A drone flying between two high-rises at 200 feet loses its satellite fix for 4 seconds. At cruising speed, that is roughly 80 meters of blind flight. For a delivery vehicle carrying a medical package to a rooftop pharmacy, those 4 seconds used to be a hard wall. The aircraft would hover, attempt signal recovery, and if recovery failed, return to its launch point or land in place. Years of sensor engineering have compressed that failure window to near zero, and the result is a generation of unmanned aircraft that can hold centimeter-level positional awareness in places where GPS goes dark. That capability is what separates a proof-of-concept drone demo from a working urban delivery network, and it is now operating at commercial volume.

The GPS Problem in Cities

Satellite positioning was designed with open sky in mind. Urban environments break the assumptions behind it. Tall structures bounce signals off walls and windows, producing multipath errors that can place a drone tens of meters from where it actually is. Narrow streets between buildings restrict the number of satellites the receiver can see, and below certain elevation angles the geometry of available satellites degrades accuracy further.

These conditions are not rare edge cases. They describe the default state of flight in most downtown cores. A drone relying on GPS alone in these areas would spend a large fraction of its route operating with degraded or absent position data. For years this was the primary bottleneck for autonomous delivery in dense areas.

How Drones Learn Where They Are Without GPS

Urban corridors filled with tall buildings, bridges, and overpasses create dead zones where satellite signals drop out entirely. Drone operators compensate by fusing LiDAR point clouds with inertial measurement units and stereo camera feeds, running them through extended Kalman filters that keep positional accuracy within centimeters. The same filtering logic applies when routing around temporary obstacles like construction cranes or event scaffolding.

Flight planning systems pull from location intelligence software, updated terrain models, and real-time airspace feeds to build routes that account for signal gaps before a drone ever leaves the ground. Zipline’s 120 million autonomous miles by January 2026 confirm that pre-computed routing paired with onboard sensor correction works at commercial scale.

What Extended Kalman Filters Actually Do

The phrase “sensor fusion” gets used loosely, so it is worth being specific. An extended Kalman filter takes noisy, contradictory inputs from several sensors and produces a single best estimate of the aircraft’s position, velocity, and orientation. When GPS feeds clean data, the filter weights that heavily. When GPS drops out, the filter increases its reliance on inertial measurements and visual odometry from stereo cameras, while using LiDAR point clouds to correct the drift that inertial sensors accumulate over time.

This is a continuous, probabilistic process. The filter runs dozens of times per second, updating its confidence in each sensor source based on the consistency of recent readings. The practical outcome is that the drone does not “know” it lost GPS in any discrete sense. It transitions smoothly between sensor regimes without interrupting its flight path.

The Regulatory Side Is Catching Up

The FAA has proposed Part 108, a rule intended to replace the current waiver system under Part 107 for operations beyond visual line of sight. The proposed framework covers aircraft up to 1,320 pounds and introduces 5 risk categories organized by population density, along with 2 approval tiers for operators. Technical mandates under the proposal include detect-and-avoid systems, remote identification, continuous position tracking, and integration with unmanned traffic management systems. The final rule is expected in spring 2026.

In Europe, the U-space regulatory program has gone through testing phases but has not yet reached wide implementation. Testing so far has validated that the system architecture works, and several member states are preparing for phased rollouts.

Commercial Operations Are Already Scaling

Zipline completed its 2 millionth delivery by January 2026, backed by $800 million raised in its Series H round. Wing has announced plans to expand service to 150 additional Walmart store locations through 2027, with Bay Area deliveries beginning this year. These are not pilot programs. They represent sustained commercial operations generating revenue and fulfilling real orders.

The sensor fusion stack described above underpins this scaling. Without reliable position holding in GPS-degraded areas, none of these operators could serve urban and suburban customers at the volumes they are reporting.

What This Means for Route Planning

Pre-flight route computation now accounts for known satellite dead zones, temporary airspace restrictions, and obstacle databases that include construction activity. The onboard system then handles deviations in real time. If a crane has gone up since the route was planned, stereo vision and LiDAR detect the new structure, and the aircraft adjusts its path locally while the Kalman filter maintains position accuracy.

This two-layer approach, where ground-based planning handles the known and airborne sensors handle the unknown, is what makes reliable urban delivery possible. Each layer covers the weaknesses of the other.

The Remaining Hard Problems

Wind modeling between buildings remains difficult. Downdrafts and turbulence near building edges are hard to predict and vary with weather conditions. Acoustic compliance is another open question, as cities are starting to set noise limits for low-altitude aircraft that current rotor designs struggle to meet.

Battery density also constrains payload and range in ways that limit where urban delivery is profitable. A 15-minute flight window with a 2-pound payload works for pharmacy items and small retail orders. Heavier or longer-range missions require infrastructure like intermediate charging stations that few cities have planned for.

These are engineering and policy problems, not fundamental barriers. But they will take time, testing, and regulatory cooperation to resolve.