1. Executive Summary

The unmanned combat drone has crossed the threshold from a remotely piloted tool to a weapon that acts on its own. This study surveys the scientific state of the art across seven disciplines — inertial navigation, vision-based perception, terrain-referenced positioning, swarm radio, phased-array antenna technology, defense against satellite jamming, and embedded artificial intelligence — and projects from them the functionality that autonomous strike systems can be expected to reach by the year 2036.

At its center stands a guiding thesis that lends the study its name: the fail-safe kill chain. The combat drone of 2036 identifies its target with its own onboard sensors and its own artificial intelligence, without a data link. It networks with its peers over electronically steered directional beams that are hard to intercept and resistant to jamming. And should this nervous system fail, a tiered onboard navigation carries the mission to its end on its own — from the satellite signal through vision- and terrain-based positioning down to pure inertia, which no longer needs any external source.

The examination confirms the thesis but corrects its naive form. Pure inertial navigation drifts without halt; resilience arises not from a perfect sensor but from the way the methods catch one another, so that the drone is never reliant on a single source for long enough to matter. Across the few seconds of the terminal approach, even pure inertia suffices. The loss of swarm communication costs the collective sharpness, not the capacity to act — and therein lies both the military explosiveness and the legal problem.

The study also considers the adversary: against radio-silent, autonomous swarms, electronic jamming remains ineffective, so the defense falls back on high-power microwaves, high-energy lasers, kinetic interceptors, and programmable flak — means whose worth is decided by a merciless economics. It closes with the legal and ethical assessment: the fail-safe kill chain has been built; the open question is no longer whether it works, but who retains control over the decision of life and death.

2. Introduction

2.1 Background: the drone as a turning point in warfare

For more than a century, air warfare followed a simple logic: whoever wanted to command the sky needed expensive platforms, trained pilots, and an unbroken chain of command from the operations center to the cockpit. That logic is falling apart. In its place comes an aircraft that costs little more than a midrange car, that no one has to pilot by remote, and that appears in swarms whose individual members are as expendable as cartridges. The unmanned strike drone has crossed the threshold from remotely piloted camera to a weapon that acts on its own — and it has done so faster than the doctrines of armed forces and international law could follow.

The driver of this upheaval is not a single breakthrough but the convergence of several lines of development that ran separately for decades. Microelectronics has shrunk inertial sensors from the size of a coffee can onto a chip of a few grams, without any loss of accuracy. Machine vision now processes image streams in microseconds rather than in frame rates. Radio networks organize themselves when nodes drop out. And artificial intelligence, which only a few years ago demanded entire data centers, now runs in the single-digit watt range on embedded hardware. Each of these lines is remarkable on its own; their fusion into a single, palm-sized platform changes the very nature of combat.

2.2 Research question and aim

This study asks about the state of the art — and about what grows out of it by the year 2036. Three questions guide the inquiry:

How does an autonomous combat drone navigate, communicate, and think according to the current state of research, when the relevant disciplines — inertial navigation, vision-based perception, terrain-referenced positioning, swarm radio, intercept-resistant antenna technology, defense against satellite jamming, and embedded artificial intelligence — are laid side by side?

What functionality will grow out of it? At the center stand onboard automatic target recognition, covert swarm communication over electronically steered phased-array antennas, and — as the argumentative core — resilience against the loss of precisely that communication.

Where do the limits lie — technical, economic, legal? A study that celebrates only the possible and conceals the impossible would be propaganda, not analysis.

2.3 Methodology, source base, and time horizon

The inquiry is a synthesis of the literature and the technology. Its factual base draws, wherever possible, on primary scientific publications — peer-reviewed journals, papers from recognized conferences, and vetted preprints — as well as on primary documents of international law. Manufacturer specifications, market forecasts, and industry roadmaps are used where they constitute the only available evidence for concrete performance figures; throughout the text they are marked as such and not commingled with peer-reviewed measurement results. Every substantive claim carries a short citation; the complete source list at the end provides each piece of evidence with its reference and, where available, a DOI.

The time horizon reaches from the documented present to a reasoned projection onto the year 2036. The projection extrapolates no wishful thinking but extends documented development curves — and it names where the extrapolation runs into physical or economic limits.

2.4 Definitions and scope

The term combat drone is used broadly here: it means an unmanned aircraft that produces a military effect, whether it appears as a reusable platform, as a single-use loitering munition, or as a member of a cooperating swarm. What distinguishes it from the remotely piloted reconnaissance drone is the capacity for autonomy — for independent navigation, perception, and, in the extreme case, target selection without continuous human input. Where this study speaks of an autonomous weapon system, it adopts the functional definition of the International Committee of the Red Cross: a system that, once activated, selects and engages targets on the basis of a generalized target profile, without a human determining the specific target, the location, or the timing of the attack [1]. This definition is chosen deliberately, because it marks the legally and ethically decisive threshold (Chapter 14).

2.5 Guiding thesis: the fail-safe kill chain

The study pursues a single, continuous idea that holds it together like a backbone. It runs: The combat drone of the year 2036 is built so that the loss of its connections does not stop it. It identifies its target onboard, with its own sensors and its own artificial intelligence. It networks with its peers over electronically steered directional beams that are hard to intercept and difficult to jam. And if this nervous system fails all the same — if the radio falls silent and the satellite signal is choked off — then a tiered onboard navigation carries the mission to its end on its own: from the satellite signal through vision-based and terrain-referenced positioning down to pure inertia, which no longer needs any external source.

This thesis is seductive in its clarity, and for precisely that reason it demands rigor. Pure inertial navigation drifts, and it drifts without halt, the longer it is left to its own devices [2]. The fail-safe kill chain works not because any single sensor is perfect, but because the methods catch one another: what inertia loses over time, the view of the terrain wins back, and what the satellite signal accomplishes in open sky is taken over in its absence by the map beneath the wing. The study examines this chain link by link — and it says honestly where it breaks.

The structure follows this idea. Part I surveys the state of the art: the inertial backbone (Chapter 3), the artificial eye made of vision- and terrain-based navigation (Chapter 4), the radiating nervous system of the swarm (Chapter 5), the electronic shield against satellite jamming (Chapter 6), and the thinking brain onboard (Chapter 7). Part II assembles these building blocks into the expected functionality: automatic target recognition (Chapter 8), the fail-safe kill chain (Chapter 9), and the autonomous swarm as a weapon system (Chapter 10). Part III turns the perspective around and considers the adversary — the defense against drone swarms (Chapter 11), its merciless economics (Chapter 12) — and places the whole in context: technically (Chapter 13), legally and ethically (Chapter 14), before the conclusion (Chapter 15) takes stock of the guiding thesis.

Of the three pillars of the fail-safe kill chain, each can be switched off — all but one. A phased-array mesh can be jammed, a GNSS signal spoofed, an optical camera blinded. What remains, when the radio link falls silent and the sky over the battlefield is electronically poisoned, is the drone’s ability to know its own motion from within itself. An inertial measurement unit (IMU) continuously measures linear accelerations and rotation rates and, by integration, computes from them the flight path — without a single signal from outside [3]. Therein lies its strategic value: whoever receives no radio waves is, in principle, immune to jamming and spoofing. It navigates relative to the initial state — the backbone on which all the remaining senses hang. This chapter examines how far it carries on its own.

3.1 The MEMS-IMU: Miniaturizing Precision (SWaP-C)

For decades an iron coupling held: small and frugal meant inaccurate. Microsystems technology has broken it. Today the assessment runs by SWaP-C — size, weight, power, cost — and on all four axes the silicon-based MEMS-IMU (micro-electro-mechanical systems) has advanced into the tactical accuracy class [4]. The two levers of research are the same ones that later determine the drift rate: noise density and bias stability [4]. The industry sorts gyroscopes roughly into three bands — consumer/automotive above 10 °/h, tactical grade between about 0.1 and 10 °/h, navigation grade between 0.001 and 0.01 °/h [3], [4]; what matters is that commercial MEMS parts now reach the tactical band, at a fraction of the cost of their fiber-optic predecessors.

The reference devices show the span (figures in the table). The Sensonor STIM300 was long an industry standard — semiconductor-based, ITAR-free [5]; the newer GuideNav GUIDE730 undercuts it on all four axes and is made for the space-constrained micro-drone [5]. The Advanced Navigation Motus delivers, in 26 g, a gyroscopic performance once reserved for devices in the kilogram range [6], and Analog Devices markets the ADIS16490 as “precision tactical grade” with an angular random walk of 0.09 °/√h [7]. How robust MEMS have become for ballistic loads is shown by the capacitive Safran Colibrys MS1000: per datasheet, shock-tolerant up to 6,000 g and operable up to 150 °C, on a chip of 9 × 9 mm [8]. A symbolic threshold fell in 2025, when, according to the trade press, SBG Systems unveiled the first purely MEMS-based north-seeking gyrocompass — heading below 1° entirely without GNSS feed, at under 150 g and 2 W [9]. A function that previously belonged to mechanical ring laser gyroscopes has thereby migrated to the silicon scale.

Abbildung 1: MEMS-IMU accuracy classes by gyroscope bias instability (log scale). Source: Chapter 3.

All figures are manufacturer or datasheet specifications; an independent measurement of exactly these models under a uniform protocol is not available.

3.2 IMU Arrays and Virtual Gyroscopes

When a single cheap sensor is too noisy, numbers help. The noise components of several MEMS sensors are mutually uncorrelated and can be averaged out through data fusion, so that the ensemble behaves like a single, more precise sensor — a virtual gyroscope [10]. The rule of thumb behind this is the statistical averaging of uncorrelated errors: with N mutually independent sensors, the noise standard deviation falls by 1/√N — four sensors halve the noise, a hundred cut it to one-tenth. The 3.4-fold accuracy gain noted further below accordingly lies close to the ideal value of this square-root rule for the number of sensors installed [10]. The error modeling acts in two domains: in the time domain through models such as ARMA, in the frequency domain through Allan variance and power spectral density (PSD); these feed the adaptive weightings and the extended Kalman filter (EKF) that shape a smoothed state estimate out of the raw stream [10].

The gain is concrete. At Wuhan University, calibrating an array over a three-axis turntable raised the navigation accuracy by a factor of 3.4 — close to the theoretical maximum of noise suppression for the given number of sensors; in a real-world experiment the same array drifted, over a 30-second GNSS outage in a sports car, by only 1.85 m on average [10]. The calibration needs no external hardware: gravitational acceleration alone, as a reference, permits a self-contained self-calibration and improves the measurement accuracy by about 10 dB [10]. The redundant geometry yields a second virtue — fault detection and isolation (FDI) isolates a failed individual sensor without the estimate tipping over [10]. The array is thus more precise and more failure-robust at once, and it does not stand alone: one review counts thirty mainstream algorithms for reducing stochastic MEMS errors across seven families [11].

The most recent turn abandons the logic of filtering entirely: instead of smoothing the noise after the fact, a high-quality signal is newly generated out of a poor one. A diffusion-based generative learning framework synthesizes high-resolution virtual IMU data from noisy low-cost measurements [12]. The architecture is a conditional diffusion model on a U-Net basis: high-quality measurements serve as the ground-truth prior, the low-cost raw data as the conditional input [12]. The trick sits in the noise schedule — a schedule governed by the Allan variance embeds the sensor-specific error characteristics into the forward diffusion process, and the reverse process generates from it the high-precision synthetic signal [12].

The appeal lies in the interface. The generated data are a direct drop-in replacement for the raw data in any downstream navigation filter and surpass the original measurements in both position and attitude estimation [12]. With that, the boundary of the possible shifts from the production line into the code: MEMS deficits are overcome algorithmically rather than in hardware. For the drone of 2036 this means an enticing economics — navigation quality that one used to have to buy is increasingly computed.

While MEMS conquer the tactical SWaP-C market, the heavy end of the scale remains reserved for optical gyroscopes. Ring laser gyroscopes (RLG) exploit the Sagnac effect: two counter-rotating laser beams in a closed circuit experience, under rotation, a measurable frequency difference, precisely referenced to the Earth’s rotation — motionless and with minimal long-term drift [13]. A high-quality inertial navigation system (INS) on this basis is more than a sensor; it is the platform on which all the remaining methods converge. With Kalman filters it integrates terrain reference, visual odometry, and GNSS into a seamless estimate of the Earth-coordinate state — even when the platform operates in isolation for weeks, as the submarine shows [3]. This architecture is the bridge from pure inertia to the aiding methods of the following chapters.

The practical strategic variant today is often the fiber-optic gyroscope (FOG). Naval INS on a FOG basis, such as the iXblue MARINS, are proven [14], fiber technology has markedly lowered the cost of high-precision submarine navigation since the turn of the millennium [15], and fiber-optic INS are established beyond the navy as far as underground directional drilling [16]. The frontier keeps sinking: interferometric optical gyroscopes are being driven to volumes under 100 cm³, and integrated microphotonics could, in the medium term, reach MEMS size with better jamming resistance — the dividing line between tactical-MEMS and navigation-grade-optical is blurring [17]. The view toward 2036 reaches further still: cold-atom interferometers and spin gyroscopes promise navigation-grade performance at shrinking volume and have already been demonstrated as compact, deployable sensors [18], [19].

3.5 Assessment: How Far Does Pure Inertia Carry?

Here the study must be honest, for at this point the guiding thesis is decided. The self-sufficiency of inertial navigation has a physical price: the position error of pure inertial navigation is inherently unstable and grows with time — even if the sensors were perfect. Even quantum and cold-atom sensors only shrink the error sources; they do not lift the method’s fundamental instability [2]. Pure inertia drifts, in the limit, without bound. That is the inescapable flip side of any navigation without an external reference.

How fast it drifts can be captured analytically. Wheeler et al. derive closed-form expressions for the fix displacement error (FDE) of a pure three-axis gyroscope INS — modeled as Gaussian white noise plus drift as a first-order Markov process, as a function of flight duration, velocity, noise and drift amplitude, and drift time constant, validated by Monte Carlo over long flights [20]. The work also supplies a sturdy benchmark: it confirms by calculation the values a navigation-grade gyroscope must meet to satisfy the FAA Required Navigation Performance 10 (RNP-10) — a concrete anchor for how accurate a backbone must be [20]. A methodological trap is to be noted: the minimum of the Allan deviation is wrong as a drift metric, because it does not capture the drift time constant; the maximum of the Allan deviation is correct [20].

The apparent contradiction with the guiding thesis dissolves on the timescale. The unbounded drift is a statement about the long run; the combat drone’s mission is timed in seconds and minutes. The calibrated MEMS array, with only 1.85 m of deviation over a 30-second GNSS outage, shows that across short outage windows the drift is practically negligible [10] — and in precisely those windows visual and terrain reference are aiding anyway. The honest synthesis: no single sensor on its own guarantees mission success in electronic combat; the solution is the continuous stochastic calibration of interoceptive sensors — MEMS, RLG, FOG — against exteroceptive data. The backbone delivers the low-drift short-term stability, the aiding methods the long-term absolute reference [3]. Thus pure inertia does not carry forever, but it carries robustly for exactly as long as the thesis “radio loss doesn’t matter” needs it to.

This raises the question of the corrective. If the backbone runs away over the long run, something must periodically re-anchor it to the world — a sense that does not integrate but sees. The next chapter turns to this artificial eye: to visual-inertial odometry and terrain-referenced navigation, which keep resetting the inertial drift against ground and terrain to zero.

4. The Artificial Eye: Seeing Where Satellites Fall Silent

The previous chapter ended with an uncomfortable admission: an inertial platform never knows where it is, it only knows how much it has moved since its last known position—and that estimate accrues an error with every passing second that nothing can recover from the inside. Pure dead reckoning drifts, in principle, without bound. Anyone who hopes to survive a loss of radio link therefore needs a second sense, one that does not listen inward but looks outward: exteroceptive perception that periodically re-anchors the accumulated drift to the visible world. Two families carry the drone through the GNSS silence. One measures how the world shifts relative to the camera—Visual-Inertial Odometry. The other interrogates the landscape itself for an absolute position—Terrain Referenced Navigation. Together they form the artificial eye that sees where the satellites fall silent.

4.1 Visual-Inertial Odometry: Fundamentals and Physical Limits

Visual-Inertial Odometry, or VIO for short, couples two senses that prop each other up because neither suffices alone. The camera extracts salient image points—features—and tracks how they migrate from one frame to the next; from this optical flow, the camera’s own motion can be reconstructed. The inertial measurement unit (MEMS IMU) supplies, in between, the high-rate acceleration and angular-rate data that bridge the gap between two camera frames. The fusion is necessary because pure inertial navigation inevitably runs away over the integration time, driven by bias instability and time-varying errors [21].

Two algorithmic schools divide the field. The filter-based line, usually built around an Extended Kalman Filter, gets by with little memory and short computation times. The optimization-based line solves a bundle-adjustment problem over a sliding time window (sliding-window bundle adjustment), buying higher accuracy with higher cost [22]. Whichever school carries the day, what remains decisive is the quality of the IMU. Comparative analyses show that VIO benefits substantially from low-noise, tactical-grade inertial sensors, because these bridge the short-term dynamics between two frames more reliably; a consumer IMU can be heavily aided visually, yet its absolute trajectory error remains larger [21].

The limits lie in the light. Classical frame cameras capture the full image at a fixed rate and possess a dynamic range of only about 60 dB—too little for the harsh brightness jumps of a tunnel exit or backlighting [23]. Four situations make image-based VIO stumble: low light, moving objects that corrupt the optical flow, motion blur at high angular rates, and texture-poor surfaces—water, snow, desert, monochrome walls—on which simply too few features adhere [23], [21]. Against this perceptual degradation, redundancy helps: the addition of LiDAR or learning-based descriptors that still hold where simple methods fall apart under changing illumination [24].

4.2 Event Cameras: Neuromorphic Vision

If light is the frame camera’s weakness, then the event camera is the physical answer to it. It captures no full image. Each of its pixels measures the change in logarithmic light intensity asynchronously and on its own, reporting an “event”—timestamp, pixel location, sign of the change—only when a threshold is crossed [23]. From this principle follow properties that seem made for the high-speed drone: a temporal resolution in the microsecond range, a dynamic range of 140 dB against the roughly 60 dB of the frame camera, a pixel bandwidth in the kilohertz range, and with it practically no motion blur [23]. Precisely the three conditions on which classical VIO breaks down—little light, harsh contrasts, fast motion—thus lose their terror.

The algorithms follow the asynchronous nature of the sensor. Instead of discrete frame intervals, they describe motion as a continuous-time trajectory, for instance through cubic B-splines in the space of rigid-body motions, and integrate events to microsecond precision with the IMU measurements—the first event-based VIO with a continuous-time framework arose in this spirit [25]. Newer stereo systems such as ESVO2 sample contour points according to local event dynamics, use IMU pre-integration as a motion prior against tracking dropouts, and outperform five established methods on public datasets in large outdoor scenes [26]. That such pipelines can do without labeled training material is shown by self-supervised learning of optical flow, which uses the grayscale images from the same camera as a teaching signal [27]. On the exact magnitude of the efficiency gain, restraint is warranted—individual architecture names circulating in the industry and a blanket multiple of computational savings cannot be hard-sourced from primary references. The direction, by contrast, is secured: neuromorphics, which work with sparsity and precise spike timing rather than dense full images, solve perception tasks with orders-of-magnitude lower latency and energy [28], [29]. The institutional anchor lies in Zurich: the Robotics and Perception Group around Davide Scaramuzza in the orbit of UZH and ETH is regarded as globally leading in event-based perception for agile drones [23], [26].

Abbildung 2: Event camera vs. conventional frame camera: dynamic range and response latency. Source: Chapter 4.

4.3 Ultra-Low-Power VIO at the Energy Limit

For a micro-drone, it is not accuracy alone that decides, but the energy budget—the SWaP limit of size, weight, and power. Full-fledged perception has to fit into the milliwatt range, or else the eye devours the range [30]. That this can be done is demonstrated by LEVIO, a lean embedded VIO that runs on the GAP9—an ultra-low-power system-on-chip based on RISC-V with a parallel cluster architecture. The pipeline combines feature tracking, the perspective solver EPnP, and a sliding-window bundle adjustment; it stays under 100 mW and requires about 1 MB of memory, and it shifts the Pareto front of embedded VIO noticeably in favor of energy efficiency without sacrificing accuracy [30]. From the same research line comes a related idea: a purely inertial odometry that couples a model-based filter with a learning-based component and draws on thrust and rotor-speed measurements can beat classical VIO in pose estimation error—significant because it can relieve the camera at times and thereby save additional energy [31].

4.4 Aerodynamic Self-Perception (HDVIO2.0)

A drone in the wind is no rigid body gliding through still air—it is shoved, tugged, shaken by its own downwash. Classical VIO with a crudely simplified vehicle model degrades precisely here. HDVIO meets this by combining a translational point-mass model with a learning-based component that captures the complex aerodynamic effects; from the discrepancy between actual motion and the motion predicted by the hybrid model, it simultaneously estimates the external disturbance force and the full state, fed solely from the history of thrust and IMU measurements [32]. In real-world trials with a quadrotor in wind up to 25 km/h, the system improved motion estimation by up to 33% and the estimation of the external force by up to 40% over the state of the art—and that without explicitly knowing the full state [32]. The more recent development pursues the same thrust: an aerodynamics-informed, transformer-based inertial odometry improves velocity prediction by 36.9% through inclusion of rotor speed and gains a further 22.4% through the transformer architecture’s better exploitation of temporal dependencies [33].

What such methods can carry when pushed to the physical limit, the same school has shown impressively: Swift, a reinforcement-learning system trained in simulation and real-world experience that estimates velocity and position exclusively from onboard sensors, beat human world champions in real races and flew the fastest time ever recorded [34]. On the exact figures of swarm-flown formations—circulating values around 22 m/s, 7 g, and a halved accident rate—caution is in order; they cannot be attributed to any single indexed primary work and are to be understood here only as an order of magnitude. That agile autonomy is achievable on this order of magnitude, however, Swift proves hard [34].

VIO measures motion relatively and therefore drifts, more slowly than the pure inertial platform, but it drifts. What is missing is a drift-free absolute support. It lies literally beneath the wing: in the shape of the terrain itself.

4.5.1 TERCOM and Stochastic Filters

Terrain Referenced Navigation, TRN, locates the drone absolutely in the global coordinate system by measuring the topography overflown—via radar or laser altimeter, barometer, or camera—and matching the measured elevation profile against a stored digital terrain model [35]. The classical method is TERCOM: it overflies a stretch and minimizes the mean absolute distance between measured and mapped elevation over an entire sequence of fixes; the estimated position is the argument minimum of this cost function over the search space [36]. Because terrain data are strongly nonlinear and the noise is non-Gaussian, particle and point-mass filters as well as maximum-likelihood Kalman filters today take the place of simple filters [37]. In the integration of INS, GPS, and TRN, the TRN fixes calibrate the inertial system during a GPS outage; a comparison of best fix, weighted fix via a Probabilistic Data Association Filter, and single- versus multiple-hypothesis IGMAP shows that the single-hypothesis variant offers the best balance of accuracy, robustness, and computational efficiency [36].

The Achilles’ heel is the flat land. Over ocean or desert, the terrain lacks distinctiveness, and the filter can diverge. The answer is active: path-planning algorithms autonomously alter the route so that it overflies areas of high topographic gradient and maximizes the expected information gain of the measurement [35].

4.5.2 Vision-Based TRN and the Rotation-Invariant ICP

Active altimeters are tricky for small UAVs—they weigh, and they radiate, which betrays the emission signature. The passive alternative looks down with a downward-facing camera: from stereo image pairs, semi-global matching produces a dense disparity map, from this a local 3D elevation model that is matched against the database [38]. The core problem lies in the altitude—with increasing distance, the ground resolution drops, and the classical Iterative Closest Point reacts sensitively to calibration errors. The solution is a rotation-invariant ICP that feeds the filter not the raw elevation values as a measurement model, but its robust transformation output, which mitigates the divergence risk with coarsely resolved maps [38]. The method was validated in real Cessna flights at 3,000, 4,000, and 5,000 m altitude—with nothing but an IMU, a barometer, and a FLIR Blackfly S camera; the rotation-invariant ICP improved positional accuracy consistently at all altitudes, most markedly under the difficult conditions at 5,000 m, and confirmed vision-based TRN as a passive, GPS-independent alternative for cruise missiles and strategic UAVs [38].

4.5.3 Gimbaled-Laser TRN and Information Maximization

Whoever points the laser rigidly into nadir forfeits information. A gimbaled, steerable laser can scan at deliberate angles and aim predictively at topographically informative landmarks far ahead of the aircraft [35]. The choice of pointing angle can be optimized in information-theoretic terms: via the Cramér-Rao Lower Bound, the angle is steered so that it maximizes the trace of the Fisher information matrix. The result is no marginal correction—against nadir or fixed configurations, navigation performance improves by a full order of magnitude, and mislocalizations arising from inadequate modeling of the highly nonlinear terrain matching disappear [35].

4.5.4 Maritime/Bathymetric TRN

Underwater, the same principle holds, only the map is the seafloor. A submarine or AUV that may not surface for a GNSS fix surveys a wide swath of the bottom with multibeam echo sounders and matches it against the bathymetric map. Trials by the Norwegian Defence Research Establishment with the submarine KNM Utsira demonstrated high-precision submerged position updates through terrain-referenced bathymetric navigation—without pushing a mast to the hostilely monitored surface [39]. Independent reviews confirm the methodology: terrain-relative navigation without acoustic beacons works, yet its accuracy depends on the bathymetric variability—precise over strongly structured bottom, weaker in flat zones [40]. It is the same flat-terrain problem as on land, only mirrored into the depths.

Thus arises the architecture of the artificial eye: VIO carries the high-precision relative motion, TRN the drift-free absolute support over structured terrain, the tactical IMU the backbone in between—a layered perception that outlasts the GNSS outage. Yet a single platform navigating in this way is only half the answer. The strength of the formation lies in sharing what each eye sees. How these autonomously navigating drones network covertly into a common situational picture—via mesh radio and phased array—is the subject of the following chapter.

5. The Nervous System of the Swarm: Mesh Radio and Phased Array

So far the single platform has stood at the center: a flying body that, with layered self-navigation, finds its way even when the satellite signal goes dark. A single drone, however, wins no war. The real threat of 2036 is not the device but the formation—dozens, perhaps hundreds, of autonomous platforms that see a target together, push sensor data to one another, and divide the work among themselves as though they were a single organism. For loose flying bodies to become a swarm, a nervous system is needed: a radio network that organizes itself, that withstands the constant shift of position in the air, and that operates covertly enough not to invite the adversary to react. This chapter maps that nervous system—from topology in motion through learning routing to the physical invisibility of the directional radio beams that hold the swarm together.

5.1 Flying Ad-Hoc Networks (FANET): Topology in Motion

A Flying Ad-Hoc Network (FANET) is a self-organizing, infrastructure-free radio network among small unmanned aerial vehicles—flexible, inexpensive, and quick to deploy [41]. Formally it is a subclass of the Mobile Ad-Hoc Network (MANET), but the kinship deceives. Where a MANET moves nodes across a surface and a Vehicular Ad-Hoc Network (VANET) guides vehicles along roads, the nodes of a FANET fly in three dimensions and fast. The consequence is a topology that reconfigures more rapidly than anything terrestrial networks know; routing protocols designed for MANETs sometimes fail even to follow the changes [41].

Four stresses converge and make static or purely proactive protocols collapse: the high three-dimensional mobility of the nodes, the extremely fast topology changes, strong Doppler shifts of the radio signals, and frequent, intermittent link interruptions [42]. That these peculiarities are not merely theoretical is shown by a look at trained network models: intrusion-detection and traffic models calibrated on wired benchmarks break down in a real UAV swarm because mobility, fluctuating link quality, and decentralized routing reshape the traffic distributions—quantified on the calibrated digital-twin dataset UAV-CAS with 99,492 flows from 1,024 configurations across five attack families [43].

The research response begins by telling the network its own motion. In real flight trials with two autonomous fixed-wing drones and a ground node, the FANET-specific Predictive OLSR (P-OLSR), which exploits the onboard GPS information, clearly outperforms standard OLSR in routing under frequent topology changes [41]. At the time of publication, P-OLSR was the only FANET-specific routing technique with an available Linux implementation—a sober indication of how young this field is.

5.2 Intelligent Routing

When the topology changes faster than any routing table can be maintained, no rigid address book helps anymore. The Stochastic Packet Forwarding Algorithm (SPA) draws the consequence and forwards packets on the basis of probabilistic distribution models rather than relying on fixed tables—a direct answer to the fast topology dynamics of the FANET [42]. This geo-routing is theoretically grounded in the random 1/2-disk scheme: each node selects the next relay node at random among those within transmission range that lie roughly in the direction of the destination; if the distance evolution is modeled as a Markov process, convergence conditions and bounds for the expected hop count can be derived [44].

In the military and cognitive-radio domain, the decision becomes multidimensional. Routing protocols for Cognitive Radio UAVs rely on the Central Node Resolution Factor (CNRF), which weighs four quantities at once: distance, velocity, link quality (RSSI), and the drones’ residual energy. From this weighting the network selects optimal cluster heads and integrates predictive algorithms to avoid link failures before they occur [45]. Residual energy as a coequal criterion is more than pedantry here—in a swarm meant to fly for hours, it decides which platform can still serve as a relay at all.

The next step is taken by learning routing. Instead of following heuristic rules, reinforcement learning lets the network learn adaptive decisions through ongoing interaction with the environment, decisions that minimize latency and secure stability; formally, the routing decisions are modeled as a Markov Decision Process (MDP) [46]. Survey works finally place SPA, CNRF/cluster-head, and RL/MDP approaches into a common taxonomy and compare them by topology dynamics, node failure, and link stability [47], [48].

5.3 Programmable Data Plane and Predictive Self-Healing

A swarm that notices failures only when packets are lost reacts too late. The research therefore relocates network management into the data plane itself. With the P4 language (Data Plane Programmability), the UAVs’ switches can be instructed to drastically expand their standard telemetry; P4 decouples packet processing from hardwired vendor functions and allows programmable, granular visibility into packet events—evaluated across more than 200 P4 works in a single survey [49].

The P4 FANET In-Band Telemetry (FINT) framework turns this programmability into a diagnostic system. Data packets collect network-specific metadata in real time as they transit the swarm: the Received Signal Strength Indicator (RSSI) of the radio links, the percentage CPU load of the flight computers, and the precise three-dimensional geolocation data of every single platform [50]. This telemetry flows directly into an AI system that recognizes patterns indicating imminent link failures or overload and reroutes preemptively before the connection breaks—Proactive Link Failure Forecasting that minimizes packet loss and communication breakdowns in the swarm [50]. The nervous system feels the pain before the bone breaks.

This visibility comes at a price. Classical in-band telemetry generates a transmission overhead that grows linearly with the hop count and the number of telemetry values—a serious problem for energy- and bandwidth-constrained swarms. Lightweight methods mitigate this: DLINT works deterministically, PLINT probabilistically via reservoir sampling, and Bloom filters compress the state tables without sacrificing monitoring accuracy [51]. Foresight and frugality stand here in a negotiated relationship that every real system must recalibrate anew.

5.4 Directional Radio in the Swarm

Omnidirectional antennas are an own goal in a dense swarm. They sow interference, burn energy, and drastically curtail spatial frequency reuse; directional antennas solve this but demand new medium-access protocols [52]. The crux carries a telling name: the deafness problem. A node does not receive because its antenna is turned away just then—deaf to precisely the neighbor calling it. The Location Oriented Directional MAC (LODMAC) uses the position data of neighboring drones from GPS or Visual-Inertial Odometry to aim antennas precisely at one another, thereby enabling collision-free, heavily parallelized communication [53].

The actual leap comes from the multi-beam antenna. A node that transmits simultaneously on several non-interfering beams achieves up to an m-fold throughput compared to a single-beam antenna; asynchronous packet arrival, however, aggravates deafness and MAC-layer capture, which is why a window mechanism establishes concurrent communication even with unsynchronized arrival [54]. Multi-beam antennas increase frequency reuse, extend range, improve reliability, and lower energy consumption—capabilities that the classical IEEE 802.11 DCF mechanism cannot exploit at all, which is why specialized protocols are necessary [55].

What this shift means can be quantified. The transition from single-beam to multi-beam protocols such as the Multiple-Beam Antenna Array MAC (MBAA-MAC), orchestrated with optimized routing, lowers the end-to-end latency in mobile multiframe scenarios from 700 ms to 9 ms [56]—nearly two orders of magnitude, the difference between a swarm that hesitates and one that reacts. The outlook to 2036 goes further still: DeepBeam infers the angle of arrival and the transmit beam in use from ongoing transmissions via deep learning—without pilot sequences, beam sweeping, or synchronization, with up to 96% accuracy on a five-beam codebook and up to a sevenfold reduced latency compared to the initial beam sweep of 5G NR [57]. With that, the self-organizing directional-radio swarm moves from theory into reach.

Abbildung 3: Latency gain from multi-beam directional links in the swarm. Source: Chapter 5.

5.5 Phased-Array Antennas: Electronic Beam Steering in the Ku Band

Whoever wants to redirect a directional beam in microseconds cannot rotate an antenna mechanically. The phased-array antenna solves this without a single moving part. It consists of a planar arrangement of many radiating elements; under computer control it varies the phase of the radio-frequency signal at each element so that the wavefronts bundle, through constructive and destructive interference, into a narrow main beam that is steered electronically in fractions of a microsecond [58]. For networking within the swarm, this happens in the Ku band (12 to 18 GHz): the high frequencies carry the bandwidth that sensor data fusion demands, and through pronounced directivity and atmospheric attenuation they offer inherent protection against distant interception [59].

The central design conflict is called mutual coupling. The mutual coupling of neighboring elements shifts terminal impedances and reflection coefficients and thereby affects radiation characteristics, SINR, and radar cross section; it directly limits resolution, interference suppression, and direction estimation [60]. The conflict grows especially sharp once the beam is meant to scan widely. Wide-angle beam-scanning arrays open up broad spatial coverage, critical frequency reuse, and higher system capacity, but they struggle with precisely this strong coupling and the narrow element beamwidth [61]—exactly the properties that an intra-flight array integrated into the chassis and steerable in all directions requires. That such arrays can also be field-maintainable is shown by the modular ferrite concept: interchangeable building blocks with their own driver circuitry and collimation memory that can be swapped at the depot level without recalibrating the entire array [62].

5.6 MADL: Fifth-Generation Stealth Networking

What a covert mesh looks like in practice is shown by Northrop Grumman’s Multifunction Advanced Data Link (MADL), conceived for fifth-generation fighter jets such as the F-35 Lightning II and the B-2 Spirit and increasingly envisioned as a command node for collaborative stealth drones [63]. The platform attribution itself stems from a non-academic source; the physics behind it, however, is solidly documented.

The design principle is called Low Probability of Intercept and Low Probability of Detection (LPI/LPD): an extremely narrow directional beam minimizes the stray radiation, so that the energy strikes almost exclusively the intended receiver. This is academically grounded by an electronically reconfigurable antenna with rapid sidelobe time modulation, which holds the main lobe stationary and rapidly modulates the sidelobes in time—the signal radiated in unwanted directions thereby undergoes a spread-spectrum distortion and becomes harder to detect and intercept [64]. Against electronic warfare, MADL combines spread-spectrum technique with extremely fast frequency hopping across the Ku band, which smears the signal so broadly that it vanishes into the thermal background noise; that direct-sequence and frequency-hopped spread spectrum possess precisely this LPI property belongs to the secured body of knowledge [65], [66]. The seamless integration of the antennas into the chassis allows up to eight stealth platforms to form a covert mesh in which high-resolution radar data, electro-optical target imagery, and sensor metrics are fused in real time into a synchronized situational picture—without revealing one’s own position [63]. Adaptive transmit power control makes the signal only as strong as the receiver needs it, which further complicates passive localization [66].

5.7 The Physical Invisibility

In the end the engagement shifts into physics itself, and there the eavesdropper stands at a disadvantage from the outset. The inverse-square law is relentless: the power density of a signal decreases inversely with the square of the distance. A hostile passive receiver standing farther away than the intended receiver sees the signal weaker by precisely this square—a disadvantage that no electronics, however sensitive, can compute away.

That leaves the search below the noise floor. A ground-based system would have to capture signals weaker than the natural electromagnetic background radiation. Methods such as Cyclostationary Feature Detection exploit for this the cyclostationary, that is periodic, features of modulated signals to separate them from the noise at very low signal-to-noise ratios—at a computational load similar to mere energy detection, but more robust against noise uncertainty [67], [68]. Theoretically this reaches far: under the Neyman-Pearson criterion, signals can still be reliably detected at a signal-to-noise ratio down to −20 dB [69]. The punch line is in the fine print. This limit holds only with a priori knowledge of the spectral features and a long observation time—and precisely both are denied by an LPI waveform with fast frequency hopping and adaptive power. Whoever does not know the features and is given no time to look never reaches the −20 dB. The detection limit is real, yet the key to it lies with the transmitter, who does not hand it over.

5.8 Swarm Beamforming: Collective Range

What a single antenna cannot achieve, the swarm achieves together. Drones in three-dimensional formation can cohere their signals collectively in one direction—swarm beamforming, which extends the range far beyond the possibilities of individual antennas. The formation then acts as a single Virtual Antenna Array, formed from many distributed transmitters that attune their wavefronts to one another like rowers their stroke [70].

The same coherence that creates range creates security. Cooperative beamforming over a distributively formed Virtual Antenna Array yields a secrecy rate derivable in closed form; with artificial-noise injection, the method achieves up to twice the secrecy rate compared to conventional cooperative beamforming—precisely when an eavesdropper stands close to the intended receiver [70]. The swarm thus aims its useful signal precisely at the receiver while simultaneously scattering artificial noise in those directions where an interceptor might lurk. With channel state information available, selected platforms can relay the source message via distributed beamforming, where closed-form expressions for outage probability, goodput, and total transmit power quantify the gain over purely terrestrial transmission [71].

With that the picture of the nervous system closes: a network that holds itself together despite 3D mobility, Doppler, and link dropouts, that monitors its own health predictively, that drives latency down by nearly two orders of magnitude, and that steers its beams so that the adversary can hardly find them and, should it find them after all, can hardly decrypt them. This nervous system, however, has a prerequisite it does not itself control—orientation in space. Phased-array alignment, LODMAC, FINT geolocation, and swarm beamforming all presuppose that every platform knows where it is and where its neighbors are. This is exactly where the adversary attacks. The next chapter is devoted to the GNSS war: the electronic shield of jamming and spoofing that aims to wrest from the swarm the spatiotemporal reference on which its entire nervous system depends.

6. The Electronic Shield: GNSS Under Siege

The layered self-navigation of the 2036 combat drone is only as strong as its most vulnerable link — and that link carries its weakness in the physics. Where the onboard sensors of Chapters 4 and 5 draw on camera light and on the inertia of their own mass, that is, on the immediate surroundings, the absolute anchoring in space hangs on a signal that has traveled some 20,000 km before it reaches the antenna. Along that path it loses almost everything. What remains is a whisper beneath the noise — and a whisper can be drowned out, mimicked, forged. This chapter is about the satellite signal under siege and about the multilayered defense with which an autonomous platform holds its ground within it.

6.1 Attack Vectors: Jamming, Spoofing, Meaconing

The vulnerability begins with a sober energy budget. The civilian GNSS signal is broadcast openly and without cryptographic integrity protection from an altitude of roughly 20,000 km; at the Earth’s surface its received power is on the order of about −160 dBW — weaker than the thermal noise that every receiver carries with it in any case [72]. From this follows the structural susceptibility: a low-power ground-level transmitter easily outshines the distant satellite, all the more so because the spreading sequence of the open-service code is publicly documented, and thus predictable and reproducible, while cryptographic protection has only been ramping up with Galileo OSNMA since 2023 [73].

The first attack vector is the most brutal. In Jamming, the attacker floods the GNSS bands — L1 at 1575.42 MHz, L2 at 1227.60 MHz, L5 at 1176.45 MHz — with high-energy broadband noise; the carrier-to-noise-density ratio (C/N0) collapses, satellite tracking is lost, and the navigation solution goes dark: a classic denial-of-service [72].

Subtler and more dangerous is Spoofing. The attacker generates artificial, highly structured GNSS signals and transmits them at slightly higher power than the real satellites; the receiver locks onto the false signal and computes a dictated PVT solution of position, velocity, and time. This is the most critical vector, because it allows undetected course deviations and the hijacking of autonomous systems [74]. The most advanced variant, the Seamless Takeover, synchronizes the false signal in precise phase with the genuine one and, once the phase has been accepted, steers it away slowly and minimally — which defeats the Extended Kalman Filter safeguards of commercial off-the-shelf drones [75].

A simpler version is Meaconing: authentic signals are received, delayed, and rebroadcast elsewhere — a mere replay that succeeds with simple hardware repeaters [76]. At the upper end of the scale stands Secure Code Estimation and Replay (SCER), in which the attacker estimates and replicates the secret code components in real time and thereby challenges even future protective systems such as Navigation Message Authentication and Spreading Code Authentication [77].

6.2 The Spoofing Epidemic

What was long regarded as a state capability has become a commodity. With freely available software-defined-radio hardware — a HackRF platform runs, as an order of magnitude, under 100 euros — GNSS spoofing is no longer an exclusive capability of state actors, and the review literature explicitly names this democratization as a driver of the threat [76], [74]. The consequences are measurable: research documents an exponential increase in massive spoofing incidents in civil aviation, clustered over conflict zones such as the Black Sea and the Middle East [78].

The most effective countermeasure arose not in the drone but in the sky above it. In crowdsourcing via ADS-B, aircraft transmit their GNSS position together with signal-quality metrics such as the Navigation Integrity Category (NIC), which a ground-based infrastructure continuously evaluates. The Crowd-GPS-Sec system detects spoofing globally in less than two seconds and localizes the attacker after 15 minutes to within 150 m — without any change to infrastructure or onboard receivers [79]. Algorithms scan the data for abrupt position jumps, and a Kalman filter checks the derived velocities; if a passenger aircraft mathematically exceeds 650 m/s — as a rule of thumb, about Mach 2 — the point is deemed spoofed, and the intersection of the overlapping radar horizons also pinpoints the jammer [80]. The same principle of cross-verification carries beyond aviation: for connected vehicles, detection succeeds over WiFi in roughly six seconds and over cellular in roughly thirty, at a false-positive rate below 0.01 [81].

6.3 Machine Learning for Interference Detection

The step from the rigid threshold to the learning receiver shifts the defense from heuristics to statistics. The features remain the same — the C/N0 ratio and the behavior of the preamplifier’s Automatic Gain Control — but now models evaluate them in context: Support Vector Machines, for instance, read the signal entropy and distinguish a natural signal blockage by a building from targeted jamming [72]. More significant for the flying platform is that this intelligence fits onto embedded hardware. Long Short-Term Memory networks can be trained on Raspberry Pi platforms and u-blox EVK-F9P sensors; they learn the temporal dependencies of normal observation codes in the RINEX format (Receiver Independent Exchange Format) and, when anomalies arise, selectively discard individual frequencies rather than sacrificing the entire navigation solution [78].

The depth of the models drives accuracy to the limit of what can be measured — most telling precisely where the attack turns subtle and supervised learning fails.

PerDet combines features from accelerometer, gyroscope, magnetometer, GPS, and barometer: interoceptive redundancy exposes the false signal [84]. Where the algorithmics reach their limits, the hardware hardens further — Controlled-Reception-Pattern antennas place adaptive nulls into the direction of interference via null-steering and, through cyclostationary methods, simultaneously estimate that direction of arrival [85].

6.4 SemperFi: Active Signal Recovery

All the methods so far share a quiet assumption: they detect the attack in order to avoid it — the spoofed signal is discarded, and the platform falls back on its inertial navigation. SemperFi, developed at Northeastern University and ETH Zurich, breaks with this logic and recovers the genuine signal. Two components interlock, both designed for a single-antenna receiver and implemented in the open-source software GNSS-SDR.

The Adversarial Peak Identifier turns the drone’s body into an instrument of verification. In a suspicious signal environment, SemperFi forces the platform into a brief, highly dynamic flight maneuver. Because the false signal originates from a single ground-based transmitter, its phase changes differently from that of the spatially distributed real satellites. SemperFi exploits the excellent short-term stability of the onboard IMU and cross-correlates the physically experienced acceleration vectors of the maneuver with those reported by the spoofed GPS; a divergence identifies the Adversarial Peak beyond doubt and counters precisely the Seamless Takeover [75].

Rather than discarding the exposed signal, SemperFi activates the Legitimate Signal Retriever. It relies on Successive Interference Cancellation: it computes an exact, inverted replica of the false signal and adds it to the antenna signal, so that the interfering wave cancels itself mathematically and the extremely weak genuine satellite signal beneath it reemerges, along with its Time-of-Arrival [75]. In the canonical NDSS evaluation, flight patterns of under 100 m suffice to identify the Adversarial Peak, and the true position is reconstructed within 0.54 s on a Jetson Xavier — secure against naive as well as stealthy spoofers [75]. An earlier preprint version cited diverging values (under 50 m, around 10 s), which reflect different levels of maturity and must not be conflated [86].

With this, SemperFi accomplishes the paradigm shift that this chapter sums up: from passive detection to active, autonomous signal recovery in real time. The electronic shield is no longer a mere filter that rejects the false, but an instrument that restores the truth. Yet this shield does not orchestrate itself. Which maneuver is forced and when, how a detector’s output is weighted, which sensor source is trusted in case of conflict — these decisions fall in split seconds and without radio contact to the ground station. They demand a brain capable of judgment on board. To it the next chapter is devoted.

7. The Brain on Board: Local AI by 2036

The preceding chapters took the drone apart into its senses and its voice: the layered eye of the sensor suite, the concealed phased-array mesh, the defenses against jamming and deception. But an eye does not see, and an antenna does not decide. Above all of this there must reside an authority that interprets the stream of perception, weighs it against the mission objective, and commands, in every millisecond, what happens next. This chapter describes that authority—the brain on board—and closes Part I. If the guiding thesis holds, that the drone of 2036 carries its mission to completion even without radio contact, then that thesis rests not on sensors and antennas but on the question of whether cognition can be compressed into the wattage of an aircraft.

7.1 Cognitive Edge Computing: The Break with the Cloud

For decades, progress ran in one direction: data travels to the data center, the data center thinks, and the result returns. For a drone over contested territory, that is a death sentence. Three forces compel the reversal. Latency—a decision routed over radio and cloud arrives too late. Data sovereignty—reconnaissance imagery that leaves the device can be compromised. And, sharpest of all in the military context, resilience against electronic warfare: onboard inference makes the aircraft independent of the radio link, capable of acting under jamming and loss of connection [87]. What long seemed an economic argument becomes, here, an operational and tactical one.

The shift has a target figure. To sustain the progress of the coming decade, a thousandfold increase in overall efficiency—intelligence per joule—is held to be necessary; it is not a measurement but a consortium industry forecast (confidence C), and it decomposes multiplicatively into a factor of ten from algorithms, twenty from silicon, and five from system architecture [88]. Where the path leads is shown by a more honest diagnosis: in local inference of large language models, as much as ninety percent of the energy goes not to computing but to moving data [87]. This bottleneck is attacked on three fronts—silicon, memory, algorithms.

7.2 Semiconductor Evolution

The first front is the substrate. The transition from the FinFET to the Nanosheet transistor with a Gate-all-around structure opens up new logic density; the 2-nanometer node (N2) forms the foundation of the A-class, followed by A14 at one and a half and A10 at one nanometer in the early 2030s. At the limits of channel length, the architecture shifts to the Complementary FET (CFET), which stacks n-type and p-type transistors vertically; beyond that, channels made of two-dimensional materials below ten nanometers begin to take shape. All of these nodes and dates come from manufacturer roadmaps and are explicitly an industry forecast (confidence C) [89], [90].

The real revolution lies in the third dimension. Where the monolithic chip reaches its limit, heterogeneous 3D integration and Chiplets take its place: specialized tiles of logic, compute, and memory, fused by Hybrid Bonding and packaging technology such as CoWoS and SoW-X. The roadmaps hold out the prospect of more than two hundred billion transistors monolithically by 2030, and multi-chiplet packages with more than one trillion transistors by the end of the decade—throughout an industry forecast (confidence C) [89], [91]. For the drone, this yields a compact system-on-chip of NPU chiplets, CPU cores, and memory—the physical basis for a brain in a palm-sized form factor.

7.3 Overcoming the Von Neumann Bottleneck

The fastest silicon is of little use as long as data must traverse the long road between memory and compute. An access to external DRAM costs, as an order of magnitude, about 640 pJ per 32-bit word, a Multiply-Accumulate operation roughly 0.9 pJ—so the transfer is about two hundred times more expensive than the computation itself [92]. The asymmetry matches the order of magnitude established since Horowitz’s ISSCC analysis: classical architectures pay the lion’s share of their energy to remember, not to think.

Three answers interlock. The Unified Memory Architecture lets CPU, GPU, and NPU share a common memory pool instead of shoveling weights across a slow bus. More radical is Processing-in-Memory: the compute unit migrates into the memory array, and the vector-matrix multiplication takes place in analog form within the Crossbar—multiplication by Ohm’s law, summation by Kirchhoff’s node rule, in parallel within a single clock cycle [93]. The third answer is MRAM: nonvolatile, standby current zero. From this follows a property valuable for a weapon system—instant-on: the weights remain in memory, and the brain wakes without load time, in milliseconds.

Sobriety is called for here. Industry reports cite, for analog in-memory chips, peak values above 2,000 TOPS/W versus 100 to 300 for digital high-end GPUs (confidence C) [94]—values that hold for the array, not for the system. Parasitic resistances in the Crossbar distort the ideal weighted sum and demand careful mapping [95]; above all, the analog-to-digital conversion at the periphery devours a hidden share of the energy—targeted pruning can lower the ADC energy by up to 7.13× [96]. The system efficiency therefore falls below the potential of the core. This honesty protects the argument.

7.4 Algorithmic Metamorphosis

Silicon and memory furnish the stage. The greatest lever—the factor of ten from the roadmap—lies in the algorithm. Four movements reshape thinking so that it fits into an aircraft.

7.4.1 1-Bit and Ternary Language Models (BitNet b1.58)

BitNet b1.58 reduces every weight parameter to three states—minus one, zero, plus one, log₂(3) ≈ 1.58 bits, hence the name [97]. With that, a multiplication is no multiplication at all but a subtraction, a null operation, or an addition: the energy-hungry floating-point multipliers, the heaviest component of any AI accelerator, become superfluous, and what remains are integer additions. The leap is not bought at the cost of accuracy—the model attains, in perplexity and downstream performance, the level of a full-precision Transformer of equal size, and defines a scaling law of its own [97]. With BitNet b1.58 2B4T, the first open-source native 1-bit model at production scale is now at hand—two billion parameters, four trillion tokens, on a par with the leading full-precision models in its class while requiring markedly less memory, energy, and decoding latency [98]. The engine bitnet.cpp demonstrates the gain losslessly: 2.37× to 6.17× on x86, 1.37× to 5.07× on ARM [99]. On dedicated ASICs, ternary billion-parameter inference moves below seven watts—within the budget of an aircraft (confidence C).

7.4.2 State Space Models and Mamba

Mamba replaces the attention mechanism with selective State Space Models with input-dependent parameters and a hardware-aware parallel scan—linear in the sequence length, roughly fivefold throughput compared with a Transformer of equal size [100]. For the edge, one property counts most: the inference memory stays constant, no matter how long the mission and the context grow. No KV cache that grows along with it, and therefore no memory overflows after hours in the field [100]. The limit: on exact copying and chain-of-thought tasks, a Mamba of constant size falters, and the advantage fades the moment one forces it to grow along [101]. The gold standard is therefore hybrids—Mamba for the linear stream of perception, selective attention where precise remembering counts.

7.4.3 Liquid Neural Networks

Liquid Neural Networks are, inspired by the nervous system of the roundworm C. elegans, a continuous subclass of recurrent networks: their time constant depends on the input—with dense information the system accelerates its internal dynamics, with sparse information it slows down [102]. Their state evolves as an ordinary differential equation, which long made their inference expensive. The breakthrough is the closed form: the Closed-form Continuous-time model replaces the numerical ODE solution with an analytical approximation and becomes one to five orders of magnitude faster in training and inference—peer-reviewed in Nature Machine Intelligence [103]. It is precisely this family that solves vision-based drone navigation from raw pixels robustly, where established recurrent models fail—causal control over short and long range, tracking of static as well as moving objects [104]. Add to this an extraordinary robustness against distribution shift, and that at a fraction of the parameters—predestined for TinyML and mission-critical embedded systems [102].

7.4.4 Dynamic Context Pruning and “Sifting Agents”

The quietest movement keeps the system mentally clean. Irrelevant data acts like cognitive noise—it lowers the density of intelligence and provokes hallucination and reasoning drift. Against this, extremely fast 1-bit filter models can be placed upstream, so-called Sifting Agents, which test every fragment for relevance before it reaches the reasoning model; in addition, methods compress the KV cache down to a few bits. These values are secondary, to be taken as concept and order of magnitude [87]. Their use for the drone is nonetheless clear: an upstream filter saves energy and latency where both are scarce.

7.5 Neuromorphics and Silicon Photonics

Beyond the algorithms there waits a hardware that rethinks computing itself. Neuromorphic circuits break with the global system clock: Spiking Neural Networks operate event-driven and in parallel—a neuron fires and consumes energy only when information actually changes, just as the brain does with its roughly twenty watts [93]. The hardest quantitative evidence comes from Intel’s Loihi 2: a co-designed SNN achieves, in online continual learning, 113× lower latency and 6,600× lower energy than the strongest edge-GPU baseline—0.05 mJ instead of 333 mJ per inference, one part from the algorithms, the lion’s share from the neuromorphic hardware co-design [105]. Equally decisive is the sensing: an event-based camera transmits only the pixels with a change in contrast—in static scenes often less than a tenth of the image—which pushes consumption down into the micro- to milliwatt range. Loihi has shown its practical viability, from gesture recognition with a DVS camera to the locomotion of a six-legged robot [106], [107]. Neuromorphics and Liquid Neural Networks share the same gift—asynchronous, sparse time series at a minimal thermal budget.

Silicon photonics belongs in its honest place. Where copper loses energy as waste heat at high data rates, light transmits a bit for only 0.05 to 0.2 pJ—values from industry, plausible in their order of magnitude, no peer review (confidence C) [108]. Its use today lies primarily inside the package, in the bandwidth between chiplets and memory. For the aircraft of 2036, the caveat holds unmistakably: a standalone edge photonics inside the drone is not yet a production reality—a promise for the next decade, not a component of this one.

Abbildung 4: Energy per inference: neuromorphic computing vs. edge GPU (log scale). Source: Chapter 7.

7.6 The AI Ecosystem of the Autonomous Platform in 2036

Put the strands together, and what emerges is no monolith that knows everything, but an ecosystem of heterogeneous edge agents, each tailored to its task and fused with its hardware. A Mamba-based language model carries the higher-level reasoning at constant memory [100]. Liquid Neural Networks interpret the sensor stream in real time and conduct the causal navigation—the only directly documented path to vision-based control of an aircraft [104]. Upstream 1.58-bit Sifting Agents keep the context clean and the watt budget lean [97]. Above it all, the weights remain nonvolatile in MRAM and in the memory itself—instant-on, matrix computation by Ohm’s law in the array [93].

The load-bearing figures bundle into a coherent watt budget. Ternary billion-parameter inference runs below seven watts (confidence C); a neuromorphic SNN consumes 0.05 mJ per inference and is thus 6,600× more frugal than an edge GPU [105]; the event-based perception moves in the micro- to milliwatt range; the closed form of the Liquid networks saves one to five orders of magnitude in compute time per decision [103]. From this follows the argument of this chapter: cognition in the single-digit to low double-digit watt range is, in 2036, no speculation but a documented constellation. With it, data sovereignty stands secure—the sensitive data does not leave the device, the cloud remains at most a training backbone, and the inference is entirely on board. This is precisely what makes the drone capable of acting under radio jamming, and it closes the effect chain that this study asserts.

The brain is thus built. What it has not yet done is act. An authority that sees, decides, and navigates within the watt range becomes a weapon system only when it names what it has perceived: friend from foe, target from camouflage, human from machine. Part II therefore begins where this chapter ends—with the first and most consequential application of onboard thinking: automatic target recognition.

8. Automatic Target Recognition

Part I took the measure of the building blocks: an inertial backbone, an artificial eye, a transmitting nervous system, an electronic shield, and a brain aboard. Part II assembles them into the capability this study places at its center. We begin with the faculty that decides whether a flying machine remains a remotely piloted tool or becomes an effector that acts on its own: the ability to recognize a target by itself.

8.1 From Detection to Classification: Sensor Fusion Meets Artificial Intelligence

Automatic target recognition — ATR in the technical idiom — denotes the machine-driven detection, localization, classification, and identification of military targets from sensor data. It is not a solved problem. For the radar-based variant, no closed theoretical framework yet exists that would fully connect sensor, target, environment, and signal processing; precisely for that reason, deep machine learning counts as the practical solution that has prevailed [109]. The generic processing chain — from detection through feature representation and object proposal to classification — has been fundamentally overturned by learning methods; a survey of more than three hundred works distinguishes one-stage and two-stage detectors along the way [110].

The drone of 2036 recognizes not through a single sensor but through their combination. A multilayered chain pairs radar for the long-range, weather-independent first detection with infrared sensors for the thermal signature and a high-resolution optical fine-tracker for precise pursuit. To this is added LiDAR as a three-dimensional sensory organ: in field trials, the LiSWARM research system captured swarms of 150 to 500 aircraft with 98 percent accuracy and tracked individual trajectories in space — the precondition for assigning effectors to individual targets at all [111]. Through the mesh described in Chapter 5, these sensor images from several platforms can be fused into a shared picture without any single one of them betraying its position.

8.2 Onboard Target Recognition Without a Data Link

The decisive step lies in performing this recognition aboard, with no radio link to an evaluation station. In contested environments this is not a convenience but a condition: a drone forced to first transmit its sensor data to a ground station would go blind the moment the adversary smothered the radio. This is exactly where the brain from Chapter 7 earns its keep. Terrain-referenced perception runs on event-based cameras and a RISC-V system-on-chip under one hundred milliwatts; the cognitive evaluation rests on architectures built for an aircraft’s power budget — ternary language models that eliminate floating-point multiplication, state-space models with constant memory footprint, and continuous-time networks whose closed form speeds inference by orders of magnitude [105]. That such continuous-time networks robustly solve vision-based drone navigation where conventional recurrent models fail is not speculation but documented fact [104].

Image recognition itself rests on deep neural networks that learn discriminative features directly from training data instead of having them specified by hand — on radar imagery as well as on electro-optical and infrared material [112], [113]. Their Achilles’ heel is data scarcity: measured target data are expensive and rare, which is why transfer learning, synthetic data generation, and few-shot methods dominate the field [114].

8.3 Cooperative Target Assignment in the Network

A swarm does not merely recognize; it distributes. The problem of which member engages which target — weapon-target assignment in the technical vocabulary — can be cast as a Markov decision process and solved with reinforcement learning methods that surpass classical optimizers in adaptability and computational efficiency [115]. Extended multi-agent methods couple target assignment and path planning into collision-free three-dimensional trajectories in dynamic environments and scale with swarm size [116]. That cooperating, sensor-fusing drones can outperform professional human pilots in pure flight control is established in research on autonomous racing — at speeds and accelerations no human can master any longer [34]. The peak values reported for military swarms — upward of twenty-two meters per second and as much as seven g — should be read as an order of magnitude, not as a measured constant; the direction, however, is beyond question.

8.4 The Decision to Strike: Levels of Autonomy

Recognition is not decision. Between the moment a system classifies a target and the moment it applies force lies a threshold that can be captured technically in three levels. With human-in-the-loop, a human authorizes each individual weapon engagement. With human-on-the-loop, the human supervises and can abort. With human-out-of-the-loop, the system selects and engages its targets without intervention [1]. The driving force that pushes the human out of this loop is time: whoever compresses the kill chain from hours to seconds inevitably shifts the decision from human reaction speed to the machine [117]. Therein lies precisely the danger that nominal oversight “on the loop” degenerates, when the chain is compressed, into a mere act of confirmation — a finding Chapter 14 takes up.

Yet before one entrusts the decision to the machine, one must be able to trust its recognition — and this is exactly where the sober limit lies. Even slightly manipulated inputs, indeed perturbation patterns physically introduced into the world, reliably produce misclassifications [118]; a pasted-on trigger can deceive a classifier on purpose, and such backdoors sometimes survive retraining [119]. More serious still is the domain gap: a model trained on synthetic data does not generalize readily to measured data — the distribution of operational reality is a different one from that of training [120]. These two weaknesses, physical deceivability and the domain gap, are the technical substance of the legal problems of predictability and distinction. A drone that chooses its target autonomously is only as reliable as the weakest point of its perception — and that point can be attacked. How deeply the platform that recognizes and decides in this way is secured against the loss of its external supports is shown by the chapter that follows.

9. The Fail-Safe Kill Chain

This is the chapter the study has been driving toward. It draws together what the navigation chapters unfolded one by one, and it tests the claim that gives the entire investigation its name: that the combat drone of 2036 is built so that the loss of its links does not stop it.

9.1 The Principle of Layered Resilience

Resilience here arises not from one indestructible sensor but from a ladder of methods down which the drone climbs as the electromagnetic environment grows more hostile. On the topmost rung it navigates like any civilian device: by the satellite signal. If that signal is jammed or spoofed, the positioning does not collapse at once; an active defense engages instead — a receiver that exposes the false signal through a forced flight maneuver and then mathematically cancels it to lay the genuine signal beneath it bare again [75]. Should that fail as well, the artificial eye takes the sky’s place: visual-inertial navigation measures the drone’s own motion relatively and drifts only slowly [23], while terrain-referenced navigation recovers the absolute reference from the topography beneath the wing — confirmed up to five thousand meters of altitude in real flight tests [38]. On the lowest rung, finally, there remains pure inertia, which no longer needs any external source.

Abbildung 5: The fail-safe kill chain: layered navigation resilience. Source: Chapter 9.

What is decisive about this ladder is that no rung replaces the others; they catch one another. What inertia loses over time, the terrain gives back. What the satellite signal accomplishes under an open sky, the stored map takes over in its absence. This mutual support — sensor fusion in the technical vocabulary — is the true backbone of self-sufficiency.

9.2 When the Mesh Falls Silent

Here lies the core of the thesis, and it demands a distinction that often blurs in public debate. The mesh network from Chapter 5, carried by electronically steered phased-array antennas, is low-intercept and jam-resistant — but it is no vital nerve whose severing kills the drone. It is an amplifier. As long as it holds, several platforms fuse their sensor images into a superior shared picture; if it fails, the swarm loses that collective sharpness — but each individual drone retains what it needs to complete its mission.

For the decisive step has already been taken in Chapter 8: target recognition happens aboard, without a data link. A drone that recognizes its target with its own sensors and its own artificial intelligence, and that derives its position from inertia and terrain, needs for the strike neither a ground station nor its swarm companions. This is precisely why the loss of communication is, from the standpoint of mission success, a matter of indifference: the adversary can disrupt the nervous system, but in doing so he disrupts only the coordination, not the capacity to act. It is the same autonomy that immunizes the drone against radio jamming and, at the same time, as Chapter 14 shows, makes it so legally problematic — the two sides of the same coin.

9.3 Autonomous Terminal Guidance Through Inertia and Terrain

The thesis shows itself most sharply in the terminal phase of the approach. Here what counts is not accuracy over hours but over the final seconds. And over such short windows even pure inertial navigation is remarkably accurate: a calibrated tactical sensor array drifts, over a thirty-second loss of the satellite signal, only 1.85 meters on average [10]. A drone that has acquired its target optically in the terminal approach and tracks it with event-based, microsecond-fast image processing no longer needs any external positioning in this window. Inertia bridges the seconds; the eye guides it onto the target. The radio failure that would render a remotely piloted drone blind and useless in this moment leaves the autonomous platform untouched.

9.4 Sensor Fusion as the Backbone

Technically, this interplay rests on an architecture that has formed the gold standard of high-grade navigation for decades: an inertial core system that, through a Kalman filter, continuously fuses all available aiding sources — satellite signal, visual and terrain-referenced positioning — into a single, gap-free state estimate [3]. The inertial system supplies the low-drift short-term stability, the aiding methods the absolute long-term reference. If one source fails, the filter reweights the remaining ones; when it returns, the filter integrates it again. The ladder from Section 9.1 is therefore no rigid switching but a sliding transition that the filter rebalances at every instant.

9.5 The Error Budget Over a Complete Mission — and the Honest Limit

It would be easy to close the thesis here in triumph. Rigor demands the opposite. Pure inertial navigation is inherently unstable: its position error grows with time, and it would do so even if the sensors were perfect, because the method integrates error over time [2]. Quantum and cold-atom sensors will shrink the error sources, yet they do not abolish the inherent drift. How far an inertial backbone drifts can be quantified analytically: closed-form models derive the position error from flight duration, velocity, noise amplitude, and the gyroscope’s drift parameters, anchoring it to the civilian navigation requirement RNP-10 — using, correctly, the maximum rather than the minimum of the Allan deviation as the drift measure [20].

From this follows the true shape of the thesis. The fail-safe kill chain works not because pure inertia would carry indefinitely — it does not — but because it never has to remain on its own for long. Over minutes and hours the terrain re-anchors the drone again and again; over the final seconds of the strike, the drift is smaller than the target radius. The chain breaks precisely when both come together: a long flight over featureless terrain — over open sea, over desert — on which terrain-referenced aiding fails and pure inertia drifts away unchallenged. This is exactly why predictive trajectory planners modify the route so that the drone overflies areas of high topographic information content [35]. The resilience is real, but it is earned — a property of the system design, not a gift of physics. With this qualification, and only with it, the guiding thesis holds.

10. The Autonomous Swarm as a Weapon System

The fail-safe kill chain of the previous chapter describes the single drone. Its full military significance unfolds only in the collective. The swarm is not the sum of its drones but a weapon system in its own right, with properties no single platform possesses — and with a vulnerability lower than that of any of its components.

10.1 Distributed Intelligence and Formation Flight

A swarm is robust because it has no center. Where a single platform is lost with the failure of its flight computer, the swarm survives the loss of individual members without abandoning its task — provided the intelligence is distributed and not bundled in one node. This is exactly what multi-agent reinforcement learning methods accomplish: drones that share visual and inertial data fly in tight formation, anticipate the aerodynamic turbulence of their neighbors, and avoid collisions proactively instead of reacting after the fact. That autonomous agents push to the physical limits in pure flight control and thereby outperform human pilots is established for solo flight [34]; for the cooperative formation, scalable learning methods show that this performance can be carried over to growing swarm sizes [121].

10.2 Self-Organization, Role Assignment, and Cluster Heads

A formation that cannot be steered centrally must organize itself. The radio networks from Chapter 5 supply the foundation for this: routing protocols continuously select optimal nodes — cluster heads — by a combined measure of distance, velocity, link quality, and residual energy, bundle the traffic through them, and replace them the moment one fails [122]. The programmable data plane makes the network predictive on top of this: it gathers telemetry in flight on signal strength, computing load, and the position of each drone, and predicts looming link failures before they occur in order to reroute traffic preemptively [50]. Hierarchical learning methods, finally, master the exponentially growing decision problem of large formations by combining macro-actions and human directives [123]. The swarm distributes not only its intelligence; it manages its own structure.

10.3 Saturation Attacks and Magazine Depth

From this architecture grows the tactically most dangerous property of the swarm: its capacity for the saturation attack. A defense that reliably engages a single target becomes useless when twenty or two hundred targets arrive at once — not because it grows inaccurate, but because it runs out of time and out of rounds. The swarm transforms quantity into a qualitatively new threat. This logic inverts the old cost ratio of attack and defense: on the attacker’s side stand many cheap, replaceable members; on the defender’s side, a few expensive effectors with limited magazine depth. The economic depth of this asymmetry — and the defense’s attempts to meet it with area effects and low-cost interceptors — is the subject of Chapters 11 and 12.

10.4 Stealth Swarm and Covert Sensor Fusion

The most demanding form combines the distributedness of the swarm with the invisibility from Chapter 5. Over low-intercept phased-array directional beams, several platforms form a covert mesh in which they fuse high-resolution radar data, electro-optical target images, and sensor metrics in real time into a shared picture — without any one of them giving away its position. Distributed beamforming, in which the swarm acts as a single virtual antenna array, thereby extends not only the range but can be deployed deliberately to fend off eavesdroppers: with injected artificial noise, cooperative beamforming reaches up to twice the eavesdropping-secure data rate compared with conventional methods [70]. The stealth swarm sees more than any single drone and at the same time shows less — it is the collective embodiment of the fail-safe kill chain. Facing it now stands the defender, whose responses the third part of this study takes the measure of.

11. The Arms Race of Defense: Countering Drone Swarms

The first two parts of this study described a weapon that, ideally, no longer needs any radio link at all: it recognizes its target onboard, keeps navigating via visual-inertial odometry and inertial sensing when every connection drops, and coordinates covertly over a phased-array mesh. That is precisely what makes it the hardest of all adversaries. A defense that relies on jamming the radio grasps at empty air the moment the swarm flies under radio silence; whoever severs the operator link strikes a drone that long ago ceased to have an operator. The defense must therefore take other routes—physical rather than logical, ones that reach the airframe independently of its software. This chapter maps the technical battlefield of this new defense; the next surveys its economics.

11.1 The Asymmetric Threat Landscape

The problem begins with the physics of the targets. Drones in the lower classes are “low, slow, and small”—they fly low, slowly, and compactly, throw off barely any infrared signature, and possess a radar cross-section (RCS) that renders them nearly invisible to conventional air-defense radars. Add to this their maneuverability and a unit price that pushes an FPV drone into the range of a few hundred dollars. This combination is difficult to counter both physically and economically [124].

The more dangerous concept is the true swarm. It is decentrally organized, knows no single point of failure, and exchanges telemetry and targeting data autonomously over an ad hoc mesh. Its purpose is the saturation attack: not the individual hit but the sheer number, which temporarily overloads any point defense [124], [125]. Analysts speak of the “second drone age”—an era in which cheap, intelligent mass displaces the assumption of lasting air superiority [126], [127]. The defense thus faces a twofold task: to strike many targets at once, and to do so affordably.

11.2 The Electromagnetic Vulnerability of Unmanned Systems

The Achilles’ heel of these machines lies not in their aerodynamics but in their electronics. Commercial drones are certified to civilian EMC standards with an immunity threshold of merely 10 V/m—a value already exceeded by urban cellular fields reaching up to 61 V/m [128]. The anechoic chamber reveals what this means: above roughly 30 V/m, critical flight instability sets in, followed by an uncontrolled crash; against a base station with 40 dBW ERP, only a safe operating distance of about 20 m remains [128].

What matters is how the drone falls. It does not crash because its semiconductors burn out, but because its flight control is paralyzed. An induced differential noise violates the logic thresholds in the PWM signal between the flight controller and the electronic speed controller (ESC); the preferred coupling path runs through the internal wiring harnesses, which act as unintended receiving antennas—so-called back-door coupling. This effect takes hold long before any junction burnout occurs [129]. Research on intentional electromagnetic interference (IEMI) classifies the field further: into conducted versus radiated coupling, into front-door (via the antenna) versus back-door (via cables and enclosure gaps), and it identifies the sensor modules as the preferred target of attack [130], [131]. The drone, in other words, stops flying before it stops functioning—and this is exactly where the microwave weapon comes in.

11.3 High-Power Microwaves (HPM/HPEM): Area Effect Against Swarms

The high-power microwave (HPM) couples electromagnetic energy directly into the circuits—independent of software architecture, communication protocol, or degree of autonomy. That is its strategic trump card: it works even against the radio-silent swarm immune to jamming, because it attacks not the communication but the physics of the electronics [132]. The reference is a source at 2.45 GHz in the ISM band, chosen for the 65 to 75 percent efficiency of the cavity magnetron and because this wavelength matches the geometric resonance of typical UAV wiring. The ESC signal lines are especially telltale: at 5 to 8 cm, they sit at half-wave resonance and pick up five to ten times the voltage of other cable lengths [132].

The quantitative weight of this section is carried by a Monte Carlo simulation over 10,000 runs. In continuous-wave (CW) operation at 25 kW, the kill probability is 51.4 ± 1.0 percent at 20 m and falls to 13.1 ± 0.7 percent at 40 m. Switching to pulsed operation at 500 kW peak power with a 1 percent duty cycle—the mean power, at 5 kW, remains thermally manageable—the 90 percent kill range grows from about 18 m to roughly 88 m [132]. The real advantage over the laser is geometry: the beam cone measures about 7.5 m in diameter at 30 m and captures several drones at once—drastically lowering the requirement for precise mechanical tracking. This area effect comes at a price in safety zones: in 25 kW CW operation, the exclusion radius under ICNIRP 2020 reaches 72 m for personnel and 161 m for the civilian population in the beam direction [132].

Abbildung 6: High-power microwave vs. swarms: effect over distance. Source: Chapter 11.

That such pulses can be generated technically is experimentally documented: a MILO source (magnetically insulated transmission line oscillator) delivers 3 GW pulses in the L band at 10 to 15 percent power efficiency—a blueprint for vehicle- or UAV-mounted microwave weapons [133]. Strategically, the HPM excels against electronics and swarms and is more atmospherically robust than the laser, but it suffers from limited range and from the possibility that the adversary hardens its electronics [134]. On the industrial side stand Diehl Defence’s HPEM effector SkyWolf, embedded in the modular Sky Sphere system, as well as the U.S. reference programs AFRL THOR and its successor Mjölnir, and the GaN solid-state emitter Epirus Leonidas, which is explicitly marketed as swarm defense (all manufacturer claims or established industry knowledge without independent confirmation, confidence level C) [135].

11.4 High-Energy Lasers (HEL): The Physics of Ablation and the State of the Art

Where the microwave creeps into the circuit, the high-energy laser (HEL) burns material away. Its near-infrared beam (1.06 to 1.55 µm) is absorbed in a micrometer-deep layer; the heat conducts onward, melts, vaporizes, and at extreme power density a plasma forms [136]. The effect depends decisively on the material. Aluminum alloys initially reflect the NIR light strongly and conduct the heat well, which makes them tough: penetrating a 2 mm shell demands about 44.4 kJ, and only beyond 10 kW/cm² does the beam perforate within seconds [137]. Carbon-fiber-reinforced polymer (CFRP) behaves in the opposite way—it absorbs superbly but barely conducts the heat laterally, so that a localized thermal runaway vaporizes the epoxy matrix. The laminate fails through delamination and spallation, the flaking-off of entire plies, markedly faster than metal [138].

The state of the art is real but more modest than the marketing suggests. Rheinmetall and MBDA tested a 20 kW naval demonstrator for more than a year aboard the frigate “Sachsen” under harsh maritime conditions; naval operational readiness is projected for 2029 [139], [140]. EOS states a scalability of 50 to 150 kW for its land-based Apollo (manufacturer claim, level C) [141]; the U.S. Navy had already tested the fundamentals with the 30 kW LaWS system. Counted as established industry knowledge without peer review (level C) are Rafael’s Iron Beam (~100 kW), the British DragonFire (~50 kW), and Lockheed Martin’s HELIOS (~60 kW). Across all designs, fiber lasers possess the highest maturity for operational deployment [136].

11.5 The Laser’s Swarm Dilemma

As cleanly as the laser dismantles a single target, it just as clearly fails against the mass. Its fundamental shortcoming is sequentiality: each effector engages exactly one target at any given moment [142]. Each of them demands a dwell time—at 50 kW, about 1.3 s against a Class 1 drone, up to 4.4 s against a Class 2 [141]. On top of this comes a paradoxical kinematics: as the swarm approaches, the required dwell time does fall, but the angular separation between the drones grows, and the slew time needed to swing over rises drastically. Slew and dwell add up until the saturation swarm reaches its target before it has been fully destroyed [137]. Even with tracking optics, the platform’s energy store would remain the limit—it, not laser power alone, determines the magazine depth against swarms [143]. Then there is the atmosphere: humidity, rain, and fog attenuate the infrared, and the laser heats the traversed air enough that it defocuses the beam like a bad lens—thermal blooming [136], [137].

The answer from research is no longer power but smarter target selection. The NPS engagement study formulates strategies against sequentiality: Prioritize Proximity engages the nearest first, Take Out the Shooter the armed nodes, Focus on Information Nodes the mesh relays, and Maximize Soft Kill blinds sensors rather than destroying structures. The right strategy substantially reduces the success of heterogeneous decoy attacks [137]. The conclusion is sobering: against very large swarms, multiple lasers are required—reliance on a single, monolithic beam weapon is ruled out [142].

11.6 Kinetic and Hybrid Defense

Where energy reaches its limits, matter returns. Kinetic defense relies, first, on interceptor drones that confront the adversary in flight. Their breakthrough is called image-based visual servoing (IBVS): it decouples the drone dynamics from the camera orientation and combines them with proportional navigation guidance and a delayed Kalman filter to counter image-processing latency. In simulation, the method achieves a hit accuracy (CEP) of 0.089 m, in reality over 80 percent interception success in wind below 4 m/s and a terminal velocity of 20 m/s [144]. More recent work extends the range: a planar-sector LOS guidance intercepts agile targets in real wind at up to 138 m [145]. Translated into the market, these are Diehl’s electrically terminal-guided CICADA, the DLR project CUSTODIAN, which neutralizes the adversary by ramming [146], as well as—as level-C systems—the MARSS interceptor, capable of up to 290 km/h [135], and the U.S. counterparts Coyote and Roadrunner.

From these interceptors a new architecture takes shape that meets the swarm’s mass with mass: the containerized loitering magazine. Interceptor drones wait on the ground in weatherproof launch canisters; when the early warning detects the incoming swarm and estimates its strength, a threat-matched salvo launch releases exactly as many interceptors as there are attackers. Rather than climbing straight into the attack, they enter a loitering mode, link up over a mesh radio network, and distribute the targets among themselves in a decentralized manner—each drone takes exactly one attacker, with double assignments and self-collisions resolved by the formation itself, without a central authority and resilient to the loss of individual nodes [147], [148]. Only at the right moment, when the assigned target enters the optimal intercept geometry, does the hunter strike under image-based terminal guidance [144]. In reality this is emerging in reusable, canister-launched interceptors of the Coyote and Roadrunner class. What is striking is that this method turns the autonomous drone’s strengths against it: it works mechanically and is therefore EW-independent, it needs no signal from the radio-silent target, and through mesh rather than central control it is hardened even against jamming.

Second, classic anti-aircraft fire is reviving, made programmable. Rheinmetall’s Skynex fires 35 mm AHEAD ammunition (Advanced Hit Efficiency And Destruction): the muzzle system measures the velocity of each projectile and inductively transmits the exact detonation time, so that the round breaks apart shortly before the intercept point and ejects 152 heavy tungsten sub-projectiles as a cone-shaped cloud (manufacturer claim, level C) [149]. Against densely flying formations, a single well-placed burst is superior to the sequential laser.

11.7 Electronic Warfare as Defense

The cheapest defense remains the electronic one. RF jamming oversaturates the operator link at 2.4 or 5.8 GHz, GNSS spoofing overlays the GPS signal—both drive the drone into its fail-safe mode, into hover, landing, or return-to-home, at very low cost and over large areas [124]. Yet here lies the absolute limit of electronic warfare, and it coincides exactly with the guiding thesis of this study: a swarm that navigates optically and inertially and flies under radio silence is immune to RF jamming [124]. Whoever needs no radio loses nothing when the radio is jammed.

A subtler counter-tactic even targets the microwave weapon itself. A forward-sent vanguard UAV—a “movement-to-contact”—triggers the HPM weapon and thereby warns the following drones, even under total communications failure. For the HPM betrays itself through its own destructive radiation with a detectable signature that cheap, attritable drones can exploit [150]. The defense that strikes everything electronic is thus betrayed by its own effect.

11.8 Detection and Tracking: LiDAR Swarm Reconnaissance

None of these weapons works on what it has not first seen—and seeing is precisely what is difficult against LSS swarms. Radar mistakes plastic drones for birds and clutter; passive RF scanners fail against radio-silent formations [124]. LiDAR circumvents both: it scans the scene with near-infrared pulses at 905 or 1550 nm, detects via avalanche photodiodes and time-to-digital converters, and delivers high-resolution 3D point clouds independent of daylight [151]. The LiSWARM system combines these data with priority-aware clustering, neural object detection, and trajectory tracking. In field experiments at two drone light shows with 150 and 500 UAVs, it achieved up to 98 percent detection accuracy with reliable single-trajectory tracking; the scalability analysis extends to 15,000 drones [152]. These point clouds are the prerequisite for steering an interceptor precisely into the swarm and for assigning HPM or HEL optimally—without clean detection, every effector choice remains a shot in the dark [152].

With this, the technical arsenal of the defense has been described: from the electromagnetic area weapon through the surgical laser and the kinetic drone to the seeing LiDAR. Left out was the question that, in the end, may decide victory and defeat more powerfully than any kill probability: what does an intercept cost—and who, in the exchange of mass for mass, runs out of money first? The next chapter turns to the economics of this defense.

12. The Economics of the War of the Machines

The previous chapter showed that defending against autonomous swarms is technically solvable—with microwaves that strike entire formations, with lasers that melt through aluminum in seconds, with interceptors that ram their target on the approach. But the question of whether a weapon works is not the question of whether it pays off. A Patriot missile destroys a Shahed with near-complete certainty; it does so at a price that exceeds the drone’s a thousandfold. The value of a defense system is therefore decided not by its hit rate but by a balance sheet—and in this balance sheet, attack and defense have turned upside down in the second drone age.

12.1 Cost-per-Kill and the Cost Calculus of Saturation

The central metric of this war is called cost-per-engagement: what it costs to neutralize a single incoming target. A comparative model that evaluates nineteen C-UAS systems against reported conflict data from the years 2022 to 2026 and a Monte Carlo simulation arrives at a sobering finding: this metric spreads across more than five orders of magnitude—from a fraction of a cent up to several million dollars per intercept [153]. The figures stem from a non-peer-reviewed preprint and are to be read as an estimate, not a measurement; their order of magnitude, however, coincides with what independent observers report from the most recent conflicts.

Abbildung 7: The economics of defense: cost per engagement (model estimate, log scale). Source: Chapter 12.

(Values as a model estimate per [153]; not peer-reviewed.)

This table holds the entire asymmetry of the conflict. The jammer, at roughly a cent per use, is almost free—and against a radio-silent, optically navigating swarm of the latest generation, ineffective. The laser shoots for the price of a coffee but can hold only a single target at any one moment. And the guided missile, which hits practically any target, ruins the defender the moment the targets arrive in swarms. The model reports a cost-exchange ratio of up to 190:1 to the defense’s disadvantage when a surface-to-air missile costing several million intercepts a drone in the three-digit dollar range [153].

From this follows an inherent logic of attack. The saturation attack aims first not at destruction but at exhaustion: it is meant to bleed out the defender’s magazine and budget by forcing him to answer cheap with expensive, until his expensive means run out [153], [127]. The drone mass becomes an economic weapon—one that prevails not through its effect on the target but through the bill it imposes on the other side.

12.2 Magazine Depth and Reloadability

This is precisely where the decisive quantity shifts—away from the hit rate, toward magazine depth: how many targets a system can engage before it runs dry, and how quickly it is refilled. For the high-energy laser, this magazine is bounded not by ammunition but by electrical power. Models for ship-based lasers show that it is not the laser power itself but the platform’s energy store that caps the number of shots against a swarm [143]. A laser may shoot cheaply—but when the stores are empty, it does not shoot at all, and against the second wave moving up, it stands mute.

This brings to the center a factor that appears on no effector’s data sheet: industry. Sustained conflicts reward those states that can produce, repair, and reload at scale—be it 155 mm ammunition, interceptors, or FPV kits [127]. A country’s magazine depth is not the number of missiles in its bunkers but the speed with which its factories replace them. Whoever reloads faster than the adversary expends wins the endurance duel—regardless of which system is superior in the individual engagement. Magazine depth and reload capacity together thus become the strategic quantity on which the viability of any defense architecture is decided.

12.3 The Economic Inversion of Attack and Defense

For decades, air defense was the expensive business: sophisticated interceptor missiles against few, precious attackers. The drone has inverted this relationship. Now cheap mass stands against expensive precision—and the defender is the expensive one. This inversion, however, can be inverted a second time. The interceptor drone, available for around 3,500 USD, breaks open the disadvantageous cost curve: it confronts the adversary’s mass with its own, comparably cheap and arbitrarily scalable mass [153]. Mass against mass, at prices of the same order of magnitude—that is the economic answer to the swarm.

Subtler, but no less consequential, is the vanguard economics of directed energy. A microwave weapon betrays itself through its own effect: the destructive microwaves generate a detectable signature. A cheap, expendable vanguard UAV can trigger this defense through mere approach and report its position to the following drones—even under total radio failure, entirely without the expensive apparatus of coordinated swarm behavior [150]. The strength of the high-energy weapon, its overwhelming physical effect, thus becomes its economic weakness: it pays for every shot with a betraying signature that the adversary exploits for a few thousand dollars.

The synthesis of these findings is unambiguous—and it is a refusal of the search for the one wonder weapon. No single system bears the burden alone: the laser fails on sequentiality, the missile on price, the jammer on the targets’ autonomy, the microwave on its own signature. Viable is solely a diversified portfolio, a layered air defense in which each means does what it can do most cheaply [125], [134]. High-volume swarms belong before the barrels of the anti-aircraft gun and in the cone of directed energy, where the individual intercept costs almost nothing; the expensive missile interceptors remain reserved for the high-value targets—cruise missiles, combat aircraft—for which their price makes sense [134]. It is not the flawless weapon that decides, but the clever orchestration of many imperfect ones: sensor fusion, directed energy, electronic warfare, and kinetic interceptors, brought together in a network.

With this, the technical arc closes and a larger one opens. If the defense of the machine is, in the end, a question of the balance sheet, then it is no longer a purely military one—it touches the industrial, the strategic, and, once machines decide over life and death by cost curves, the legal and the ethical. To these further questions the concluding discussion now turns.

13. Discussion: Convergence Toward the 2036 Drone

The preceding chapters surveyed disciplines that long developed in isolation. The discussion assembles them into a single picture, examines how mature that picture truly is, and demonstrates it through a mission account—before naming the limits of its own method.

13.1 The Interplay of the Four Pillars

The autonomous combat drone of 2036 rests on four pillars whose value lies not in their individual performance but in their interlocking. Navigation (Chapters 3, 4, 6) provides the positioning that no adversary can switch off, because it cascades from external support down to pure inertia. The brain (Chapter 7) turns sensor data into decisions, and it does so in the single-digit watt range onboard, not in a distant cloud. Communication (Chapter 5) binds the platforms into a swarm whose covert directional beams deny the networking to enemy reconnaissance. And effect (Chapters 8, 9, 10) fuses these three into the capacity to identify and engage a target autonomously.

Only convergence produces the property this study calls its guiding thesis. A drone with excellent navigation but without onboard recognition would remain a remotely guided projectile. A drone with brilliant artificial intelligence but without resilient positioning would be blind in a jammed environment. The fail-safe kill chain is a system property, not a component—it arises from the interplay, and it collapses the moment a pillar is missing.

Abbildung 8: Convergence of the four pillars into the autonomous combat drone of 2036. Source: Chapter 13.

13.2 Technology Readiness Levels and Remaining Bottlenecks

An honest synthesis distinguishes the proven from the hoped-for. Highly mature and amply documented are the tactical MEMS inertial sensors, visual-inertial navigation including event-based cameras, terrain-referenced navigation up to high altitudes, the active defense against satellite spoofing, and the efficient AI architectures—ternary language models, state-space models, and continuous-time networks—whose suitability for drones is documented in peer-reviewed work [34], [75], [104]. On the defensive side, too, the core physical values of high-power microwaves, image-guided interceptors, and LiDAR swarm reconnaissance have been verified.

Still immature or projected, by contrast, are several building blocks that public accounts readily treat as given. Cold-atom inertial sensors exist as compact demonstrators, not as series components. The spectacular energy-efficiency figures of analog in-memory compute units apply to the memory array, not to the overall system with its converter periphery. Silicon photonics moves data within compute packages, not within the airframe. And on the defensive side, many of the systems named—from Iron Beam through DragonFire to individual manufacturers’ microwave weapons—rest on industry claims, not on independent measurement. The greatest remaining bottleneck is no single component but the reliability of perception under adversarial deception: the domain gap and physical adversarial patterns remain the unsolved Achilles’ heel of autonomous target recognition (Chapter 8).

13.3 Scenario 2036: A Mission in Denied Airspace

To make the interplay tangible, let a mission be sketched—explicitly as a constructed illustration that rests solely on the values documented in this study, not as the report of a real event.

A formation of twelve platforms launches at dusk. Up to the edge of the contested zone it navigates by the satellite signal, networked over narrow, barely interceptable directional beams in the Ku band. As it enters the denied airspace, the adversary begins to jam. The satellite signal goes first—not through crude noise, but through a precise deception that slowly shifts the reported position. The onboard receiver responds as Chapter 6 describes: it forces each drone into a brief evasive maneuver, detects the forgery from the discrepancy between inertial acceleration and reported course, and restores the true position within half a second [75]. Where the deception grows too strong, the filter discards the satellite signal entirely and steps down one rung: to visual and terrain-referenced navigation, which guides the formation along a route that a forward-looking path planner has deliberately laid over contour-rich terrain.

Then the radio goes dark. A broadband jam severs the mesh; the formation loses its shared picture. For a remotely guided formation this would be the end. For this swarm it is an inconvenience. Each drone already carries its target onboard—recognized by its own sensors, classified by its own artificial intelligence in the single-digit watt range. It navigates on by inertia and terrain, and in the terminal run, over the final seconds, pure inertial navigation suffices: its drift stays under two meters within this window [10], while the event-based camera fixes the target in its sights with microsecond-fast acuity. The loss of communication has cost the coordination, not the effect.

The defender, whose means Chapter 11 surveys, faces the most uncomfortable variant of the threat: a radio-silent, autonomously navigating swarm against which electronic jamming remains ineffective. What remains to him are the area weapon of the high-power microwave, which engages several members at once, and the programmable shrapnel cloud of the flak gun—means whose economics decide between victory and exhaustion (Chapter 12).

13.4 Limits of the Study and Open Research Questions

This investigation is a synthesis of the literature, not an experiment. Its force ends where its sources end. Some of the performance figures—especially on the industry and defensive side—derive from manufacturer claims and non-peer-reviewed model calculations; they are marked as such and carry less weight than the verified core numbers. The projection to 2036 extends documented curves, yet it cannot foresee discontinuities: a breakthrough in quantum sensing no more than a ban under international law that would sever lines of development. Three questions in particular remain open, which this study raises but does not answer: How reliable will autonomous target recognition under deliberate deception ever be? How far can the magazine depth of the defense be economically scaled up against the mass of the swarm? And who bears responsibility when the fail-safe kill chain strikes the wrong thing? The last of these questions leads beyond technology—and into the chapter that follows.

14. Legal, Ethical, and Security-Policy Assessment

What the preceding chapters described as technically feasible—a drone that recognizes targets onboard, networks covertly, and carries the mission to completion on its own even in the radio shadow—shifts the decisive question from engineering to norms. A weapon that triggers itself raises problems no data sheet solves: who distinguishes combatant from civilian when the distinction takes place within a web of weights, who is liable for the machine’s error, and what does it mean for the stability between states when software decides over life and death by the second? This chapter situates the technology within prevailing law, the ethical debate, and security policy—weighing the arguments and without suppressing the sober counterposition.

14.1 Autonomous Weapon Systems and International Law

The very term is contested. “Autonomous weapon systems” (lethal autonomous weapon systems, LAWS) have no agreed definition under international law; competing interpretations are often less the product of technical analysis than of geopolitical discourses that recast divergent understandings of AI into strategic assets, thereby impeding common standards [154]. As a working definition, the functional determination of the International Committee of the Red Cross has prevailed: systems that, after activation, “select and apply force to targets without human intervention,” triggered by sensor information and a generalized “target profile,” without the user any longer knowing the specific target, the timing, or the location [1]. It builds on the familiar taxonomy—human-in-the-loop (the human authorizes each use), human-on-the-loop (the human supervises and can abort), human-out-of-the-loop (the system selects and engages without intervention)—and identifies the last stage as the genuinely problematic one.

International humanitarian law (IHL) applies to such systems without restriction: the first of the eleven guiding principles of the CCW group of experts makes clear that it “continues to apply fully to all weapons systems, including the potential development and use of LAWS” [155]. From this follow the three cardinal obligations—distinction between combatants and civilians, proportionality (no collateral damage excessive in relation to the military advantage), and precautions in attack—whose compatibility with an autonomous target selection by generalized profiles is, as a matter of legal doctrine, regarded as limited at best [156]. To this is added the procedural obligation under Article 36 of Additional Protocol I to determine, in the development and acquisition of new weapons, means, or methods, whether their use would be unlawful under international law—reaffirmed by CCW principle (e) [155]. Where treaty law is silent, the Martens Clause takes hold with its protection through the “principles of humanity” and the “dictates of public conscience”; it provides normative orientation, but on account of its abstractness it does not serve as an independent legal basis [157] and operates above all as an ethical hinge—there, where the ICRC warns against replacing human life-and-death decisions with “sensor, software, and machine processes” [1].

This is where the concept of Meaningful Human Control (MHC) enters, a central theme of the CCW debate, because it addresses the accountability gap—the problem that, for the wrongful act of an autonomous machine, perhaps no one is liable. Its status under IHL remains unclear: systematic interpretation anchors MHC in Article 36 and Article 57 AP I, yet the reservations of individual great powers prevent a uniform state practice; these powers are nonetheless not “persistent objectors,” since they do not expressly reject human control [157]. Against this, CCW principle (b) holds that human responsibility must be retained across the “entire life cycle,” “since accountability cannot be transferred to machines” [155]. How it would be operationalized is contested—for instance through a clear allocation of functions between human and machine [158]; a pointed counterposition doubts that genuine control in combat is possible at all and proposes “Meaningful Human Certification,” whereby responsibility would shift to testing and authorization prior to deployment [159].

The primary documents trace the state of negotiations. On May 12, 2021, the ICRC recommended new legally binding rules on three pillars: first, a ban on unpredictable autonomous weapons whose effects cannot be sufficiently understood and which therefore operate indiscriminately; second, a ban on their use against persons; third, a regulation of all remaining systems through limits on target types, duration, scope, and situation of use, as well as requirements for human–machine interaction (effective oversight, timely intervention, the ability to deactivate) [1]. The CCW group affirmed its eleven guiding principles in 2019, which the ICRC commended as a “useful basis” for an effective normative and operational framework [160]. The provisional high point is marked by resolution A/RES/78/241 of the UN General Assembly of December 22, 2023, the body’s first standalone engagement with LAWS: it reaffirms the applicability of international law—the UN Charter, IHL, human rights—voices concern over possible consequences for global security and international stability, and tasks the Secretary-General with seeking the views of states [161]. Binding prohibitive norms have not yet grown from it.

14.2 Proliferation and the Democratization of Precision

The second shift is economic in nature. Software-defined radio and consumer drones lower the threshold of entry to precision effect so far that capabilities once reserved for great powers move within reach of small actors. This becomes visible in a dramatic cost asymmetry: first-person-view drones can be had from 300 to 500 USD, while the interceptors to defend against them lie in the millions. From recent conflicts, cost-exchange ratios of up to 190:1 to the defender’s disadvantage have been reported [153]. These figures are to be read as conflict observations, not as peer-reviewed measurement—yet their order of magnitude explains the driving force behind the mass proliferation of autonomous precision. Without international coordination, a loss of control and a race to the bottom loom [162].

Against the dramatization, a sober correction is in order. The attention paid to lethal autonomy is overdrawn; in fact, armed forces have so far used artificial intelligence above all for data processing and for accelerating targeting—as a force multiplier for reconnaissance, not as an autonomous shooter [117]. This accords with the finding that fully self-governing weapons, removed from human control, are “more a conceptual possibility than a military reality” and that the discourse risks fixating on imaginaries rather than on real systems [154]. The regulatory task therefore consists in anticipating a foreseeable development without preempting it.

14.3 Escalation Dynamics and Strategic Stability

The third shift concerns the stability between states—and closes the arc back to human control. Even where a human nominally remains “on the loop,” their oversight withers to mere confirmation when the kill chain is compressed to seconds. Operators tend to overtrust machine outputs—automation bias—which undermines control precisely in time-critical loops [158]; the displacement of responsibility onto testing and authorization describes combat more honestly than the notion of effective real-time oversight [159].

The real driver of this development is reaction time. The machine’s perception-decision-action loop closes in milliseconds, whereas a human needs fractions of a second to several seconds for the same—to perceive, decide, and act. It is this asymmetry of orders of magnitude that generates the pull toward the fully autonomous, out-of-the-loop drone: where the approach is faster than human cognition, the machine appears to be the only actor that can respond in time. And this is precisely the normative crux. That the machine can perceive and propose faster does not justify ceding to it the release of force as well; speed is an argument for machine-accelerated reconnaissance, not for abolishing the human decision over life and death. Meaningful human control over the effect therefore remains a deliberate stance against the pull of reaction time—not a technical default, but a decision to be defended [159], [1].

At the strategic level, this heightens the danger of unintended escalation. AI-augmented conventional systems count among the strategic non-nuclear weapons that shift the calculus of nuclear stability [163]. Between nuclear powers, machine acceleration can raise the risk of inadvertent escalation—through the security dilemma, the fog of war, and doctrinal misreadings that can no longer be corrected within the compressed time window [164]. As a countermeasure, “asymmetric arms control” offers itself, mitigating the stability risks of new technologies, provided the concepts of arms control and strategic stability are adapted to the changed situation [165]. That the danger is not a purely security-policy one is underscored by the ICRC when it cites the “risks of conflict escalation” as a consequence of the loss of control—the bridge between the legal and the strategic argument [1].

With this, the normative debate connects to the technical Achilles’ heel of the preceding chapters. The domain gap between training and deployment distribution and the susceptibility to physical adversarial attacks are precisely what the ICRC recommendation means by “unpredictable” weapons: a system whose misclassification cannot be reliably predicted cannot fulfill the obligation of distinction. The law’s predictability problem is thus no abstract reservation but the legal obverse of a measurable engineering problem—the ground for the conclusion that brings together technical feasibility and normative limit.

15. Conclusion and Outlook

15.1 Core Findings

This study surveyed the state of the art across seven disciplines and condensed it into a single system picture. Three findings carry the result. First, the self-sufficiency of navigation is no longer a vision: tactical MEMS inertial sensors, visual and terrain-referenced methods, and the active defense against satellite spoofing have each been proven individually and combine into a cascaded positioning that the adversary cannot switch off entirely. Second, the brain onboard has become reality: ternary language models, state-space models, and continuous-time networks bring cognitive performance into the watt budget of an airframe, and for vision-based drone navigation this is directly documented. Third, the defense has as yet no cheap answer to the radio-silent threat that thus arises: it must reach for area weapons and low-cost interceptors whose value is decided less by their physics than by their economics.

15.2 The Fail-Safe Kill Chain—Assessment of the Guiding Thesis

The guiding thesis has withstood scrutiny, but not in its naive form. It is not the case that pure inertia would compensate for the loss of all external support indefinitely—it does not, for its drift grows inexorably [2]. What is true is more precise and more interesting: the combat drone of 2036 is designed so that it is never reliant on a single source long enough for that source’s failure to break the mission. Onboard target recognition makes the data link dispensable for the engagement run; the cascaded navigation catches the failure of any single rung with the next; and over the few seconds of the terminal run, even pure inertia is accurate enough. The loss of the mesh communication over the phased-array antennas costs the swarm its collective acuity—but not its ability to act. In precisely this sense, and only in it, does the communication failure no longer matter. The resilience is real, yet it is earned: a property of the system design that tears where a long flight over featureless terrain denies the terrain-referenced support.

15.3 Outlook to 2036 and Beyond

The lines of development point in one direction, in which the threshold of entry falls further and autonomy rises further. What is cutting-edge research today—event-based vision, ternary models, neuromorphic compute units—will by 2036 become an embedded matter of course, just as satellite-based navigation became one within a single generation. With this, the decisive question shifts from technical feasibility to controllability. The sober assessment of the research cautions against treating full lethal autonomy too hastily as given: militarily, artificial intelligence has so far served above all to accelerate the targeting process, not to wholly disempower the human [117]. Whether it stays that way is not a technical decision but a political and legal one—and it brooks no delay, for the technology does not wait. The fail-safe kill chain is built. The question of 2036 is not whether it works, but who keeps a hand on its trigger.

Annex A — Abbreviations and Acronyms

Annex B — Glossary of Key Technologies

Fail-safe kill chain. The guiding concept of this study: the system design principle of building a combat drone so that the loss of any single supporting source (satellite signal, radio link, swarm) does not break the mission, because onboard target recognition and layered self-contained navigation catch each other.

Event camera. A neuromorphic image sensor whose pixels respond asynchronously and independently, reacting only to local changes in brightness. It delivers microsecond latency, roughly 140 dB of dynamic range, and virtually no motion blur — the physical answer to the classic limits of frame-based perception.

Flying Ad-Hoc Network (FANET). A self-organizing, infrastructure-free radio network among drones, characterized by three-dimensional mobility, rapid topology changes, and intermittent connectivity.

Terrain Referenced Navigation (TRN). Absolute positioning achieved by matching the measured topography beneath the aircraft against stored elevation models — independent of the satellite signal, and most effective over terrain rich in contour.

Inertial navigation (INS/IMU). Navigation derived solely from measured acceleration and angular rate, without any external reference; inherently immune to jamming and spoofing, but subject to drift over time.

LPI/LPD waveform. A radio transmission with a low probability of intercept and detection, achieved through narrow directional beams, frequency hopping, and spreading, so that the signal vanishes into the noise.

Meaningful Human Control (MHC). A legal and ethical guiding concept under international law that demands effective human control over the decision to employ a weapon and addresses the accountability gap of autonomous systems.

Multifunction Advanced Data Link (MADL). A low-probability-of-intercept stealth data link of the fifth generation that ties several low-observable platforms together into a covert mesh.

Phased-array antenna. A planar arrangement of many radiating elements whose main beam is steered electronically, in microseconds, through computer-controlled phase shifts — with no mechanical movement.

Processing-in-Memory (PIM). A computing architecture that moves the computation into the memory array itself, thereby bypassing the von Neumann bottleneck; analog crossbars perform vector-matrix multiplications in a single clock cycle.

SemperFi. An anti-spoofing architecture for drones that exposes a counterfeit signal by forcing a flight maneuver and then mathematically cancels it through successive interference cancellation, recovering the genuine satellite signal.

Ternary language model (BitNet b1.58). A neural network whose weights take only the values −1, 0, and +1; it eliminates floating-point multiplication and makes billion-parameter inference possible within a single-digit wattage.

Visual-Inertial Odometry (VIO). The fusion of camera imagery and inertial data to estimate relative self-motion; together with terrain referenced navigation, it forms the artificial eye of the drone.

References

[1] International Committee of the Red Cross — ICRC position on autonomous weapon systems. 12.05.2021. https://www.icrc.org/en/document/icrc-position-autonomous-weapon-systems

[2] Wright, M.; Anastassiou, L.; Mishra, C.; Davies, J.; Phillips, A. M.; Maskell, S.; Ralph, J. F. — Cold atom inertial sensors for navigation applications. Frontiers in Physics 10:994459, 2022. DOI 10.3389/fphy.2022.994459. https://doi.org/10.3389/fphy.2022.994459

[3] El-Sheimy, N.; Youssef, A. A. — Inertial sensors technologies for navigation applications: state of the art and future trends. Satellite Navigation 1:2 (Springer Open), 2020. DOI 10.1186/s43020-019-0001-5. https://doi.org/10.1186/s43020-019-0001-5

[4] Langfelder, G.; Bestetti, M.; Gadola, M. — Silicon MEMS inertial sensors evolution over a quarter century. J. Micromech. Microeng. 31(8):084002, 2021. DOI 10.1088/1361-6439/ac0fbf. https://doi.org/10.1088/1361-6439/ac0fbf

[5] GuideNav — What Are the Key Strengths and Limitations of the STIM300 MEMS IMU? Blog (Herstellerangabe; STIM300 & GUIDE730). https://guidenav.com/blog/what-are-the-key-strengths-and-limitations-of-the-stim300-mems-imu/

[6] Advanced Navigation — Inertial Measurement Unit (IMU) – An Introduction. Tech-Article (Herstellerangabe; Motus-Kennwerte). https://www.advancednavigation.com/tech-articles/inertial-measurement-unit-imu-an-introduction/

[7] Analog Devices — Precision Tactical Grade MEMS IMU Delivers Breakthrough System Level Advancements for Positioning and Navigation Applications. Investor/Press Release (Herstellerangabe; ADIS16490). https://investor.analog.com/news-releases/news-release-details/precision-tactical-grade-mems-imu-delivers-breakthrough-system/

[8] Safran Navigation & Timing — MEMS Accelerometers und MS1000 Datasheet (Datenblattangabe; Colibrys MS1000). https://safran-navigation-timing.com/solution/mems-accelerometers/ ; https://www.safran-group.com/sites/default/files/2022-04/safransensingtechnologies-30sms1000f0921.pdf

[9] LIDAR Magazine — SBG Systems Unveils World’s First MEMS-Based Gyrocompassing IMU. 1. April 2025 (Fachpresse/PR; < 1° Heading, < 150 g, 2 W, GNSS-frei). https://lidarmag.com/2025/04/01/sbg-systems-unveils-worlds-first-mems-based-gyrocompassing-imu/

[10] Xuan, J.; Zhu, T.; Peng, G.; Sun, F.; Dong, D. — A Review on the Inertial Measurement Unit Array of Microelectromechanical Systems. Sensors 24(22):7140, 2024. DOI 10.3390/s24227140. https://doi.org/10.3390/s24227140

[11] Han, S.; Meng, Z.; Omisore, O. M.; Akinyemi, T. O.; Yan, Y. — Random Error Reduction Algorithms for MEMS Inertial Sensor Accuracy Improvement — A Review. Micromachines 11(11):1021, 2020. DOI 10.3390/mi11111021. https://doi.org/10.3390/mi11111021

[12] Lv, J.; Zhu, F.; Zhang, X. — Overcoming the Intrinsic Performance Limitations of MEMS IMU via Diffusion-Based Generative Learning. arXiv:2605.16391, 2026. DOI 10.48550/arXiv.2605.16391. https://arxiv.org/abs/2605.16391

[13] State-of-the-Art and Development Trends of Inertial Navigation Systems Based on the Ring Laser Gyroscopes. Eco-Vector Journals Portal, 2024. https://journals.eco-vector.com/1993-7296/article/view/642242

[14] Carr, K.; Greer, R.; May, M. B.; Gift, S. — Navy testing of the iXBlue MARINS Fiber Optic Gyroscope (FOG) Inertial Navigation system (INS). IEEE/ION PLANS 2014. DOI 10.1109/plans.2014.6851515. https://doi.org/10.1109/plans.2014.6851515

[15] Heckman, D. W.; Baretela, L. M. — Improved affordability of high precision submarine inertial navigation by insertion of rapidly developing fiber optic gyro technology. IEEE PLANS 2000. DOI 10.1109/plans.2000.838332. https://doi.org/10.1109/plans.2000.838332

[16] Xu, H.; Wang, L.; Zu, Y.; Gou, W.; Hu, Y. — Application and Development of Fiber Optic Gyroscope Inertial Navigation System in Underground Space. Sensors 23(12):5627, 2023. DOI 10.3390/s23125627. https://doi.org/10.3390/s23125627

[17] Dell’Olio, F.; Natale, T.; Wang, Y.-C.; Hung, Y.-J. — Miniaturization of Interferometric Optical Gyroscopes: A Review. IEEE Sensors Journal 23(23), 2023. DOI 10.1109/jsen.2023.3327217. https://doi.org/10.1109/jsen.2023.3327217

[18] Fang, J.; Qin, J. — Advances in Atomic Gyroscopes: A View from Inertial Navigation Applications. Sensors 12(5):6331, 2012. DOI 10.3390/s120506331. https://doi.org/10.3390/s120506331

[19] Lee, J. et al. — A compact cold-atom interferometer with a high data-rate grating magneto-optical trap and a photonic-integrated-circuit-compatible laser system. Nature Communications 13, 2022. DOI 10.1038/s41467-022-31410-4. https://doi.org/10.1038/s41467-022-31410-4

[20] Wheeler, J. M.; Chamoun, J. N.; Dangui, V.; Digonnet, M. J. F. — Analytic Method for Estimating Aircraft Fix Displacement from Gyroscope’s Allan-Deviation Parameters. arXiv:2202.09360 (eess.SY), 2022. DOI 10.48550/arXiv.2202.09360. https://arxiv.org/abs/2202.09360

[21] Surace et al. — An Evaluation of MEMS-IMU Performance on the Absolute Trajectory Error of Visual-Inertial Navigation System. Sensors / PMC9024873, 2022. https://pmc.ncbi.nlm.nih.gov/articles/PMC9024873/

[22] van Goor, P.; Mahony, R. — An Equivariant Filter for Visual Inertial Odometry. arXiv:2104.03532, 2021. https://arxiv.org/abs/2104.03532

[23] Gallego, G.; Delbrück, T.; Orchard, G.; … Scaramuzza, D. — Event-Based Vision: A Survey. IEEE TPAMI, 2020. DOI 10.1109/tpami.2020.3008413 — https://doi.org/10.1109/tpami.2020.3008413

[24] Choi, E.; Hong, J.; Lee, D.; et al. — Benchmarking Visual Feature Representations for LiDAR-Inertial-Visual Odometry Under Challenging Conditions. IEEE Access, 2026. DOI 10.1109/ACCESS.2026.3667112 — https://arxiv.org/abs/2603.18589

[25] Mueggler, E.; Gallego, G.; Rebecq, H.; Scaramuzza, D. — Continuous-Time Visual-Inertial Odometry for Event Cameras. IEEE Transactions on Robotics, 2018. DOI 10.1109/TRO.2018.2858287 — https://arxiv.org/abs/1702.07389

[26] Niu, J.; Zhong, S.; Lu, X.; Shen, S.; Gallego, G.; Zhou, Y. — ESVO2: Direct Visual-Inertial Odometry with Stereo Event Cameras. IEEE Transactions on Robotics, 2025. DOI 10.1109/TRO.2025.3548523 — https://arxiv.org/abs/2410.09374

[27] Zhu, A. Z.; Yuan, L.; Chaney, K.; Daniilidis, K. — EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. Robotics: Science and Systems XIV, 2018. DOI 10.15607/rss.2018.xiv.062 — https://doi.org/10.15607/rss.2018.xiv.062

[28] Davies, M.; Wild, A.; Orchard, G.; et al. — Advancing Neuromorphic Computing With Loihi: A Survey of Results and Outlook. Proceedings of the IEEE, 2021. DOI 10.1109/jproc.2021.3067593 — https://doi.org/10.1109/jproc.2021.3067593

[29] Bartolozzi, C.; Indiveri, G.; Donati, E. — Embodied neuromorphic intelligence. Nature Communications, 2022. DOI 10.1038/s41467-022-28487-2 — https://doi.org/10.1038/s41467-022-28487-2

[30] Efficient Visual-Inertial Perception on Resource-Constrained Systems (LEVIO / GAP9). ETH Zürich Research Collection. https://www.research-collection.ethz.ch/items/37e0067a-0b9c-4a2d-916c-cd1ca182a766

[31] Cioffi, G.; Bauersfeld, L.; Kaufmann, E.; Scaramuzza, D. — Learned Inertial Odometry for Autonomous Drone Racing. arXiv:2210.15287, 2022/2023. https://arxiv.org/abs/2210.15287

[32] Cioffi, G.; Bauersfeld, L.; Scaramuzza, D. — HDVIO: Improving Localization and Disturbance Estimation with Hybrid Dynamics VIO. arXiv:2306.11429, 2023. https://arxiv.org/abs/2306.11429

[33] Cui, J.; Yu, F.; Zhang, L.; Hu, Y.; Zou, D. — AI-IO: An Aerodynamics-Inspired Real-Time Inertial Odometry for Quadrotors. arXiv:2603.00597, 2026. https://arxiv.org/abs/2603.00597

[34] Kaufmann, E.; Bauersfeld, L.; Loquercio, A.; Müller, M.; Koltun, V.; Scaramuzza, D. — Champion-level drone racing using deep reinforcement learning. Nature 620(7976):982–987, 2023. DOI 10.1038/s41586-023-06419-4 — https://doi.org/10.1038/s41586-023-06419-4

[35] Carroll, J. D.; Canciani, A. J. — Terrain-referenced navigation using a steerable-laser measurement sensor. NAVIGATION 68(1):115–134, 2021. DOI 10.1002/navi.406 — https://doi.org/10.1002/navi.406

[36] Groves, P. D.; Handley, R. J.; Runnalls, A. R. — Optimising the Integration of Terrain Referenced Navigation with INS and GPS. Journal of Navigation 59(1):71–89, 2006. DOI 10.1017/s0373463305003462 — https://doi.org/10.1017/s0373463305003462

[37] Turan, B. — Comparison of Nonlinear Filtering Methods for Terrain Referenced Aircraft Navigation. IEEE/ION PLANS, 2020. DOI 10.1109/plans46316.2020.9109984 — https://doi.org/10.1109/plans46316.2020.9109984

[38] Kim, T.; Nam, S.; Lee, H.; Oh, J. — Verification of Vision-Based Terrain-Referenced Navigation Using the Iterative Closest Point Algorithm Through Flight Testing. Sensors 25(18):5813, 2025. DOI 10.3390/s25185813 — https://www.mdpi.com/1424-8220/25/18/5813

[39] Terrain Referenced Navigation of AUVs and Submarines Using Multibeam Echo Sounders (FFI / KNM Utsira). NavLab. https://www.navlab.net/Publications/Terrain_Referenced_Navigation_of_AUVs_and_Submarines_Using_Multibeam_Echo_Sounders.pdf

[40] Melo, J.; Matos, A. — Survey on advances on terrain based navigation for autonomous underwater vehicles. Ocean Engineering, 2017. DOI 10.1016/j.oceaneng.2017.04.047 — https://doi.org/10.1016/j.oceaneng.2017.04.047

[41] Rosati, S.; Kruzelecki, K.; Heitz, G.; Floreano, D.; Rimoldi, B. — Dynamic Routing for Flying Ad Hoc Networks. IEEE Transactions on Vehicular Technology, 2014/2015. DOI 10.1109/TVT.2015.2414819. https://arxiv.org/abs/1406.4399

[42] Cong Pu. — A Stochastic Packet Forwarding Algorithm in Flying Ad Hoc Networks: Design, Analysis, and Evaluation. IEEE Access, 2021. https://congpu.github.io/document/paper/ieee_access_2021.pdf

[43] Mishra, S.; Bhargava, B.; Liu, Z.; Islam, S. — UAV-CAS: A Calibrated Digital-Twin Dataset for Intrusion Detection in UAV Swarm Networks. arXiv:2606.17845, 2026. https://arxiv.org/abs/2606.17845

[44] Banaei, A.; Cline, D. B. H.; Georghiades, C. N.; Cui, S. — On the Random 1/2-Disk Routing Scheme in Wireless Ad Hoc Networks. arXiv:1102.5739, 2011/2013. https://arxiv.org/abs/1102.5739

[45] — A Robust Routing Protocol in Cognitive Unmanned Aerial Vehicular Networks (Central Node Resolution Factor, CNRF). PMC11478788, NIH. https://pmc.ncbi.nlm.nih.gov/articles/PMC11478788/

[46] — Reinforcement Learning-Based Routing Protocols in Flying Ad Hoc Networks (FANET): A Review. Mathematics (MDPI) 10(16):3017, 2022. https://www.mdpi.com/2227-7390/10/16/3017

[47] — Review and Comparison of Emerging Routing Protocols in Flying Ad Hoc Networks. Symmetry (MDPI) 12(6):971, 2020. https://www.mdpi.com/2073-8994/12/6/971

[48] — A survey on FANET routing from a cross-layer design perspective. Computer Communications (Elsevier), 2021. https://ruomoplus.lib.uom.gr/bitstream/8000/180/1/1-s2.0-S1383762121001934.pdf

[49] Kfoury, E.; Crichigno, J.; Bou-Harb, E. — An Exhaustive Survey on P4 Programmable Data Plane Switches: Taxonomy, Applications, Challenges, and Future Trends. IEEE Access, 2021. DOI 10.1109/ACCESS.2021.3086704

[50] — P4 FANET In-band Telemetry (FINT) for AI-assisted wireless link failure forecasting and recovery. Computer Networks (Elsevier), 2024. https://www.iris.sssup.it/retrieve/b1eb8cd6-b3c0-4e19-adf5-fd3699b46724/1-s2.0-S1389128624004316-main.pdf

[51] Papadopoulos, K.; Papadimitriou, P.; Papagianni, C. — Deterministic and Probabilistic P4-Enabled Lightweight In-Band Network Telemetry (DLINT/PLINT). IEEE TNSM, 2023. DOI 10.1109/TNSM.2023.3301839

[52] — A Multichannel MAC Protocol without Coordination or Prior Information for Directional Flying Ad hoc Networks. Drones (MDPI) 7(12):691, 2023. https://www.mdpi.com/2504-446X/7/12/691

[53] Temel, S.; Bekmezci, İ. — LODMAC: Location oriented directional MAC protocol for FANETs. Computer Networks (Elsevier), 2015. https://www.researchgate.net/publication/273399929

[54] Garg, S.; Venkatraman, N.; Bentley, E. S.; Kumar, S. — An Asynchronous Multi-Beam MAC Protocol for Multi-Hop Wireless Networks. IEEE ICCCN, 2022. DOI 10.1109/ICCCN54977.2022.9868910. https://arxiv.org/abs/2111.10073

[55] Wang, G.; Qin, Y. — MAC Protocols for Wireless Mesh Networks with Multi-beam Antennas: A Survey. 2019. DOI 10.1007/978-3-030-12388-8_9. https://arxiv.org/abs/1910.00772

[56] Medjo Me Biomo, Jean-Daniel. — MBA-DRR: A Delay-Reducing Routing Protocol for Multi-Beam Directional Antennas in Multi-Hop Ad Hoc Networks. PhD Thesis, Carleton University. https://www.csit.carleton.ca/~msthilaire/Thesis/JeanDaniel%20PhD.pdf

[57] Polese, M.; Restuccia, F.; Melodia, T. — DeepBeam: Deep Waveform Learning for Coordination-Free Beam Management in mmWave Networks. ACM MobiHoc, 2021. DOI 10.1145/3466772.3467035. https://arxiv.org/abs/2012.14350

[58] Chaloun, T.; Boccia, L.; Arnieri, E.; Fischer, M.; Valenta, V.; Fonseca, N. J. G.; Waldschmidt, C. — Electronically Steerable Antennas for Future Heterogeneous Communication Networks: Review and Perspectives. IEEE Journal of Microwaves, 2022. DOI 10.1109/JMW.2022.3202626

[59] Rappaport, T. S.; Xing, Y.; Kanhere, O.; et al. — Wireless Communications and Applications Above 100 GHz: Opportunities and Challenges for 6G and Beyond. IEEE Access, 2019. DOI 10.1109/ACCESS.2019.2921522

[60] Singh, H.; Sneha, H. L.; Jha, R. M. — Mutual Coupling in Phased Arrays: A Review. International Journal of Antennas and Propagation, 2013. DOI 10.1155/2013/348123

[61] Li, M.; Chen, S.-L.; Liu, Y.; Guo, Y. J. — Wide-Angle Beam Scanning Phased Array Antennas: A Review. IEEE Open Journal of Antennas and Propagation, 2023. DOI 10.1109/OJAP.2023.3296636

[62] — Proceedings of the 1990 Antenna Applications Symposium, Volume 2 (modulare Ferrit-Phased-Arrays). DTIC ADA237057. https://apps.dtic.mil/sti/tr/pdf/ADA237057.pdf

[63] — Multifunction Advanced Data Link. Grokipedia. https://grokipedia.com/page/Multifunction_Advanced_Data_Link — nur für unstrittige Plattform- und Herstellerzuordnung (F-35, B-2, Northrop Grumman); physikalisch-technische Aussagen via [Zhao2024], [Simon1994], [Zou2016] abgesichert.

[64] Zhao, J.; Qiao, S.; Booske, J. H.; Behdad, N. — Low-probability of Intercept/Detect (LPI/LPD) Secure Communications Using Antenna Arrays Employing Rapid Sidelobe Time Modulation. arXiv:2406.11229, 2024. https://arxiv.org/abs/2406.11229

[65] Simon, M. K.; Omura, J. K.; Scholtz, R. A.; Levitt, B. K. — Spread Spectrum Communications Handbook (Kap. Low probability of intercept communications). McGraw-Hill, 1994/2002. https://doi.org/10.5860/choice.32-3356

[66] Zou, Y.; Zhu, J.; Wang, X.; Hanzo, L. — A Survey on Wireless Security: Technical Challenges, Recent Advances, and Future Trends. Proceedings of the IEEE, 2016. DOI 10.1109/JPROC.2016.2558521

[67] Karthik, A. K.; Jameer Ali, M. S.; Khan, M. Z. A.; Bhagavathi Rao, A. — A Novel Method for Spectrum Sensing of Linear Modulation Schemes (Cyclostationary Feature Detection). arXiv:2002.03451, 2020. https://arxiv.org/abs/2002.03451

[68] Saggar, H.; Mehra, D. K. — Cyclostationary Spectrum Sensing in Cognitive Radios Using FRESH Filters. arXiv:1312.5257, 2013. https://arxiv.org/abs/1312.5257

[69] Quan, Z.; Shellhammer, S. J.; Zhang, W.; Sayed, A. H. — Spectrum sensing by cognitive radios at very low SNR. IEEE GLOBECOM, 2009. DOI 10.1109/GLOCOM.2009.5426262. https://arxiv.org/abs/0907.1992

[70] Jang, G.; Kim, D.; Lee, I.-H.; Jung, H. — Cooperative Beamforming With Artificial Noise Injection for Physical-Layer Security. IEEE Access, 2023. DOI 10.1109/ACCESS.2023.3252503

[71] Bithas, P. S.; Moustakas, A. L. — Generalized UAV Selection With Distributed Transmission Policies. IEEE Transactions on Communications, 2022. DOI 10.1109/TCOMM.2022.3229665

[72] Khan, S. Z. / Radoš, K. et al. — GNSS Spoofing and Jamming Mitigation: A Comprehensive Review. University of Vaasa / Osuva, 2023. https://osuva.uwasa.fi/bitstreams/70bb2825-1a24-4520-b63a-92007649bb21/download

[73] Egea-Roca, D.; Arizabaleta-Diez, M.; Pany, T.; Antreich, F.; López-Salcedo, J. A.; Paonni, M.; Seco-Granados, G. — GNSS User Technology: State-of-the-Art and Future Trends. IEEE Access 10, 2022. DOI: 10.1109/ACCESS.2022.3165594 — https://doi.org/10.1109/ACCESS.2022.3165594

[74] Radoš, K.; Brkić, M.; Begušić, D. — Recent Advances on Jamming and Spoofing Detection in GNSS. Sensors 24(13):4210, 2024. DOI: 10.3390/s24134210 — https://doi.org/10.3390/s24134210

[75] Sathaye, H.; LaMountain, G.; Closas, P.; Ranganathan, A. — SemperFi: Anti-spoofing GPS Receiver for UAVs. NDSS Symposium, 2022. DOI: 10.14722/ndss.2022.23071 — https://doi.org/10.14722/ndss.2022.23071

[76] RDI (Radiocommunications Agency Netherlands) — GNSS spoofing. 2019. https://www.rdi.nl/site/binaries/site-content/collections/documenten/2019/07/16/gnss-spoofing/GNSS+spoofing.pdf

[77] Gallardo, F.; Pérez Yuste, A. — SCER Spoofing Attacks on the Galileo Open Service and Machine Learning Techniques for End-User Protection. IEEE Access 8, 2020. DOI: 10.1109/ACCESS.2020.2992119 — https://doi.org/10.1109/ACCESS.2020.2992119

[78] Park, S. et al. — Global Navigation Satellite System Spoofing Attack Detection Using Receiver Independent Exchange Format (RINEX) Data and Long Short-Term Memory Algorithm. Information (MDPI) 16(6):502, 2025. https://www.mdpi.com/2078-2489/16/6/502

[79] Jansen, K.; Schäfer, M.; Moser, D.; Lenders, V.; Pöpper, C.; Schmitt, J. — Crowd-GPS-Sec: Leveraging Crowdsourcing to Detect and Localize GPS Spoofing Attacks. IEEE Symposium on Security and Privacy (S&P), 2018. DOI: 10.1109/SP.2018.00012 — https://doi.org/10.1109/SP.2018.00012

[80] Inside GNSS — Q: What has been learned recently about GNSS RF jamming and spoofing events? What tools are available online to track and investigate these events? 2022/2023. https://insidegnss.com/q-what-has-been-learned-recently-about-gnss-rf-jamming-and-spoofing-events-what-tools-are-available-online-to-track-and-investigate-these-events/

[81] Oligeri, G.; Sciancalepore, S.; Ibrahim, O. A.; Di Pietro, R. — GPS spoofing detection via crowd-sourced information for connected vehicles. Computer Networks 216, 2022. DOI: 10.1016/j.comnet.2022.109230 — https://doi.org/10.1016/j.comnet.2022.109230

[82] Sun, Y.; Yu, M.; Wang, L.; Li, T.; Dong, M. — A Deep-Learning-Based GPS Signal Spoofing Detection Method for Small UAVs. Drones (MDPI) 7(6):370, 2023. DOI: 10.3390/drones7060370 — https://doi.org/10.3390/drones7060370

[83] Iqbal, A.; Aman, M. N.; Sikdar, B. — A Deep Learning Based Induced GNSS Spoof Detection Framework. IEEE Transactions on Machine Learning in Communications and Networking 2, 2024. DOI: 10.1109/TMLCN.2024.3386649 — https://doi.org/10.1109/TMLCN.2024.3386649

[84] Wei, X.; Wang, Y.; Sun, C. — PerDet: Machine-Learning-Based UAV GPS Spoofing Detection Using Perception Data. Remote Sensing (MDPI) 14(19):4925, 2022. DOI: 10.3390/rs14194925 — https://doi.org/10.3390/rs14194925

[85] Zhang, J.; Cui, X.; Xu, H.; Lu, M. — A Two-Stage Interference Suppression Scheme Based on Antenna Array for GNSS Jamming and Spoofing. Sensors 19(18):3870, 2019. DOI: 10.3390/s19183870 — https://doi.org/10.3390/s19183870

[86] Sathaye, H.; LaMountain, G.; Closas, P.; Ranganathan, A. — SemperFi: A Spoofer Eliminating GPS Receiver for UAVs. arXiv:2105.01860, 2021. https://arxiv.org/abs/2105.01860

[87] Cognitive Edge Computing: A Comprehensive Survey on Optimizing Large Models and AI Agents for Pervasive Deployment. arXiv:2501.03265, 2025. https://arxiv.org/abs/2501.03265

[88] AI+HW 2035: Shaping the Next Decade. arXiv:2603.05225 (Konsortial-Roadmap, Vertrauen C). https://arxiv.org/abs/2603.05225

[89] TSMC — Logic Technology Roadmap (Hersteller-Roadmap, Vertrauen C). https://www.tsmc.com/english/dedicatedFoundry/technology/logic

[90] IMEC — Introducing 2D-material based devices in the logic scaling roadmap (Vertrauen C). https://www.imec-int.com/en/articles/introducing-2d-material-based-devices-logic-scaling-roadmap

[91] TSMC 1-Trillion-Transistor-Roadmap (Sekundärberichte, Vertrauen C). https://wccftech.com/tsmc-over-1-trillion-transistors-3d-packaged-200-billion-monolithic-chips-2030/

[92] PromptQuorum — MRAM In-Memory Computing 2026 (Sekundärquelle der 640 pJ / 0,9 pJ / 200×-Grössenordnung; deckt sich mit Horowitz, ISSCC 2014; Vertrauen C). https://www.promptquorum.com/local-llms/mram-in-memory-computing-local-ai-2026

[93] Mehonic, A.; Sebastian, A.; Rajendran, B.; Simeone, O.; Vasilaki, E.; Kenyon, A. J. — Memristors — from In-memory computing, Deep Learning Acceleration, Spiking Neural Networks, to the Future of Neuromorphic and Bio-inspired Computing. arXiv:2004.14942, 2020. https://arxiv.org/abs/2004.14942

[94] Dataintelo — Analog In-Memory AI Compute Market Research Report 2034 (Quelle der >2000-TOPS/W-Spitzenwerte; Vertrauen C). https://dataintelo.com/report/analog-in-memory-ai-compute-market

[95] Zhang, F.; Hu, M. — Mitigate Parasitic Resistance in Resistive Crossbar-based Convolutional Neural Networks. arXiv:1912.08716, 2019. https://arxiv.org/abs/1912.08716

[96] Ibrayev, T.; Garg, I.; Chakraborty, I.; Roy, K. — Pruning for Improved ADC Efficiency in Crossbar-based Analog In-memory Accelerators. arXiv:2403.13082, 2024. https://arxiv.org/abs/2403.13082

[97] Ma, S.; Wang, H.; Ma, L.; Wang, L.; Wang, W.; Huang, S.; Dong, L.; Wang, R.; Xue, J.; Wei, F. — The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits. arXiv:2402.17764, 2024. DOI 10.48550/arXiv.2402.17764. https://arxiv.org/abs/2402.17764

[98] Ma, S.; Wang, H.; Huang, S.; Zhang, X.; Hu, Y.; Song, T.; Xia, Y.; Wei, F. — BitNet b1.58 2B4T Technical Report. arXiv:2504.12285, 2025. https://arxiv.org/abs/2504.12285

[99] Wang, J.; Zhou, H.; Song, T.; Mao, S.; Ma, S.; Wang, H.; Xia, Y.; Wei, F. — 1-bit AI Infra: Part 1.1, Fast and Lossless BitNet b1.58 Inference on CPUs. arXiv:2410.16144, 2024. Code: https://github.com/microsoft/BitNet https://arxiv.org/abs/2410.16144

[100] Gu, A.; Dao, T. — Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv:2312.00752, 2023. https://arxiv.org/abs/2312.00752

[101] Ren, R.; Li, Z.; Liu, Y. — Exploring the Limitations of Mamba in COPY and CoT Reasoning. arXiv:2410.03810, 2024. https://arxiv.org/abs/2410.03810

[102] Hasani, R.; Lechner, M.; Amini, A.; Rus, D.; Grosu, R. — Liquid Time-constant Networks. arXiv:2006.04439, 2020. https://arxiv.org/abs/2006.04439

[103] Hasani, R.; Lechner, M.; Amini, A.; Liebenwein, L.; Ray, A.; Tschaikowski, M.; Teschl, G.; Rus, D. — Closed-form Continuous-time Neural Models. arXiv:2106.13898, 2021; publ. Nature Machine Intelligence, 2022. DOI 10.1038/s42256-022-00556-7. https://arxiv.org/abs/2106.13898

[104] Vorbach, C.; Hasani, R.; Amini, A.; Lechner, M.; Rus, D. — Causal Navigation by Continuous-time Neural Networks. arXiv:2106.08314, 2021. https://arxiv.org/abs/2106.08314

[105] Hajizada, E.; Rager, D.; Shea, T.; Campos-Macias, L.; Wild, A.; Hüllermeier, E.; Sandamirskaya, Y.; Davies, M. — Online Continual Learning on Intel Loihi 2 via a Co-designed Spiking Neural Network. arXiv:2511.01553, 2025. https://arxiv.org/abs/2511.01553

[106] Massa, R.; Marchisio, A.; Martina, M.; Shafique, M. — An Efficient Spiking Neural Network for Recognizing Gestures with a DVS Camera on the Loihi Neuromorphic Processor. arXiv:2006.09985, 2020. https://arxiv.org/abs/2006.09985

[107] Polykretis, I.; Michmizos, K. P. — An Astrocyte-Modulated Neuromorphic Central Pattern Generator for Hexapod Robot Locomotion on Intel’s Loihi. arXiv:2006.04765, 2020. https://arxiv.org/abs/2006.04765

[108] Lam Research Newsroom — Silicon Photonics: Powering the Next Revolution in AI (Quelle der 0,05–0,2 pJ/Bit; Vertrauen C). https://newsroom.lamresearch.com/silicon-photonics-ai-energy-efficiency

[109] Jiang, W.; Wang, Y.; Li, Y.; Lin, Y.; Shen, W. — Radar Target Characterization and Deep Learning in Radar Automatic Target Recognition: A Review. Remote Sensing 15(15):3742, 2023. https://doi.org/10.3390/rs15153742

[110] Liu, L.; Ouyang, W.; Wang, X. et al. — Deep Learning for Generic Object Detection: A Survey. IJCV 128:261–318, 2019. https://doi.org/10.1007/s11263-019-01247-4

[111] Mostofi, N. et al. (UMass Amherst) — Detection and Tracking of Drone Swarms using LiDAR (LiSWARM). ACM MobiSys, 2025. https://people.cs.umass.edu/~phuc/papers/2025_MobiSys_LiSWARM.pdf

[112] Morgan, D. A. E. — Deep convolutional neural networks for ATR from SAR imagery. Proc. SPIE 9475, 2015. https://doi.org/10.1117/12.2176558

[113] Li, J.; Yu, Z.; Yu, L.; Cheng, P.; Chen, J.; Chi, C. — A Comprehensive Survey on SAR ATR in Deep-Learning Era. Remote Sensing 15(5):1454, 2023. https://doi.org/10.3390/rs15051454

[114] Huang, Z.; Pan, Z.; Lei, B. — Transfer Learning with Deep CNN for SAR Target Classification with Limited Labeled Data. Remote Sensing 9(9):907, 2017. https://doi.org/10.3390/rs9090907

[115] Li, S.; He, X.; Xu, X.; Zhao, T.; Song, C.; Li, J. — Weapon-Target Assignment Strategy in Joint Combat Decision-Making Based on Multi-Head Deep Reinforcement Learning. IEEE Access 11, 2023. https://doi.org/10.1109/access.2023.3324193

[116] Kong, X.; Zhou, Y.; Li, Z.; Wang, S. — Multi-UAV simultaneous target assignment and path planning based on deep reinforcement learning in dynamic multiple obstacles environments. Frontiers in Neurorobotics 17:1302898, 2024. https://doi.org/10.3389/fnbot.2023.1302898

[117] King, A. — Digital Targeting: Artificial Intelligence, Data, and Military Intelligence. Journal of Global Security Studies 9(2):ogae009, 2024. https://doi.org/10.1093/jogss/ogae009

[118] Kurakin, A.; Goodfellow, I.; Bengio, S. — Adversarial Examples in the Physical World. 2017/2018. https://doi.org/10.1201/9781351251389-8

[119] Gu, T.; Liu, K.; Dolan-Gavitt, B.; Garg, S. — BadNets: Evaluating Backdooring Attacks on Deep Neural Networks. IEEE Access 7, 2019. https://doi.org/10.1109/access.2019.2909068

[120] Inkawhich, N. et al. — Bridging a Gap in SAR-ATR: Training on Fully Synthetic and Testing on Measured Data. IEEE JSTARS 14:2942–2955, 2021. https://doi.org/10.1109/jstars.2021.3059991

[121] Liu, B.; Wang, S.; Li, Q.; Zhao, X.; Pan, Y.; Wang, C. — Task Assignment of UAV Swarms Based on Deep Reinforcement Learning (Ex-MADDPG). Drones 7(5):297, 2023. https://doi.org/10.3390/drones7050297

[122] — A Robust Routing Protocol in Cognitive Unmanned Aerial Vehicular Networks (CNRF). PMC11478788, 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC11478788/

[123] Wang, B.; Li, S.; Gao, X.; Xie, T. — UAV Swarm Confrontation Using Hierarchical Multiagent Reinforcement Learning. International Journal of Aerospace Engineering 2021:3360116. https://doi.org/10.1155/2021/3360116

[124] Kang, H.; Joung, J.; Kim, J.; Kang, J. & Cho, Y. S. (2020): Protect Your Sky: A Survey of Counter Unmanned Aerial Vehicle Systems. IEEE Access 8. DOI 10.1109/access.2020.3023473. https://doi.org/10.1109/access.2020.3023473

[125] Nallamalli, R.; Singh, K. & Kumar, I. D. (2023): Technological Perspectives of Countering UAV Swarms. Defence Science Journal 73. DOI 10.14429/dsj.73.18695. https://doi.org/10.14429/dsj.73.18695

[126] Rogers, J. (2021): Future threats: Military UAS, terrorist drones, and the dangers of the second drone age. JAPCC. https://www.japcc.org/wp-content/uploads/A-Comprehensive-Approach-to-Countering-Unmanned-Aircraft-Systems.pdf

[127] Grachauskas, A. (2026): Modern Warfare Tendencies and Implications for the Baltic Region: Adapting to a New Era of Conflict. Military Science Journal (Mokslo zurnalas) 41(1). DOI 10.47459/mz.2026.41.1.3. https://doi.org/10.47459/mz.2026.41.1.3

[128] Investigation on Electromagnetic Immunity of Unmanned Aerial Vehicles in Electromagnetic Environment. (2025) Electronics 14(21):4332, MDPI. https://www.mdpi.com/2079-9292/14/21/4332

[129] Analysis of High-Power Electromagnetic Pulses Effect on Unmanned Aerial Vehicles. (2026) Drones 10(4):272, MDPI. https://www.mdpi.com/2504-446X/10/4/272

[130] Kim, S.-G.; Lee, E.; Hong, I.-P. & Yook, J.-G. (2022): Review of Intentional Electromagnetic Interference on UAV Sensor Modules and Experimental Study. Sensors 22(6):2384. DOI 10.3390/s22062384. https://doi.org/10.3390/s22062384

[131] Jie, H. et al. (2024): A review of intentional electromagnetic interference in power electronics: Conducted and radiated susceptibility. IET Power Electronics. DOI 10.1049/pel2.12685. https://doi.org/10.1049/pel2.12685

[132] Jafari, A. A. & Anbarjafari, G. (2026): A Multi-physics Simulation Framework for High-power Microwave Counter-unmanned Aerial System Design and Performance Evaluation. arXiv:2602.08477. https://arxiv.org/abs/2602.08477

[133] Min, S.-H. et al. (2021): Analysis of Electromagnetic Pulse Effects Under High-Power Microwave Sources. IEEE Access 9. DOI 10.1109/access.2021.3117395. https://doi.org/10.1109/access.2021.3117395

[134] Christie, L. (2026): Shaping Modern Warfare: The Strategic Role of High-Power Microwave Directed Energy Weapons in Multi-Domain Operations. Defence Science Journal. DOI 10.14429/dsj.21114. https://doi.org/10.14429/dsj.21114

[135] Diehl Defence: Drohnenabwehr | Counter-UAS (SKY SPHERE, SkyWolf, CICADA). https://new.diehl.com/defence/de/produkte/drohnenabwehr-counter-uas

[136] Karkadakattil, A. (2026): Laser-Based Directed Energy Weapons: Technological Capabilities, Material Interaction, and Strategic Deployment Pathways. Defence Science Review (PNO). DOI 10.37055/pno/216776. https://doi.org/10.37055/pno/216776

[137] Counter-Unmanned Aerial Vehicles Study: Shipboard Laser Weapon System Engagement Strategies for Countering Drone Swarm Threats. NPS/DTIC AD1165019. https://apps.dtic.mil/sti/trecms/pdf/AD1165019.pdf

[138] Siora, O.; Lukashenko, V. & Bernatskyi, A. (2025): An interdisciplinary study of the effect of laser radiation on carbon fiber-reinforced polymer, in the context of counteracting unmanned aerial vehicles. History of Science and Technology 15(1). DOI 10.32703/2415-7422-2024-15-1-195-215. https://doi.org/10.32703/2415-7422-2024-15-1-195-215

[139] MBDA Deutschland: Hochenergie-Lasereffektor. https://www.mbda-deutschland.de/produkte/marine/hochenergie-lasereffektor/

[140] produktion.de (2024): Laserwaffen: System von Rheinmetall und MBDA vor Marktreife. https://www.produktion.de

[141] Electro Optic Systems: Apollo High Energy Laser Weapon. Broschüre 2025. https://eos-aus.com/wp-content/uploads/2025/08/Apollo-brochure-AUS-Aug2025_web.pdf

[142] Gupta, T.: The role of High Energy Laser as a U.S. Army Counter-small Unmanned Aircraft System (UAS) Weapon. DTIC AD1230685. https://apps.dtic.mil/sti/trecms/pdf/AD1230685.pdf

[143] Michnewich, D. A. (2018): Modeling Energy Storage Requirements for High-Energy Lasers on Navy Ships. NPS. http://hdl.handle.net/10945/59554

[144] Yan, H.; Yang, K.; Cheng, Y.; Wang, Z. & Li, D. (2024): Precise Interception Flight Targets by Image-based Visual Servoing of Multicopter. arXiv:2409.17497. DOI 10.48550/arxiv.2409.17497. https://arxiv.org/abs/2409.17497

[145] Liu, L. Y.; Yang, K.; Zou, H. et al. (2026): Planar-Sector LOS Guidance for Interception of Agile Targets with Lifting-Wing Quadcopters. arXiv:2606.10639. https://arxiv.org/abs/2606.10639

[146] DLR (2025): Reliable drone defence (CUSTODIAN-Projekt). https://www.dlr.de/en/latest/news/2025/reliable-drone-defence

[147] Brust, M. R.; Danoy, G.; Stolfi, D. H.; Bouvry, P. (2021): Swarm-based counter UAV defense system. Discover Internet of Things 1. https://doi.org/10.1007/s43926-021-00002-x

[148] Chipade, V. S.; Wang, X.; Panagou, D. (2021): IDCAIS: Inter-Defender Collision-Aware Interception Strategy against Multiple Attackers. arXiv:2112.12098. https://arxiv.org/abs/2112.12098

[149] Rheinmetall: Oerlikon Skynex Air Defence System (35 mm AHEAD). Broschüre B200e0424. https://www.rheinmetall.com

[150] Cumpson, P. J. (2026): Movement To Contact and Vanguard UAVs: Strategies for Swarm UAV Battlefield Economics in the era of Microwave Directed Energy Weapons. Security and Defence Quarterly. DOI 10.35467/sdq/221147. https://doi.org/10.35467/sdq/221147

[151] LiDAR Technology for UAV Detection: From Fundamentals and Operational Principles to Advanced Detection and Classification Techniques. (2025) Sensors 25(9):2757, MDPI. https://www.mdpi.com/1424-8220/25/9/2757

[152] Abir, T. A.; Le, V.; Kuantama, E. et al. (2025): Detection and Tracking of Drone Swarms using LiDAR (LiSWARM). ACM MobiSys. DOI 10.1145/3711875.3729156. https://doi.org/10.1145/3711875.3729156

[153] Cost-Effectiveness Analysis of Counter-UAS Technologies: A Comparative Study of Kinetic, EW, and Directed Energy Countermeasures (2022–2026). ResearchGate-Preprint (Multi-Layered Defense Economics Model), nicht peer-reviewed. https://www.researchgate.net/publication/401707891

[154] Bächle, T. C., Bareis, J. (2022): “Autonomous weapons” as a geopolitical signifier in a national power play: analysing AI imaginaries in Chinese and US military policies. European Journal of Futures Research 10:20. https://doi.org/10.1186/s40309-022-00202-w

[155] CCW Group of Governmental Experts (2019): Guiding Principles affirmed by the GGE on Emerging Technologies in the Area of LAWS, CCW/MSP/2019/9 Annex III. https://ccdcoe.org/uploads/2020/02/UN-191213_CCW-MSP-Final-report-Annex-III_Guiding-Principles-affirmed-by-GGE.pdf

[156] Amoroso, D. (2020): A Legality “Test” for Autonomous Weapons Systems. The (In)compatibility of Autonomous Targeting with IHL and IHRL. In: Autonomous Weapons Systems and International Law, Nomos, S. 31–120. https://doi.org/10.5771/9783748909538-31

[157] Zhu, L., Hu, X., Han, Y. (2025): The Status of Meaningful Human Control of Lethal Autonomous Weapons System in International Humanitarian Law. Brawijaya Law Journal 12(2):206–228. https://doi.org/10.21776/ub.blj.2025.012.02.02

[158] Canellas, M., Haga, R. (2015): Toward Meaningful Human Control of Autonomous Weapons Systems through Function Allocation. IEEE ISTAS 2015. https://doi.org/10.31228/osf.io/uvwd9

[159] Cummings, M. L. (2019): Lethal Autonomous Weapons: Meaningful Human Control or Meaningful Human Certification? IEEE Technology and Society Magazine 38(4):20–26. https://doi.org/10.1109/mts.2019.2948438

[160] ICRC (2020): ICRC commentary on the ‘Guiding Principles’ of the CCW GGE on LAWS. UNODA. https://documents.unoda.org/wp-content/uploads/2020/07/20200716-ICRC.pdf

[161] UN-Generalversammlung (2023): Resolution A/RES/78/241 — Lethal autonomous weapons systems, 22.12.2023. https://digitallibrary.un.org/record/4033027 · PDF: https://digitallibrary.un.org/nanna/record/4033027/files/A_RES_78_241-EN.pdf

[162] Longpre, S., Storm, M., Shah, R. (2022): Lethal autonomous weapons systems & artificial intelligence: Trends, challenges, and policies. MIT Science Policy Review 3. https://doi.org/10.38105/spr.360apm5typ

[163] Futter, A., Zala, B. (2021): Strategic non-nuclear weapons and the onset of a Third Nuclear Age. European Journal of International Security 6(3). https://doi.org/10.1017/eis.2021.2

[164] Johnson, J. (2021): Inadvertent escalation in the age of intelligence machines: A new model for nuclear risk in the digital age. European Journal of International Security 6(4). https://doi.org/10.1017/eis.2021.23

[165] Williams, H. (2019): Asymmetric arms control and strategic stability: Scenarios for limiting hypersonic glide vehicles. Journal of Strategic Studies 42(6):789–813. https://doi.org/10.1080/01402390.2019.1627521

Parameter	Sensonor STIM300	GuideNav GUIDE730	Adv. Navigation Motus	Safran Colibrys MS1000
Gyro bias stability	0.3 °/h	0.2 °/h	high-precision (n/a)	— (accelerometer)
Acc bias stability	—	—	—	15–24 µg
Volume / dimensions	compact	28 × 28 × 10 mm	~16 cm³	9 × 9 mm (chip)
Weight	compact	15 g	26 g	< 1 g (chip)
Power	relatively high	< 0.5 W	1.4 W	~10 mW
Primary application	tactical air/sea	UAV/robotics (SWaP)	SWaP-C UAV	AHRS, 6,000 g, MWD

Method	Dataset/Data basis	Detection	Source
PCA-CNN-LSTM	real UAV flights with jammer	0.9949 (10-fold CV)	[82]
VAE + GAN (one-class)	TEXBAT DS-7 (subtle)	~95% vs. 44.1% supervised	[83]
PerDet (sensor fusion)	real flight data	99.69%	[84]

Category	Examples	Cost-per-Engagement (estimated)	Assessment Against Mass Swarms
Classic guided missiles	Patriot PAC-3, IRIS-T SLM	1.0–4.75 million USD	Highest hit probability, economically untenable against mass
Attacker drones	Shahed derivatives, FPV	300–80,000 USD	FPV cost-benefit ratio estimated up to 250,000:1
Interceptor drones	MaXon, MARSS, CICADA	~3,500 USD	Inverts the cost curve; scalable magazine depth
High-energy lasers	Rheinmetall/MBDA, Apollo	<10 USD (energy only)	Lowest per-shot cost, limited by sequentiality
AHEAD / anti-aircraft gun	Skynex 35 mm	several thousand USD/burst	Highly effective against bunched swarms, best value for money at close range
Electronic warfare	RF jammer	~0.01 USD	Virtually free of charge, ineffective against autonomous latest generation

Abbreviation	Meaning
AGC	Automatic Gain Control
AHEAD	Advanced Hit Efficiency And Destruction (programmable fragmentation munition)
ARW	Angular Random Walk (gyroscope noise measure)
ATR	Automatic/Aided Target Recognition
AUV	Autonomous Underwater Vehicle
C/N0	Carrier-to-Noise-density ratio
CEP	Circular Error Probable
CFC	Closed-form Continuous-time (neural networks)
CFET	Complementary Field-Effect Transistor (vertically stacked transistor)
CFRP	Carbon-Fiber-Reinforced Polymer
CNRF	Central Node Resolution Factor (cluster-head election metric in a FANET)
CRLB	Cramér-Rao Lower Bound (lower bound on estimator variance)
CRPA	Controlled Reception Pattern Antenna (adaptive anti-jam antenna array)
DEM / DTED	Digital Elevation Model / Digital Terrain Elevation Data
DEW	Directed Energy Weapon
EKF	Extended Kalman Filter
EO/IR	Electro-Optical / Infrared
EPnP	Efficient Perspective-n-Point (pose estimator)
ESC	Electronic Speed Controller
EW	Electronic Warfare
F2T2EA	Find, Fix, Track, Target, Engage, Assess (kill chain)
FANET	Flying Ad-Hoc Network
FDE	Fix Displacement Error (position error of pure inertial navigation)
FINT	FANET In-Band Telemetry (programmable swarm telemetry)
FOG	Fiber-Optic Gyroscope
FPV	First-Person View
GAAFET	Gate-All-Around Field-Effect Transistor
GNSS	Global Navigation Satellite System
HEL	High-Energy Laser
HPM / HPEM	High-Power Microwave / High-Power Electromagnetics
ICP	Iterative Closest Point (point-cloud registration algorithm)
IEMI	Intentional Electromagnetic Interference
IMU	Inertial Measurement Unit
INS	Inertial Navigation System
LAWS	Lethal Autonomous Weapon Systems
LiDAR	Light Detection and Ranging
LNN	Liquid Neural Network
LODMAC	Location Oriented Directional MAC (position-based directional radio protocol)
LPI/LPD	Low Probability of Intercept/Detection
LSTM	Long Short-Term Memory (recurrent neural network)
MAC	Media Access Control
MADL	Multifunction Advanced Data Link (fifth-generation stealth data link)
MARL	Multi-Agent Reinforcement Learning
MBAA	Multiple-Beam Antenna Array (MAC protocol)
MDP / POMDP	(Partially Observable) Markov Decision Process
MEMS	Micro-Electro-Mechanical Systems
MHC	Meaningful Human Control
MILO	Magnetically Insulated Line Oscillator (HPM source)
MRAM	Magnetoresistive Random-Access Memory (non-volatile memory)
MWIR	Mid-Wavelength Infrared
NIC	Navigation Integrity Category (integrity measure in ADS-B)
OODA	Observe–Orient–Decide–Act (Boyd’s decision loop)
OSNMA	Open Service Navigation Message Authentication (Galileo signal authentication)
P4	Programming Protocol-independent Packet Processors (data-plane language)
PIM / CIM	Processing-in-Memory / Compute-in-Memory
PNT	Positioning, Navigation and Timing
PWM	Pulse-Width Modulation
RCS	Radar Cross-Section
RINEX	Receiver Independent Exchange Format (raw GNSS data format)
RLG	Ring Laser Gyroscope
RNP	Required Navigation Performance
RSSI	Received Signal Strength Indicator
SAR	Synthetic Aperture Radar
SCER	Secure Code Estimation and Replay (advanced spoofing class)
SDR	Software-Defined Radio
SGM	Semi-Global Matching (stereo disparity method)
SIC	Successive Interference Cancellation
SIGINT	Signals Intelligence
SNN	Spiking Neural Network
SoC	System-on-Chip
SSM	State Space Model (e.g., Mamba)
SVM	Support Vector Machine
SWaP-C	Size, Weight, Power, and Cost
TERCOM	Terrain Contour Matching
TOPS/W	Tera-Operations per Second per Watt (compute efficiency)
TRL	Technology Readiness Level
TRN	Terrain Referenced Navigation
UAV / UAS	Unmanned Aerial Vehicle / System
UMA	Unified Memory Architecture
VIO	Visual-Inertial Odometry
WTA	Weapon-Target Assignment
AP I	Additional Protocol I to the Geneva Conventions (1977)