What GPS-Denied Drones Taught Me About Building Monitoring

There is a drone flying over ETH Zurich's robotics lab right now that has no idea where it is — in the GPS sense. It doesn't know its global coordinates. It doesn't have a persistent map of the building it's navigating. It has never talked to a server to confirm its position.

And it's flying just fine.

The HUNT framework, published this month by researchers at ETH, solves GPS-denied autonomous flight by doing something that seems almost obvious once you hear it: instead of maintaining a global location, the drone rebuilds its understanding of the world at every moment from the signals it can directly observe right now. Barometric altitude. Visual motion cues. Inertial attitude. Target geometry. No external infrastructure required. No cloud handshake. No map that needs to stay current.

The architecture has a name in the paper: instantaneous relative frames. And it's the same architecture I've been running in a utility closet for two years.

The GPS Problem Is a Building Problem

Most IoT-based building monitoring systems are GPS-dependent in a metaphorical sense. They depend on an external reference point to know what "normal" means.

That reference point is usually a cloud service. The sensor sends its reading. The cloud compares it to a historical baseline stored in a database on someone else's servers. The cloud decides whether to send an alert. The alert arrives on your phone, eventually, if the API is responding.

This architecture has a fragility problem that reveals itself at the worst times. A flood starts at 3am. Internet is down because the same storm that caused the flood knocked out your router. The cloud service your system depends on had a maintenance window. The alert never comes.

The drone doesn't have this problem because the drone was designed from the start to not have it. The HUNT designers made a deliberate architectural choice: remove the dependency on global infrastructure. Make the system reason from what it can locally observe. Build resilience into the design, not as a fallback.

That's edge-first architecture. And it's the only kind I build.

The Sump Pump That Proved the Point

Two years ago I put a $25 Shelly smart plug on a sump pump. The plug reports wattage every few seconds to a local Home Assistant instance running on a small computer in the same building. A local AI model watches that data stream.

The model doesn't know what a sump pump "should" look like in any global sense. It doesn't call a cloud API to ask. It builds a baseline from what it observes over the first few days of operation — the pump's characteristic power draw during normal cycles, the duration of each cycle, the frequency at which cycles occur given rainfall. That baseline is rebuilt locally. The model reasons from directly observable local signals.

That's the drone in the building. Instantaneous relative frames. Local signal processing. No GPS required.

Why This Architecture Is About to Get Much More Powerful

This week NVIDIA released Cosmos Reason 2 on Hugging Face. It's a vision-language model designed for machines to "see, understand and act in the physical world." It runs on the Jetson T4000 — the 1,200 TFLOPS, 70-watt compute module we talked about last week that costs under $2,000 at production volumes.

This matters for building monitoring in a specific way. Until now, edge-first building AI was mostly numerical. Power draw. Temperature. Humidity. Runtime patterns. Numbers that a local model could watch and reason about without any external dependency.

Cosmos Reason 2 adds vision. Point a camera at a mechanical room and ask the model to describe what it observes — a condensation pattern on a pipe that suggests an insulation failure, a belt that's misaligned in a way that creates a characteristic vibration signature, a water stain that appeared between Tuesday and Wednesday's scan. All of that reasoning happens on the Jetson. None of it touches a cloud service.

The HUNT drone doesn't just navigate. It understands its environment from locally-observable signals and takes action. A building monitoring system with Cosmos Reason 2 in the utility closet does the same thing for the equipment that keeps your building running.

What "Few-Shot" Means for Predictive Maintenance

One of the consistent objections I hear from building owners is a data objection: "We don't have enough historical failure data to train an AI system."

This was a legitimate concern two years ago. Training a useful anomaly detection model required months of baseline data, ideally including examples of actual failures to train against. Most small buildings don't have that data. And they're not going to accumulate it by waiting.

Edge AI research in 2026 has largely solved this problem. Few-shot and transfer learning approaches mean a model can build a useful baseline from days of observation rather than months. The model doesn't need to have seen your specific pump fail before. It needs to understand the general physics of rotating machinery under load — and you can fine-tune that understanding on your specific equipment in a short commissioning window.

The 40-device community center building I monitor went from installation to meaningful anomaly detection in under two weeks. The HVAC pattern that surfaced — running on weekends with near-zero occupancy — didn't require failure data. It required a model that understood the relationship between occupancy schedules and HVAC runtime. Once the model had two weeks of occupancy-correlated data, the anomaly was obvious.

Fixing that scheduling mismatch cut energy costs 42%.

The Principle That Connects All of This

The drone paper, the sump pump, the community center HVAC, the new NVIDIA open-source models — they're all expressions of the same underlying principle. Systems that reason from what they can locally observe are more reliable than systems that depend on external reference points.

Don't maintain a global map that can go stale. Rebuild your understanding of the world at each moment from the signals that are directly observable right now.

For a drone, those signals are altitude, motion, and geometry. For a building, they're power draw, temperature gradients, humidity, runtime patterns, and now, with vision models running at the edge, the physical state of your equipment as seen by a camera.

The enterprise building automation vendors — Siemens, Johnson Controls, Honeywell — have been building GPS-dependent systems for twenty years. Their intelligence lives in the cloud. Their baselines live on servers. Their anomaly detection depends on API calls. That architecture costs $8 to $15 per square foot to install and comes with ongoing subscription fees that you're locked into as long as the vendor stays in business.

Edge-first building monitoring costs a fraction of that. The intelligence lives in the building. The baseline lives in the building. The decisions happen in the building. The only thing that leaves is an alert — to your phone, to your email, to whatever endpoint you want — when something actually requires your attention.

What to Do With This

If you manage a building — a church, an office, a retail space, a light industrial facility — and your monitoring strategy is "we'll notice when something breaks," you are flying GPS-dependent in a GPS-denied environment. You're betting that the internet is up, the cloud service is running, and the alert reaches the right person before the failure becomes expensive.

That bet fails. Not always. But with enough buildings, over enough time, it fails.

The hardware to deploy edge-first building monitoring — the kind that reasons locally, alerts reliably, and doesn't depend on external infrastructure to function — starts at a few hundred dollars for a small deployment. The Jetson Orin Nano Super developer kit is $249. The sensors that matter most cost between $15 and $25 each. The software stack is open-source.

What takes expertise isn't the hardware. It's knowing which signals to watch, how to build the baseline correctly, and how to tune the anomaly detection so it catches real problems without flooding you with noise. That's the work I do.