This week, during National Robotics Week, NVIDIA released a set of open AI models they're calling the foundation of embodied AI. The centerpiece is a concept Jim Fan — NVIDIA's Director of AI and co-lead of Project GR00T — calls the shift from "next-word prediction" to "next-physical-state prediction."
The idea is that language models predict the next token. A truly physical AI predicts the next state of the physical world. What will the robot arm be holding in 0.3 seconds? Where will the autonomous vehicle be when that car door opens? How much current will the motor draw under tomorrow morning's load?
My sump pump has been doing that for two years.
What "Physical State Prediction" Actually Looks Like at Ground Level
I monitor a residential property with a sump pump. The setup is not exotic: a $25 smart plug, a local AI model running on a small always-on machine, and a script that watches power draw patterns in real time. No cloud dependency. No subscription service. No vendor lock-in.
What the model does is exactly what NVIDIA is now building billion-dollar infrastructure to train humanoid robots to do: it learns what normal looks like, identifies deviations, and predicts what happens next if the deviation continues.
About a year ago, around 3am during a rain event with two inches of rainfall in the forecast, the model flagged an anomaly. The pump's power draw was spiking and releasing in a pattern that didn't match normal cycling. The float switch was sticking intermittently, forcing the motor to run under load without pumping. Not hard enough to trip a breaker. Exactly the kind of slow stress that burns out a motor over six to twelve months.
The system ran 97 automated recovery cycles before the float freed itself. No flood. No motor failure. No emergency call at 2am.
That is next-physical-state prediction. Small scale. Cheap hardware. Real outcome.
The Architecture Is the Same — The Scale Is Different
When NVIDIA describes what GR00T and the new Cosmos world models are doing, the architecture reads like a cleaned-up version of what practitioners have been cobbling together in buildings for years:
- Sensor fusion across multiple data streams (IMU, cameras, contact sensors — or in our case, power draw, temperature, humidity, door state)
- Local inference without constant cloud round-trips
- A learned model of "normal" physical state that flags deviations
- Prediction of future states based on current trajectory
The difference is that NVIDIA is running this on a robot with twenty-three degrees of freedom and a physics simulation that compresses ten years of training into two hours. I'm running it on a building with forty sensors and a model that fits in 4GB of RAM.
The distance between those two things is closing fast.
The Building That Taught Me This
The second system I run is a 40-device edge AI network inside a community center. Before the monitoring system, the building's energy bill was a number that arrived once a month and was paid without much analysis. After six weeks of data collection, the AI surfaced a pattern no one had seen: the HVAC was running on weekends when occupancy was near zero.
Not because the system was broken. Because nobody had ever correlated the HVAC schedule to the booking calendar. The physical state of the building on Saturday morning (empty, doors closed, no events) was not matching the predicted state the HVAC system was acting on (treat it like a Tuesday morning).
Fixing the correlation cut energy costs 42%. That's a physical state prediction problem. The sensor data described a physical reality. The AI predicted what that reality implied for cost and comfort. The output was an action.
This is what embodied AI looks like when the body is a building.
Why This Week's News Matters for Your Building
NVIDIA's announcements are not abstract research. They are investments in the open infrastructure that small practitioners will use in the near term.
Newton 1.0 is an open-source physics engine. It exists to make robot simulation accurate enough that models trained in simulation work in the real world. But physics simulation that models how things vibrate, wear, and fail is directly applicable to predictive maintenance in buildings. A motor that's failing has a characteristic vibration signature. A valve that's wearing has a pressure differential pattern. Newton is being built to model exactly that kind of physical behavior.
Cosmos world models are designed to generate synthetic training data at scale — letting you train a robot on a billion scenarios it's never physically encountered. The same approach works for a building system that's never experienced a winter power outage, a tenant renovation, or a HVAC compressor failing mid-summer. You don't need to wait for the failure to train for it.
These tools are being open-sourced because NVIDIA's business model is selling the compute that runs them. That's good for practitioners. Open models commoditize the intelligence layer. The value moves to the people who know how to deploy it in real physical environments.
What You Should Do About It
If you manage a building — or several — here's the practical takeaway from this week.
The research world is proving, with enormous investment and attention, that physical state prediction from sensor data is both feasible and valuable. They're proving it in robots. The same proof applies to your building.
You don't need NVIDIA's simulation infrastructure to get started. You need a $25 smart plug on your highest-risk equipment, a cheap always-on machine running a local model, and someone who knows how to wire them together and interpret what the data is telling you.
That's what I do. The sump pump didn't need a humanoid robot to save it. It needed 24 hours of baseline power data and a model that knew what normal looked like.
"Physical AI is not about the robot. It's about closing the loop between the physical world and the decision-making system — wherever that loop exists."
Your building is a physical system. It has state. It deviates from normal. Those deviations predict failures, waste, and cost overruns — before they happen. The research is catching up to what practitioners have been quietly doing for years.
The question is whether you're running this on your building yet.
See What Physical State Prediction Looks Like in Your Building
I work with building owners to deploy edge AI monitoring that runs locally, costs less than a maintenance call, and tells you what's happening before something fails. No cloud dependency. No vendor lock-in.
Learn About Physical AI for Buildings →