GPS is a single point of failure that jamming, spoofing, or a dense city skyline can take down. The drones that keep flying without it don’t carry a better receiver. They carry software that can recognize where they are from a camera, and that layer is quietly becoming foundational.

For roughly 18 months, parts of Israel had no reliable civilian GPS. In Ukraine, both sides jam the spectrum so heavily that a drone which loses its signal loses the mission, and often the aircraft with it. GPS is dependable infrastructure that took decades to build, and the world got so used to it that almost every device now assumes it’s there. That dependence is the problem. When one signal underpins everything, it becomes the first thing an adversary switches off. The military has a name for the fix: assured positioning, navigation, and timing, or assured PNT, a position you can trust even when the signal can’t be.

GPS doesn’t just get jammed

Jamming is the obvious threat, but it isn’t the only one. GPS gets spoofed, fed a false position that’s worse than no position at all. Part of why both are so easy is physics. By the time the signal reaches the ground from about 20,000 kilometers up, it’s fainter than the background radio noise. A jammer the size of a cigarette pack can bury it across a wide area, and a stronger fake can slip into its place. And GPS fails with no adversary involved at all: an urban canyon between tall buildings, a valley in the mountains, and the signal degrades or disappears. So redundancy isn’t a nice-to-have for anything that matters. A vehicle in the air that relies on a single signal is a vehicle you can lose the moment that signal drops.

On the commercial side, GPS-free navigation used to be one box to tick for redundancy. Ukraine changed that. For a reconnaissance drone under electronic warfare, it stopped being redundant and became the difference between a working system and a useless one.

How a machine sees where it is

When GPS drops, the fallback is the camera. The system compares what it sees now against a stored reference map and works out its position from the match. That sounds simple. It is one of the harder problems in the field.

The core of it is real-time registration: matching a live image to a map fast enough to fly on, using very little computation. Some of the most useful work in the field started right there, on how to match two photos in real time efficiently, and the difficulty compounds from there. Match a thermal camera at night against a map made in daylight. Match a map that’s three years out of date against ground that has changed. Match across the angles a real pilot points the camera, not the clean overhead view of a lab test. Each of those is a separate technical problem, and solving them on a small power budget is where the defensibility lives.

Vision doesn’t work alone, and GPS isn’t going anywhere. Inertial navigation systems are steady but they drift over time, accumulating error. Vision corrects the drift. The strongest systems fuse all three, GPS when it’s available, inertial always, and vision to anchor it, so no single failure ends the mission. A flying taxi over a city will need exactly that stack, because the safety bar leaves no room for a single point of failure.

The moat is software that scales, not the sensor

Here is the part that matters for anyone deciding what to build or back. The camera and the processor are commodities. The value is the autonomy stack that runs on them, and the winning version is software-first and modular.

Modular matters more than it sounds. A small tactical drone has no spare payload, so you put the computer on the ground and keep the air side hardware-free, leaning on the communication link. A larger fixed-wing flying at altitude has the weight and power to spare, so you put the full camera-and-processor payload onboard and let it navigate as a self-contained sensor. Same software, different configuration, depending on what the platform can carry. That flexibility is what lets one product serve a whole mixed fleet.

The real challenge is scale and cost, not capability in isolation. Autonomy was available years ago, but it lived on expensive, high-end UAVs that could carry redundant systems. The hard problem is putting it on thousands of cheap, off-the-shelf drones, plug-and-play, at a price that works across a fleet. It’s a deployed footprint: systems already running on thousands of drones across platforms, from off-the-shelf quadcopters to high-altitude fixed-wing. That kind of fielded scale is the signal to look for.

Assisted beats autonomous, for now

There is a temptation to chase full autonomy, Level 5, the drone that does the whole mission untouched. That is not what the market wants here and now, and it isn’t what it will pay for. Picture the actual user: a drone pilot in a battlefield, flying at night with the aircraft’s lights off, under electronic warfare that’s stripping their position, running a reconnaissance mission, all at the same time. They don’t need the mission fully automated. They need to be hands-free enough to focus on it and not lose the drone.

That’s why the first place autonomy pays off is takeoff and landing, which is where most accidents happen in crewed and uncrewed aircraft alike. Take the highest-risk, highest-workload moments off the pilot’s plate, match the autonomy level to the cost and weight the platform can bear, and you have something people can actually deploy. A perfect system nobody can afford or install is worth nothing in the field.

For more on how this plays out in the field, the UAX podcast goes deeper with the people building it.