Forecasting the Drone Future: Keirin Joyce’s Doctoral Thesis on Technology Readiness

A UNSW doctoral thesis by an RAAF officer focusing on how to turn drone technology forecasting from expert opinion into a repeatable method, and hands the ADF a three-way decision rule: acquire, invest, or wait.

Keirin Joyce’s doctoral thesis, First Person View to Autonomous Collaborative Platforms: Forecasting the maturity of aerial drone technologies (UNSW, August 2026, supervised by Keith Joiner, Vince Daria and John Young), lands at a moment when everyone in the Australian defence debate is talking about drones and very few people are disciplined about when a given drone capability will actually be ready. That is the gap Keirin goes after.

I respect his work and have followed it for some time, and precisely because I do, I want to engage with it seriously rather than simply applaud it. Page references below are to the thesis’s printed page numbers.

The full text is open access on UNSWorks.

The question and the method

The thesis asks two linked questions: “Can informed bibliometric activity mapping and analysis provide an effective means of forecasting UAS technology readiness?” and “What does an optimised UAS program look like for the Australian Defence Force?” (p. 3).

The method works like this. Decompose a platform into its enabling sub-systems rather than grading it as a single object. For each sub-system, use AI-assisted citation chaining in Litmaps, with Google Scholar as the citation source, to find the year a technology reached critical mass of high-impact research. Triangulate against industry engagement: patents, demonstrators, standards, regulatory milestones, and conversations with chief technology officers. Then add a lag drawn from historical analogues, and a penalty where regulation and certification are immature. The formula is stated plainly: TOR = T(breakthrough) + Lag, where TOR is “Time to Operational Reality” (p. 29).

Three case studies test it against progressively harder terrain. FPV strike drones are an open-source, bottom-up ecosystem with a visible citation trail. Collaborative combat aircraft are a closed, state-controlled domain where, as Keirin says, citation counts lose their value as a proxy for readiness. Large logistics UAS sit where commercial parcel delivery meets military supply. Each case is an embedded paper that went through peer review or conference scrutiny (Defense & Security Analysis; the Osaka intelligent and autonomous systems conference; Technologies 14, 155), and a further paper on regulatory barriers is enclosed as an appendix.

What the thesis gets right

Platforms do not mature; sub-systems do, and at different speeds. This is the core insight and it is a good one. The Ghost Bat airframe is mature while the machine-learned tactical choice models and manned-unmanned teaming loops are not (p. 132). In logistics, commercial routing optimization is mature while liquid-hydrogen powertrains and GPS-denied landing-zone perception are not. A single readiness score for the whole system hides exactly the bottleneck a program manager most needs to see.

Regulation is treated as a variable, not a footnote. Keirin’s “invisible infrastructure penalty” takes seriously what engineers routinely ignore. The airworthiness paradox (holding an attritable drone to a 10^-9 failure standard designed for airliners) and the software-validation problem for non-deterministic machine learning are real gates, and the thesis puts years on them (p. 21). His Table 1 has civilian certification of the relevant technologies running to a worst case of 2040, with military-controlled airspace compressing that considerably. Whatever one thinks of the arithmetic, putting a number on regulatory lag forces a conversation Defence needs to have.

The FPV case is a useful corrective to vendor claims. The thesis traces human-in-the-loop FPV control back to the Mellinger and Mahony work of 2011-12, through Betaflight in 2015, to Ukraine in 2022, and shows the science was settled long before the battlefield needed it. Autonomy is a different story. Keirin judges the claimed 80 percent terminal-guidance hit rates to be probably “bootleg” systems that mimic pilot input rather than robust integrated autonomy, and puts robust autonomous drone racing technology three to five years from scale adoption (p. 52). In a market full of inflated autonomy claims, a method for separating a genuine step from a hacked one has value.

It is candid about its own limits. Chapter 4 carries a reflexive critique of how the method must change when it moves from open ecosystems to closed ones (p. 63). Chapter 7 concedes that accuracy is bounded by public literature, that wartime imperatives and state capital can compress timelines unpredictably, and that the penalty factors depend on “careful qualitative calibration” (p. 123). It also offers falsification indicators, such as publication of the Autonomy Government Reference Architecture or a surge in open-source citations (p. 125). That is the right scientific instinct.

The decision rule is usable. Acquire when sub-system maturity, manufacturing readiness and the regulatory footprint support it. Invest, with open eyes, when the strategic weight justifies pushing ahead of the technology, which is how the thesis reads the Ghost Bat. Wait when sub-system and regulatory hurdles are severe, but use the waiting period to fund bounded experimentation, which is how it reads Project Camel Train and the ASCA mechanisms (pp. 133-134, 138). Phase 1 would require Gate Zero and Smart Buyer reviews to decompose platforms and score sub-systems separately. The recommendation that waiting is an active posture, not an idle one, is the most practically valuable thing in the document.

Where I would push

The evidence does not yet support the language of “empirical science.” Chapter 7 says the framework shifts forecasting “from an exercise in subjective speculation towards being an empirical science” (p. 123). Look at the published prediction tables, though. In every row, the prediction equals the lag plus the years that cannot be run concurrently: 8 + 2 = 10 for hydrogen-electric propulsion, 5 + 2 = 7 for autonomous landing-zone sensing, 3 + 2 = 5 and 8 + 6 = 14 for military and civilian certification.

That is a transparent and sensible structure, but it rests on lag and concurrency estimates that are heuristic judgments, delivered as single years with no uncertainty ranges. The lag itself is calibrated largely on one analogue (the three-to-four-year Betaflight lag) and then reused to project autonomous drone racing (pp. 62-63). The FPV paper was, by the thesis’s own account, written “before the TOR methodology was formalised” and then mapped to it afterward (p. 62). The two forward-looking cases cannot yet be tested. This is a disciplined structured-judgment method with an audit trail, which is valuable, but “validated” (p. 122) is stronger than three cases, one of them retrospective, can bear.

The cure is the one Keirin’s own falsification indicators point to: publish dated predictions and revisit them.

Ready is not the same as adopted. The FPV story shows the science was settled by 2015 and the battlefield adopted it in 2022 because a war created the demand, not because a clock ran out. The thesis acknowledges that crisis compresses the lag and treats this as stochastic disruption. For adoption it may be the dominant variable.

A readiness forecast tells you when a technology can be absorbed. It does not tell you when an institution will absorb it, and the gap between the two is where a great deal of defence failure lives.

The logistics verdict may conflate a hard concept with a whole capability class. The “justified deferral” of large logistics UAS (p. 133) leans heavily on the Dragonfly concept: liquid hydrogen, roughly 8 MW of hover power, 4,500 kg over 1,000 km (pp. 110-111). The thesis is honest that Dragonfly is “not a blueprint for immediate production,” and its own conclusion notes that optionally piloted conversions such as K-MAX and U-Hawk are the nearer-term path (p. 114).

But heavy sustainment capability need not wait on the hardest architecture. Optionally crewed conversions, existing commercial airframes automated as in Camel Train, and air-launched delivery all route around the powertrain and landing-zone bottlenecks. The recent Jericho Dawn work on the Aladdin air-delivery drone launched from a C-130J is a case in point. A forecast for the hardest archetype is not a forecast for the capability.

The model’s formula is narrower than its literature review. Chapter 2 makes a persuasive case that TRL alone is inadequate and that system, manufacturing and production readiness (SRL, MRL, PRL) all matter. Then it declares TRL “the ultimate, primary metric” for the predictive model (p. 17), and the formula operates on sub-system TRLs plus lag plus penalty. Integration, manufacturing and sustainment appear in the narrative, in the reverse-spiral analysis and in the MRL 7+ threshold for the Acquire tier, but not in the arithmetic.

The tool is strongest where the adversary is least hidden. Chapter 7 argues the method can reduce asymmetric surprise by detecting sub-system breakthroughs early (pp. 131-132). Yet Chapter 4 itself concedes that the approach is most effective in open ecosystems and loses power in closed, state-controlled ones (p. 63). A strategic competitor’s most consequential work sits in the latter. The recommendation to automate the method with large language models (p. 140) raises a related risk: a system that ingests the public record also ingests the hype, and the public record can be shaped.

Through the kill web

What strikes me most is how closely the thesis’s conclusion converges with arguments I have been making from a very different direction. Keirin writes that “military dominance will no longer be measured by the possession of a finite fleet of exquisite, multi-million-dollar crewed platforms, but by an organisation’s systemic agility to continuously iterate, scale, and field software-defined autonomous capabilities” (p. 141). That is the contest between intelligent mass and exquisite scarcity, reached through procurement mechanics and technology timing rather than through operational concepts. Two routes arriving at the same place is reassuring.

The reverse spiral is the thesis’s most original contribution to that convergence, and it maps directly onto what I took from the Williams Foundation’s October 1 seminar on air combat mass and depth: the next generation of airpower is an enterprise, not a platform.

Keirin shows how software-defined systems rewritten at the edge fragment configuration baselines, undermine the economic assumptions of logistics modelling, and push sustainment into permanent reaction. His arithmetic of partial upgrades is simple and devastating: fund half the fleet at each increment and the baselines multiply (p. 131).

The remedy he proposes, modular open systems and contractually secured software interface data rights (p. 134), is a sustainment-enterprise design requirement. That is the same terrain as the digital sustainment and predictive-maintenance work around the CH-53K, and the assured industrial access argument David Beaumont has put forward. The natural extension of TOR is an enterprise-level unit of analysis: not only when the airframe, autonomy and powertrain are ready, but when the data rights, depots, workforce and industrial surge capacity that turn a platform into regenerative mass are ready.

The second link is the kill web itself. The thesis decomposes vertically, platform to sub-system. Kill web capability is created horizontally, at the seams, where a drone becomes a node that sensors, shooters, command and control, and logistics can all use. The CCA matters not because the airframe flies but because of what the force can do with it as part of a network. The integration readiness work the literature review introduces is where that conversation belongs, and I would like to see it carried into the model.

Third, Keirin’s insistence that waiting should be productive is close to what I have come to call operational discovery in my work on the Nevada National Guard’s Desert Forge: capability and doctrine are found by using bounded systems under realistic conditions, not delivered by a forecast. Camel Train, ASCA challenges and military drone-racing teams are discovery mechanisms. They build the tech-literate workforce Keirin says must be in place when the regulatory and technical barriers drop. This also answers my adoption-versus-readiness concern, since institutions that have been practicing absorb a technology far faster than ones that wait for the all-clear. It is chaos management rather than crisis management.

Finally, the thesis says the method should extend to uncrewed surface and underwater vessels (p. 7). I would welcome that test. The maritime autonomy programs now moving in Australia, such as Ghost Shark, Bluebottle and Speartooth, operate under a different regulatory gate and a different industrial base, and they bear directly on the hull-gap debate. A TOR run on a maritime archetype would show whether the framework generalizes or is partly an artifact of aviation’s certification culture.

Bottom line

This is a substantial piece of work. Its best idea, that capability matures sub-system by sub-system and that regulation and sustainment belong inside the forecast, deserves to be adopted at Gate Zero whether or not the TOR arithmetic survives scrutiny.

Its weakest point is the confidence of its claims relative to its evidence, which Keirin partly acknowledges and the next phase of his research can fix by publishing dated, falsifiable predictions.

For ADF capability managers, the Acquire, Invest or Wait framing, with Wait as an active posture, is an immediately usable tool.

For the rest of us, the thesis is a reminder that the hard question about drones is no longer whether they will matter but when each piece becomes real, and what the institution does in the meantime.