- A GPS satellite's signal reaches your drone with roughly the power of a 25-watt bulb seen from 10,000 miles away. A jammer a few km off is, in path-loss terms, a million times closer. This isn't a bad GPS chip. It's the inverse-square law. The sky is contested now. š§µ
- Russia's R-330Zh "Zhitel" throws 30+ km GPS-denial bubbles across the Ukrainian front. Inside them, 50%+ of FPV drones are lost to electronic warfare. Front-line crews now build airframes with no GPS receiver at all ā because GPS is no longer assumed to be there.
- Then the procurement floor dropped out. On Dec 22, 2025 the FCC added every foreign-made drone and UAS critical component to its Covered List. Congress had asked it to act on DJI and Autel. The FCC did all of them. No new authorizations for any of it.
- The signal that followed was unmistakable: the US Army put $52M behind 2,500+ Skydio X10D drones in March 2026 ā its largest single-vendor sUAS buy ever, bid-to-award in under 72 hours. Nothing else in the cleared inventory could fly in a jammed environment.
- The civilian version isn't warfare ā it's geometry. An IMU is fast (1000 Hz) but noisy, and position comes from double-integrating acceleration. Error grows as tĀ². A consumer MEMS IMU left open-loop in a mine drifts meters in seconds. Then it finds a wall.
- The fix is Visual Inertial Odometry: fuse camera + IMU so the drone sees its own drift and corrects it. Skydio holds VIO to 984 ft. ModalAI's VOXL 2 packs it on a 15-TOPS Qualcomm module. Underground, Emesent Hovermap and Flyability Elios 3 run LiDAR SLAM in total dark.
- At LKAB's Kiruna iron ore mine, an 8-hour manual survey became a 20-minute drone flight ā inspection costs drop up to 70%. One pipeline failure runs $8.5M to clean up vs $75K to inspect, and a single mine-shaft crash totals a $10K-$50K airframe.
- The catch nobody markets: the open-source stack is real but not turnkey. SuperPoint-SLAM3 cuts KITTI drift 12x (4.15% ā 0.34%). But SuperPoint hits only 14 FPS on a Jetson Orin Nano. Without TensorRT INT8 + VPI offload you can't feed a 30 FPS control loop. That's the whole game.
- So the market splits. Skydio, Shield AI and Auterion own the closed, fully-integrated end ā we don't compete there. Even DoD's Replicator hit a wall buying software to command heterogeneous swarms. The gap is everyone else: Blue UAS OEMs with great frames and no VIO team.
- That's where we build. ORB-SLAM3 + a SuperPoint front end, TensorRT INT8 quantization, time-synced stereo calibration, PX4/ArduPilot integration on Jetson Orin NX ā delivered as a flight-tested payload. 12-24 months of robotics work, condensed onto your existing airframe.
- Honest question for people doing this: are you betting on cuVSLAM / Isaac ROS commoditizing VIO to good-enough ā or does contested-environment autonomy still need a hand-tuned SuperPoint + TensorRT stack per airframe? Where's the line? #BlueUAS #EdgeAI
- We wrote up the full landscape ā the physics, the FCC action, the vendor map, and exactly where a focused engagement fits: https://veriprajna.com/solutions/gps-denied-drone-autonomy