided to use it instead of reprinting the test from scratch.
And it is cheap precisely because of that. We are not paying to learn the dynamics of the physical world. The asset paid for that, in advance, by being a machine that repeats itself. We are only paying to store a baseline and compare against it, which runs local, on small hardware, with no cloud and no per-token bill.
Forty assets, forty free oracles
The objection is always that one pump is too simple, and that a real building needs the heavier machinery. Our 40-device site in Northampton is the test, and it makes the point sharper, not weaker. Forty assets that matter, and every one of them is its own oracle, performing its next state on its own schedule, generating its own stream of labeled cycles every single day. Forty machines, forty baselines, forty self-labeling data sources we did not have to build a simulator to understand.
Add up the cycles across forty assets over the months that site has run and you are looking at an enormous pile of ground-truth examples that cost us exactly nothing to collect, because the building generated them as a side effect of operating. A world-model lab would kill for clean labeled physical data at that scale. A building owner is sitting on it and throwing it away. The entire product is just refusing to throw it away: catch each cycle, learn each asset's normal, and watch for drift.
Where the dream model actually earns its keep
None of this means world models are hype. For a robot in an open, unlabeled world, DreamDojo is exactly the right tool, and Fan is right that this is a foundational year. The lesson is about matching the tool to the problem. A generative world model is what you reach for when the asset is unpredictable, mobile, and silent about its own state. A small drift detector is what you reach for when the asset is fixed, repetitive, and loud about its own state every few minutes.
It even rhymes with the other big story of the week. The editor of Weekly Robotics went to ICRA 2026 and found that robot hands and tactile sensing, the boring business of perceiving the physical world accurately, took something like a third of the show floor. Even at the frontier, the durable value kept landing in the sensing, not the dreaming. For a fixed asset you can skip the dreaming entirely, because the sensing already hands you the answer.
Don't simulate the machine. Listen to it.
A plumber, a property manager, and a small building owner do not need an AI that imagines what their pump might do. They have a pump that demonstrates what it does, on a reliable schedule, for years. The strongest product for that world is the one that treats every cycle as the free labeled example it is: learn the asset's normal from its own measured history, run the detector local, and raise a hand the moment normal turns into trouble.
Let the frontier teach robots to dream a future they were never handed. For the machine in your basement, there is nothing to dream. It already logs its future, one honest cycle at a time. Our job is just to be listening when it changes.
Your equipment already records its own future. We make sure someone is reading it.
Each asset that matters gets a small, dedicated detector trained on its own measured history, learning what a healthy cycle looks like and flagging the moment it drifts. It runs local, with no cloud and no per-token bill, and it never touches the controls. $99 to $199 per month, hardware under $3,000.
See how it worksSources: Jim Fan (@DrJimFan) on X, DreamDojo and DreamZero announcements ("generates the future in pixels," "Simulation 2.0," "first year that Large World Models lay real foundations for robotics"), 2026, and dreamdojo-world.github.io; Weekly Robotics #363, ICRA 2026 field report by Mat Sadowski (tactile sensing and robot hands ~30% of the floor), 2026-06-09; field deployments at The Intersecto Watertown sump pump site and Northampton 40-device building.