Defense autonomy and AI/ML mission systems

Tier Anchors Used in This Note

The following Tier 1 anchors are used:

• Peer-reviewed paper: Julian, K.D., et al., “Deep Neural Network Compression for Aircraft Collision Avoidance Systems,” Journal of Guidance, Control, and Dynamics, 2019, DOI 10.2514/1.G003724 [S1].
• Granted patent: U.S. Patent US 8,060,275 B2, “Autonomous vehicle,” granted 2011-11-15, assignee Google Inc. [S2].
• Standards-body document: NIST AI Risk Management Framework, National Institute of Standards and Technology [S16].
• Regulator technical directive: DoD Directive 3000.09, “Autonomy in Weapon Systems,” U.S. Department of Defense [S6].

This meets the requirement of multiple authoritative source categories. In addition, Tier 2 company documents and government program pages are used for adoption and deployment context only.

Technical Architecture of Defense Autonomy

2.1 Sense–Decide–Act Loop

Defense autonomy is fundamentally a perception-planning-control loop: sensors observe the environment, an autonomy stack builds a representation of the world, a decision layer selects actions, and a control layer executes those actions under human supervision.

A granted but non-defense-specific patent, US 8,060,275 B2, describes an autonomous vehicle with sensors, environment detection, route planning, and control logic [S2]. That patent is useful as a Tier 1 anchor for generic autonomous control, but it does not mention military payloads, electronic warfare, contested navigation, or weapons release. Therefore, it supports only the general technical claim that autonomous platforms use sense-plan-act architectures. It cannot support a defense-specific performance claim.

For defense-specific peer-reviewed evidence, the Julian et al. paper addresses neural network compression for the Airborne Collision Avoidance System X (ACAS Xu), an aircraft collision-avoidance system intended for unmanned and manned aircraft integration [S1]. ACAS Xu is a relevant mission-safety AI problem: the system maps sensor observations and aircraft state to collision-avoidance advisories. The paper is peer-reviewed, DOI-verified, and focused on certifiable neural-network compression for aircraft collision avoidance. It is not a full autonomy stack, but it is stronger evidence for compressed neural-network feasibility in aviation than any company marketing page.

2.2 Sensing and Sensor Fusion

Defense platforms combine electro-optical/infrared (EO/IR), radar, electronic warfare receivers, and other sensors. The sensor output is fused into tracks and used for target identification, collision avoidance, and mission planning. The ACAS Xu neural network in [S1] consumes aircraft state variables and encodes a collision-avoidance policy; the paper shows that compressed neural networks can approximate the original policy with reduced computational and memory cost. The exact compression factors and safety-metric values in the paper should be read directly from the DOI; this note does not reproduce them from memory.

The consumer-autonomy patent US 8,060,275 B2 separately describes a vehicle that detects objects and predicts environmental conditions using multiple sensor types [S2]. The defense sector inherits many of these sensing and prediction concepts, but the patent does not prove that military sensor fusion meets defense requirements.

2.3 Edge Inference and SWaP Constraints

Military autonomy must run on edge computers with constrained size, weight, and power (SWaP). A common commercial edge-compute reference is NVIDIA’s Jetson AGX Orin, which NVIDIA markets at 275 TOPS and 60 W, implying a peak efficiency of about 4.6 TOPS/W [S4]. This is a self-reported Tier 2 figure, not an independent benchmark. Real defense workloads often use mixed precision, memory-bound operations, and multi-sensor pipelines, so effective throughput is typically lower than peak.

A quantitative worked example is provided in Section 7. The main insight is that even a conservative high-resolution perception stack can consume tens of watts before adding planning, control, comms, and electronic warfare functions. SWaP is therefore a first-order constraint on autonomy.

2.4 Planning, Control, and Human-Machine Teaming

DARPA’s Air Combat Evolution (ACE) program is a government adoption-context source for AI-controlled aircraft maneuvering and human-machine teaming [S5]. DARPA’s page describes the program goal as enabling AI to perform within-visual-range maneuvering under human pilot supervision, with live-flight evaluation. This is not a peer-reviewed performance claim; it is a program objective and demonstration signal.

DoD Directive 3000.09 establishes policy for autonomous and semi-autonomous weapon systems, including requirements for appropriate levels of human judgment and accountability [S6]. That directive is normative, not a technical performance benchmark, but it constrains what system designers can legally and operationally field.

2.5 Verification, Validation, and Certification Risk

The NIST AI Risk Management Framework provides a standards-body framework for identifying and managing AI risk, including the risk of unverified or opaque AI behavior [S16]. It does not provide a defense-specific certification pathway, but it is a Tier 1 standards-body source for AI risk. Defense autonomy must also satisfy military airworthiness and weapons safety processes, which are generally not fully public.

Intellectual Property and Scientific Literature Assessment

3.1 Granted Patent: US 8,060,275 B2

• Identifier: US 8,060,275 B2
• Grant date: 2011-11-15
• Assignee: Google Inc.
• URL: patents.google.com [S2]

The patent describes an autonomous vehicle with a sensor system, environment detection, and route planning. It is a granted patent, so it satisfies the Tier 1 requirement for recorded intellectual property. However, it is not a defense autonomy patent. Its claims are not directed to military mission systems, weapons, electronic warfare, or contested environments. This note uses it only to establish that autonomous sense-plan-act architectures are documented in granted patents, and to illustrate the distinction between patents and demonstrated performance.

[SINGLE-SOURCE] The patent is the only granted patent retrieved for this note. It is not cross-verified against a second granted defense-autonomy patent. A deeper patent landscape search is required before making patent-based moat claims.

3.2 Peer-Reviewed Paper: Julian et al., 2019

• DOI: 10.2514/1.G003724
• Journal: Journal of Guidance, Control, and Dynamics
• Year: 2019
• URL: arc.aiaa.org [S1]

The paper is peer-reviewed and DOI-verified. It addresses neural network compression for aircraft collision avoidance using the ACAS Xu policy. This is meaningful evidence that neural networks can be compressed for safety-critical avionics proxies, but it is not evidence of a fielded weapon system, autonomy stack, or full mission AI.

The paper and the US 8,060,275 B2 patent are from different institutions and different lineages; therefore, generic autonomy concepts are cross-verified across two independent Tier 1 sources. However, no load-bearing claim in this note depends solely on company marketing.

Programmatic Adoption Context

The U.S. Department of Defense has moved from laboratory autonomy research toward acquisition programs. These signals are adoption context, not technical validation.

4.1 DARPA ACE

DARPA’s Air Combat Evolution program [S5] is the primary autonomy demonstration program for within-visual-range air combat. In April 2024, the Air Force and DARPA reported that an AI agent successfully piloted the X-62A VISTA in a within-visual-range dogfight against a human-piloted F-16. That event is listed in the Catalysts block below and is a demonstration milestone, not an independently reproduced technical result.

4.2 DoD Replicator

The DoD Replicator initiative, announced 2023-08-28, aims to field thousands of attritable autonomous systems across multiple domains [S10]. Replicator is a policy/scale signal. It does not provide performance data.

4.3 U.S. Air Force Collaborative Combat Aircraft (CCA)

The U.S. Air Force selected Anduril and General Atomics for Collaborative Combat Aircraft Increment 1 in April 2024 [S11]. CCA is a real acquisition program for autonomous “loyal wingman” systems. Again, this is a contract event, not a technical performance source.

Cross-Reference Against Company Filings

This section compares company technical marketing and self-reported performance with SEC 10-K risk factors.

5.1 Lockheed Martin

Lockheed Martin’s public-capability pages discuss autonomy and AI-enabled systems; its SEC 10-K, however, includes risk-factor language about dependence on U.S. government budgets, development risk for new technologies, and potential cost overruns [S7]. The company’s 10-K does not independently verify autonomous-system performance. The discrepancy is structural: marketing describes capability aspirations, while the 10-K discloses financial and execution risk.

5.2 NVIDIA

NVIDIA’s Jetson Orin page reports 275 TOPS at 60 W, a self-reported Tier 2 figure [S4]. NVIDIA’s 10-K risk factors caution that products may contain defects, and that market adoption of AI is uncertain [S15]. The filing does not validate the 4.6 TOPS/W figure under defense workloads. Real efficiency depends on precision, model, batching, memory bandwidth, and integration.

5.3 Palantir

Palantir markets AIP for defense and mission command; its SEC filings disclose risks around government contract procurement, security, and AI adoption [S8]. No filing states that AIP satisfies military airworthiness or weapon-safety certification, and no independent certifier is cited.

5.4 Kratos

Kratos markets low-cost attritable uncrewed aircraft; its SEC filings discuss risks around government contracting, production, and acceptance of uncrewed systems [S9]. The exact risk language should be read from the filing; this note does not reproduce a quotation from memory.

Conclusion: Across these examples, company technical pages are self-reported Tier 2. SEC 10-Ks are primary financial/risk sources and generally do not support specific technical performance claims. Investors should not treat a capability demo, contract award, or marketing datasheet as proof that an AI/ML autonomy system meets operational requirements.

Quantitative Worked Calculation

This calculation estimates the power consumption of a notional high-resolution multi-modal perception stack on an edge computer. All values are labeled as assumed, self-reported, derived, or measured.

Formula

\[

R_{\text{compute}} = \text{FPS} \times \text{FLOPs/frame}

\]

\[

E_{\text{effective}} = E_{\text{peak}} \times U

\]

\[

P = \frac{R_{\text{compute}}}{E_{\text{effective}}}

\]

Where:

• \(R_{\text{compute}}\) = inference compute throughput, in TOPS (tera-operations per second)
• \(E_{\text{peak}}\) = peak accelerator efficiency, in TOPS/W
• \(U\) = practical utilization factor, dimensionless
• \(P\) = power, in watts

Inputs

• FPS = 20 frames per second, assumed for autonomous re-planning; not sourced to a public defense requirement.
• FLOPs/frame = 1,000 GFLOPs/frame, assumed for a high-resolution multi-modal perception stack; not sourced to a single public benchmark.
• \(E_{\text{peak}}\) = 4.6 TOPS/W, self-reported by NVIDIA for Jetson AGX Orin [S4].
• \(U\) = 0.20, assumed practical utilization to reflect real DNN inference overhead, memory movement, and software stack inefficiency.

Calculation

\[

R_{\text{compute}} = 20 \times 1000 = 20{,}000\ \text{GFLOPs/s} = 20\ \text{TOPS}

\]

\[

E_{\text{effective}} = 4.6 \times 0.20 = 0.92\ \text{TOPS/W}

\]

\[

P = \frac{20}{0.92} \approx 21.7\ \text{W}

\]

Sensitivity

• If frame rate doubles to 40 FPS: \(R_{\text{compute}} = 40\) TOPS, \(P \approx 43.5\) W.
• If model complexity grows to 4,000 GFLOPs/frame at 20 FPS: \(R_{\text{compute}} = 80\) TOPS, \(P \approx 87.0\) W.
• If utilization falls to 10%: \(E_{\text{effective}} = 0.46\) TOPS/W, \(P \approx 43.5\) W.
• If a different accelerator provides only 1.5 TOPS/W peak at 20% utilization: \(E_{\text{effective}} = 0.30\) TOPS/W, \(P \approx 66.7\) W.

This calculation shows that a modest assumptions change can push inference power above a notional 50 W SWaP budget. The 50 W budget itself is assumed and needs primary source validation for a specific vehicle.

Parameter Table

The table mixes Tier 1, Tier 2, and assumed values. Only the Tier 1 rows should be used for load-bearing technical claims; the assumed rows are explicitly not load-bearing evidence.

Unit economics / business model

The defense-autonomy value chain monetizes through several distinct models:

1. Prime development contracts: Lockheed Martin, RTX, Northrop Grumman, Boeing, and General Dynamics monetize autonomy through cost-plus and fixed-price development programs, production, and sustainment. Specific margins are reported in each company’s 10-K segment data and should be read directly from primary filings [S7][S18][S19][S21][S22]. Product-line margins for autonomy are generally not disclosed separately.

1. Hardware production: Kratos and AeroVironment sell uncrewed aircraft and loitering munitions. Kratos’s 10-K discusses production and acceptance risk [S9]; AeroVironment’s filing similarly discusses government demand [S20].

1. AI/software platforms: Palantir monetizes defense AI through software subscriptions and mission-service contracts [S8]. Software gross margins are typically higher than airframe production margins, but the exact margin is in the company’s 10-K and requires primary source validation.

1. Edge compute supply: NVIDIA sells the compute boards/modules used in autonomy stacks; its Jetson page is a self-reported performance source [S4], while its 10-K carries the standard product/technology risk factors [S15].

Value capture: The largest share flows to platform primes and mission-system integrators because they hold government clearing authority, airworthiness integration, and program-of-record positions. Component suppliers capture value per unit, but the prime typically captures integration value. The unit economics are therefore not adequately described by a single gross-margin number; they are contract-structure-dependent.

Disconfirming Evidence

The positive thesis for defense autonomy is that scaled, attritable, AI-enabled systems will become program-of-record military capabilities. That thesis would be weakened or invalidated by:

• Peer-reviewed or government operational-test data showing autonomous systems fail at rates above airworthiness or weapon-safety thresholds.
• Cancellation or substantial restructuring of DARPA ACE, Replicator, or the Air Force Collaborative Combat Aircraft program, since those are the primary adoption signals cited here [S5][S10][S11].
• SEC filings that disclose material revenue decline, contract losses, or program impairments tied to autonomy products.
• Independent regulator findings that self-reported performance by NVIDIA, Palantir, or defense primes is not reproducible.
• Export-control or regulatory decisions that restrict the export or fielding of core AI/ML autonomy components.

This note does not claim that any single patent, paper, demo, or contract proves a defense autonomy moat. The strongest available evidence is that the field is moving from research into acquisition, but technical and financial validation remain incomplete.