The Sim-to-Real Gap Doesn't Exist in Your Building

This week, Cadence and NVIDIA closed a partnership whose entire purpose is to make simulated physics behave more like real physics. The Cadence multiphysics engines, the same ones that aerospace and chip companies have used for decades to model heat, airflow, and electromagnetic interactions, now plug directly into NVIDIA's Isaac robotics libraries and Cosmos world foundation models. The stated goal is to close the "sim-to-real gap" — the persistent problem in robotics where a policy trained in simulation does not behave the way it should once it meets actual metal, actual rubber, actual gravity.

Meanwhile, every commercial building in America has been running its own perfectly faithful simulation of itself, continuously, for decades. We call it "the building."

What the Sim-to-Real Gap Actually Costs

Sim-to-real is the central, expensive problem in modern robotics. To train a humanoid to fold a towel, or pick a part off a conveyor, or balance on uneven ground, you need millions of trials. You cannot run those trials on the real robot. The robot is too expensive, too slow, and breaks. So you run them in simulation.

Then you discover that the simulator did not model the way the gripper deforms under load. Or the way air flows around a moving arm. Or the way a thermal gradient changes a sensor reading. The robot trained for ten thousand simulated hours, walks onto the factory floor, and falls over.

This is what Cadence and NVIDIA's deal is about. Cadence's multiphysics simulators are the high-fidelity ones — they model the actual physics of fluids, heat, and electromagnetism with engineering-grade accuracy. Plugging them into Cosmos and Isaac means the synthetic world a humanoid trains in starts to look more like the real world it will eventually deploy into.

For the robotics industry, this is a serious investment in a serious problem. Worth millions. Worth years. The robot has to learn somewhere, and the only place big enough to teach it is a simulator.

The Building Is Not the Robot

Here is the thing nobody in robotics says out loud about buildings: there is no sim-to-real gap.

I deploy edge AI in a basement with a sump pump. The pump cycled 97 times last night. That is not a synthetic number. The float switch hesitated for 1.4 seconds at 11:47 PM. That is also not synthetic. The temperature in the boiler room dropped 4°F in the last hour. The pressure on gauge three has been creeping up for six days. None of this is in a simulator. None of it needs to be.

The building is already running. It has been running for years. The data it generates is the actual ground truth, generated for free, every second, with absolute fidelity to the physical reality of itself.

The robotics industry's hardest problem is making a fake world good enough to train a policy that will work in a real one. The building practitioner's starting point is the real one. We never have to bridge anything.

This sounds obvious when you write it down. It is so obvious that the robotics industry, which is currently dominating the physical AI conversation, mostly does not notice it. Their problem statement assumes the world has to be simulated, because their hardware platform is too expensive and too brittle to learn in production. Buildings are not like that. Buildings learn in production. Buildings have always learned in production. The pump that cycled 97 times taught us about the float switch in real time, on real metal, while it was actually pumping water.

What This Means for the Reasoning Models

This is not a complaint about robotics. The robotics industry is doing the right work for its problem. And it is doing work that benefits the building practitioner enormously, almost as a side effect.

NVIDIA's Cosmos Reason — the vision-language reasoning model that ships inside Isaac GR00T — crossed one million downloads this week and is now ranked first on the Hugging Face Physical Reasoning Leaderboard. Cosmos Reason is the "deep-thinking brain" of the humanoid stack. Its job is to take physical-world inputs, apply prior knowledge, common sense, and physics, and produce step-by-step plans.

That is exactly the missing layer between "my sensor saw something" and "tell the building owner what to do about it."

The 2-billion-parameter variant runs locally on a Jetson Thor or a desktop GPU. It is open-weight. The same hardware that already runs the sump pump's anomaly detection can host it. A small-building monitoring node does not need the humanoid stack around it to use the reasoning model. It just needs the model.

And here is the asymmetry that compounds: every dollar Cadence and NVIDIA spend closing the sim-to-real gap pays off for buildings, too. The reasoner that is being trained against better-and-better simulated physics is learning physics. When that reasoner runs against a building's actual sensor stream — pump cycles, temperature drops, vibration patterns, work-order history — it shows up pre-grounded. The training was hard. Using it is easy.

The Operator Who Already Has the Data Wins

When I tell a building owner that we monitor their pump, their HVAC, their panel readings — and explain that the reasoning over that data is going to get dramatically better in the next twelve months — the response is almost always the same: "So I should wait."

The exact opposite is true. The data has to be there before the reasoner can use it. The pump that has been monitored for two years has 730 days of cycle history, 730 days of float-switch behavior, 730 days of correlation against weather, occupancy, and maintenance events. When the smarter reasoner shows up, it has something to reason about. The pump that was not being monitored has nothing.

Cadence and NVIDIA closing the sim-to-real gap for humanoids is the second-most interesting story in physical AI this week. The first is what happens to small-building monitoring when the reasoning models trained against those better simulators land on the edge nodes that have been quietly accumulating real data for years.

The robotics industry is solving a hard problem. The building practitioner's job is to make sure the reasoner has something to reason about when it arrives.

Read the case studies: Making a sump pump agentic | Buildings don't need a world model — they have yesterday's data | DreamDojo doesn't live in your basement | Physical AI is not a robot