Qualcomm Patents a Self-Learning System That Figures Out Where Your Phone Is
GPS is notoriously bad indoors, but cell networks know a lot about where you are. Qualcomm is filing patents on a system that uses AI, trained on historical signal data, to figure out a phone's location by combining what the phone hears with what nearby towers hear.
How Qualcomm's AI-trained positioning actually locates your phone
Ever tried to navigate inside a big airport or mall and watched your GPS dot spin helplessly? Cell towers know roughly where you are, but "roughly" isn't very useful when you're trying to find Gate B14.
Qualcomm's patent describes a way for the cell network to send your phone a small AI model, trained on past signal measurements from that area, that helps the phone reason about where it probably is. The phone feeds in its own signal readings, and the network feeds in readings from nearby towers, and the two sets of information get fused together into a single best-guess location. The AI has already learned, from previous sessions, what patterns of signal strength and timing tend to correspond to which physical spots.
The result is a location estimate that doesn't rely on satellites at all. Instead, it leans on the radio signals already flying around you and a neural network that has, in a sense, memorized the signal fingerprint of the area.
… the at least one neural network function being generated dynamically based on machine-learning associated with one or more historical measurement procedures …
Translation: The system updates its artificial intelligence using data from past location checks.
How the cell tower and phone split the neural network work
The system has two main actors: the base station (the cell tower or network node) and the user equipment, or UE (your phone or any connected device).
The base station runs machine-learning on historical positioning measurements collected over time from that area. It produces a set of neural network functions, essentially compact AI models that encode what signals look like at different locations. These models get transmitted down to the phone as part of the normal network signaling process.
The phone then collects two types of data:
- First positioning measurement data: what the phone itself hears from towers, things like signal timing, strength, and angle of arrival.
- Second positioning measurement data: readings taken by the towers or other network nodes about the phone's transmissions, sent back to the phone as assistance information.
Each data type gets fed into a matching neural network function (one tuned for phone-side measurements, one tuned for tower-side measurements). The outputs are then fused (combined mathematically) to produce a single positioning estimate. The key idea is that instead of simply triangulating, the system uses the neural network to calculate a likelihood: given these signals, how probable is it that the phone is at this particular candidate location? The highest-probability candidate wins.
… derive a likelihood of at least one set of positioning measurement features being present at a candidate set of positioning estimates for the UE …
Translation: It calculates the probability that the device is at a specific location based on available signals.
What this means for GPS-starved indoor navigation
For most people, this matters most in the places GPS already fails: dense city blocks, underground transit, large buildings, and stadiums. If a phone can get an accurate fix from cell signals alone, apps that depend on location, from turn-by-turn directions to emergency services, become more reliable in exactly the environments where they're most needed.
Qualcomm's steady investment in AI-driven positioning reflects a broader shift in how chipmakers are thinking about the problem. The approach described here pushes real AI computation onto the device itself rather than sending raw data to a server, which has privacy and speed advantages. Whether that tradeoff holds up at scale, across millions of devices in constantly changing radio environments, is the open question.
This is the 12th Qualcomm filing we've tracked in our on-device AI privacy watchlist since July, building on earlier work around training network traffic AI and AI vision for smart glasses.
Pushing a trained location model onto the phone trades server costs and privacy risk for a different kind of fragility: the model only knows what the world looked like when it was trained. Radio signals bounce differently when a new building goes up or a stadium fills with people, and a stale model has no way to know it's wrong.
That creates a hidden maintenance burden the patent doesn't address. Someone has to decide how often to retrain and re-push these models, and getting that cadence wrong means phones confidently reporting bad locations, which is worse than admitting uncertainty.
The underlying idea, combining what the phone senses with what the towers sense to cover each other's blind spots, is sound and addresses a real gap in how location works today. Whether the accuracy gains justify the ongoing infrastructure commitment to keep models fresh is the question this design raises but leaves open.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
18 drawing sheets from US 2026/0268112 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →