# Sensor Fusion based UAVs for GPS denied Environments Topic radar · 245 documents · 2021-01-01–2026-10-23 · updated 2026-10-24 Canonical: https://slate.greyb.com/radar/sensor-fusion-based-uavs-for-gps-denied-environments/innovations ## In one minute ### The state vector is where each design differs All of these systems propagate the same core: attitude, velocity, position, inertial biases and the camera-to-inertial extrinsics. What separates them is what else they promote into the state. A defence-university filing carries two-dimensional horizontal wind alongside attitude, velocity and position, so that a pitot observation can be admitted without its noise corrupting the speed estimate [1]. A landing system carries ground altitude as its own node in a five-factor graph — visual, electromagnetic altimeter, barometer, inertial, and a layer-height random-walk factor — so the ground plane is estimated rather than assumed [2]. A swarm design carries an unknown UWB anchor's coordinates inside the sliding window and initialises them in flight from the vehicle's own visual-inertial trajectory [3]. The rule is consistent: a quantity that would otherwise be a fixed parameter is promoted into the state whenever trusting it is worse than estimating it. Wind, ground height, anchor geometry. Each promotion buys robustness and pays in observability — the estimate only converges if the trajectory excites that mode, which is why the anchor design needs a manoeuvre before the anchor is usable. 1. [CN119915297A](/?paper=CN119915297A&utm_source=radar&utm_medium=dashboard) 2. [CN116380057B](/?paper=CN116380057B&utm_source=radar&utm_medium=dashboard) 3. [CN121898372A](/?paper=CN121898372A&utm_source=radar&utm_medium=dashboard) ### Every published performance figure is the seller's own The specification language converged before the evidence did. Several sellers now quote drift as a fraction of distance flown rather than metres per hour, and quote integrity as a time to alert, which is the right shape for a buyer writing a mission requirement. The numbers themselves come from the vendor. UAV Navigation's visual system reports drift near zero over mapped terrain, about 1% of distance travelled over unmapped terrain, and bounded to 30 m where imagery and elevation data exist; a $399 module claims 2% of distance travelled, in daylight, with no night operation supported. Both are plausible. Neither is a test report. ### The marginal filing is about admitting a measurement, and almost none is about sensing The newest work in the field argues about the arbiter. A framework published in Sensors holds GNSS measurements out of the fusion window until integrity has been assessed, then re-solves them as factors. Its check runs against visual-inertial odometry rather than against surveyed truth, so the reference for deciding whether a fix is good is itself a relative estimate with its own drift. It reports detection probabilities above 96% for gradual attacks at 0.5 m/s over a 10–15 second window. ### The front end is commodity; the differentiation has moved to what gets rejected FAST corners split on a quadtree for uniform distribution, KLT optical flow on an image pyramid, PROSAC outlier rejection, DBoW loop closure over BRIEF descriptors: that is one filing's entire front end, and none of those components is distinctive to it. IMU pre-integration is the universal interface. Inertial samples are integrated between image timestamps so the inertial term becomes a residual at camera rate, minimised by Mahalanobis distance against a marginalisation prior in a sliding-window solve. This is the piece that makes tight coupling tractable on flight hardware, and it is no longer where anyone claims novelty. ### The only company in the field with audited accounts is shrinking Exyn's public filings are the sole audited revenue base in this market, and they set the scale. The company raised roughly $19.4 million gross at its May 2026 listing; in the quarter to 30 June 2026 it reported 950,000 of revenue, down 30.1%, a gross margin of 46.9%, and a net loss of 6.9 million. Cash was about 7.8 million at quarter end and roughly 4.5 million by 19 August, with the company stating that this will not cover a year of operations, and its chief executive departing after an internal investigation. ### Universities and state asset owners hold the claims, and navigation vendors hold the products The filing centre of gravity is Chinese and Indian academic and state institutions. The most active filers are IIT Hyderabad, an Indian autonomous-navigation technology hub, Northwestern Polytechnical University and State Grid, with provincial grid operators and municipal engineering institutes behind them. No European, American or Israeli inertial house appears among the significant filers. ### Height is solved on a separate channel, against a different reference The horizontal and vertical problems have different error sources — pressure, ground effect and terrain-model staleness on one side; cross-view matching failure on the other — and the filings increasingly solve them apart. One runs parallel bundle adjustment with a planar branch for horizontal position and a DEM branch for vertical, then arbitrates between them by re-solving on randomly sampled grid points and checking Euclidean consistency. Another scales the live image against the satellite tile using ground sampling distance, focal length and barometric altitude before matching at all, which turns height into an input to the horizontal solve rather than an output of it. A third corrects the barometer itself, using inertial speed and a wind estimate inferred from the differential RPM of the rotors. ### The largest public cheque in the field buys a terrestrial backup, not onboard fusion The United Kingdom committed £155 million to positioning resilience, and the split is the most informative single allocation in this market: £71 million for an enhanced terrestrial ranging system, £68 million for a national timing centre, £13 million for interference monitoring and £3 million for space-based time transfer. None of it is for onboard multi-sensor fusion. The demand is real and quantified — a single day of satellite navigation outage is estimated at £1.42 billion to the UK economy, seven days at £7.64 billion — and the money went to timing and to a ground network. ## Overview The estimation problem is no longer the constraint. The patents and papers here have converged on one architecture, and the literature reports sub-metre accuracy inside structured spaces and roughly one to two metres for short outages with relative aiding. What remains unsolved is the thing each design measures against. Every architecture is a claim on a reference the vehicle does not own — satellite imagery, a magnetic survey, a surveyed anchor set, a constellation's ephemeris, or an inertial core's own drift — and the filings spend their words on deciding whether to believe a measurement rather than on producing a better one. The estimator has become reproducible firmware, and the value, the risk and the unresolved engineering have all moved into the reference and into the evidence that the estimate was trustworthy while it was produced. ## Research ### The reference library sits in the loop, not behind it Matching against a stored reference only works if the reference is current, so the map has become a maintained artefact that the aircraft updates. After a mission, one design fuses the recovered aerial frames back into the satellite map by affine transform, so the next flight matches against a map the last flight improved. Another generates the reference from public imagery as a spatially organised multi-resolution feature database, and places a validation module between the match and the fusion update so an implausible fix is filtered before it reaches the estimator. ### The flight controller never sees the fusion layer The output is shaped as a sensor message. One filing encapsulates the optimised pose as a pseudo-GPS signal over MAVLink, so an unmodified autopilot consumes it exactly as it would a receiver fix. Another switches on measured signal-to-noise, visible-satellite count and a positioning-confidence threshold — loosely coupled Kalman with GNSS inside the loop while the thresholds pass, factor-graph multi-source optimisation when they do not. ### Compute class splits one algorithm family into two products The same tightly coupled family of estimators appears across a wide hardware range. One MCU-class design puts the whole chain on a Cortex-M7, which forces the feature culling, blur correction and lightweight optimisation to be designed in rather than added. A binocular design instead adopts MSCKF, using inertial pre-integration for state recursion and visual reprojection residuals for updates, then hands pose and a 2D occupancy grid to a DWA local planner on the same onboard computer. At the other end, a lidar-visual-inertial mapper assumes a 1000 Hz industrial IMU and an Intel NUC with 32 GB, and reconstructs a UAV-centred occupancy voxel map in a circular buffer. ## Market ### The requirement has been written sensor-agnostic A US Air Force Special Warfare market-research notice asks for a compact electronic navigation system that can integrate magnetic, inertial, gravity, visual, celestial, radar/lidar and quantum sensors, output position compatible with a team awareness application, and serve aircraft, land vehicles, maritime platforms and dismounted users. That is a buyer specifying the interface, the output and the environmental envelope, and explicitly leaving the sensing principle open. ### A 399 module, a 50,000 gyro, and a thin band between them The prices are vendor and market-report figures and should be read as bands, not quotes. The structure they reveal is the point: the market is thick at the bottom, where a prior-free module is cheaper than the airframe's flight controller, and thick at the top, where inertial quality is bought by the kilogram and the dollar. The middle is where a mid-performance fusion unit with an auditable specification would sit, and nothing in the public record occupies it convincingly. That is the opening for a supplier who can publish a characterisation rather than a claim. ### Published market sizes contradict each other and should not be used to plan Three reports build on the same $1.35 billion 2025 base for assured positioning and navigation and then diverge: one reaches $1.74 billion in 2026 at 28.9% growth, a second reaches $10.2 billion by 2035 at 28.4%, and a separate line of coverage puts the same capability at 400 million in 2022 rising to 3.5 billion by 2032. A fourth tracker prices GPS-denied military inertial navigation alone at $9.42 billion in 2025. ## Strategy ### Every commitment of the past two months buys a reference or an inertial core In September and October 2026 the announced commitments were uniformly about new bounds. A fielded visual navigation unit added terrain-referenced navigation and satellite map matching as a software update, giving an existing product two more absolute references. Northrop Grumman and SandboxAQ flew quantum magnetic navigation on a Group 3 airframe one month after setting the objective. The European Defence Agency closed out a four-year, €1.2M project that demonstrated overland flight without satellite signals. The European Space Agency funded work to extract navigation observables from Ka-band low-Earth-orbit communication signals. ### The next reference is a funded constellation with a launch schedule The signals-of-opportunity work is the clearest statement of direction. It proposes to read position from the Ka-band traffic of broadband constellations, with the EU's planned IRIS² network named as the intended source and first launches reported at 2029. These signals are not designed for positioning and carry no service commitment for it. ### Patent effort aims at embedded compute while the money aims at fibre-optic gyroscopes The recent filings are landing at the cheap end of the compute range. One puts the whole tightly coupled chain on a Cortex-M7, which forces feature culling and motion-blur correction to be designed in from the start. Another describes a binocular MSCKF design with a local planner on the same onboard computer, without naming the processor. ### Reference maintenance has no funded owner The filings spend their claims on deciding whether a match is trustworthy. The reference being matched against has no equivalent claimant. Nothing in the announced programmes funds the currency of the imagery, elevation model or magnetic survey, and no funded work tests what happens when the stored reference is wrong before the aircraft arrives. ## Areas - Multi-Sensor Fusion Algorithms & SLAM · 128 documents - Alternative Navigation Technologies · 57 documents - Defence & Contested Navigation · 28 documents - Infrastructure Inspection & Confined Spaces · 8 documents - Commercial Autonomous Operations · 13 documents - Ecosystem & Collaboration · 9 documents