early thesis pitch: drone swarming [obsolete]

An economy can be viewed as a framework for coordinating a massive multi-agent system by allowing it to express incentives that conduce to useful work over an extended timescale. Humanity, writ large, represents an adaptable, self-scaling swarm intelligence, despite being composed of fully independent autonomous agents, who share approximate incentives and are coordinated at the macro scale by economic forces. Consider what economic development looks like, at a long time scale and birds-eye view. Humans grow in fractals -- they follow incentives, gather around resource centers, and reproduce themselves to exploit surplus. This effect can be observed on timescales of hours and days with groups of dozens up to decades and centuries with populations of tens of millions: there is a direct analogy to bacterial colonies reproducing in Petri dishes and ants building nests.

Most large physical systems scale this way -- they are composed of autonomous individual units following incentive gradients. The notion of incentive and the space in which it is embedded varies greatly between physical species, from spatial nutrient density for single-celled organisms to sophisticated, high-dimensional notions of value incorporating explicit, codified social structures, language, tradition, experience, and so on, as humans do. What remains interesting about this world model, despite the differences in objective function between bacterium and human, is that humans still scale in the real world more or less like bacterial colonies once you zoom out far enough. Bacteria are contained to the Petri dish -- they are not structurally sophisticated enough to have an understanding of the world around them or meaningful ability to leap energy thresholds to exploit remote resources. Bacteria are effectively spatially confined to resource areas, and have little to no ability to expand beyond that. Bacterial colonies cannot, except accidentally, spread to an adjacent Petri dish, no matter how high the nutrient density (and hence notional incentive) because they cannot collectively sense beyond their colonies and have no structural capability to actuate in the world.

Humans, by contrast, are able to do exactly that. We build logistics networks to transport utility away from production sites, expanding our reach. Human history is a story of repeated expansion beyond a succession of these archetypal Petri dishes, as advances in science and technology enable us to reach beyond our current capabilities and express structure in the universe at grander and grander scales. The most substantial and rapid advances have been due to the commodification of microelectronics and capacity to embed and express computation ubiquitously. But this capability has not yet converged with the human species' ability to scale actuation. We use electronic computation to manage logistics and enable more rapid and wide-area communication. But our models for integrating automation into actuation on the world -- actually scaling our resource cultivation and exploitation capabilities -- have not yet come to fruition. We are starting to see the beginning of this process, but we don't see large numbers of robots participating in economic activity in a scalable way.

On the face of it, this statement appears wrong -- robotic automation is everywhere in mass manufacturing, and automated systems manage inventories, investment, logistics, and more. But they are not integrated in the recursive, fractal, log-boundary-crossing way that living systems are. Automated systems are leaves in the graph conceived and purpose-built for a single application. They are not robust in the sense that they admit of a composition of modular, independent agents that are straightforwardly replaced, which communicate and negotiate in order to produce mutually-beneficial results that conduce towards economic incentives. Assembly lines are hard-coded, prebaked, and require human intervention for all maintenance.

Research in drone and swarming applications mostly does not look at this problem, and is instead concerned with optimality and efficiency of acutation control. While the results of this research are interesting, they make the essential mistake of disregarding investigations of communications and network topology in favor of easier-to-implement centralized control. This makes most of these results functionally unusable for implementing a living system of the kind I describe, as the agents are not independent or self-motivated but limbs of a central brain.

My resources to address this problem are relatively constrained. I work in a hardware group without a strong history of drone research. We don't have terribly deep prior connections nor an existing drone lab to <...>. I don't have much funding to work with. Hence I will be strategic in my approach to proving out this perspective.

The most useful thing to build is a prototype for the concept -- something that demonstrates a simple but unsophisticated living system that can work as a demo and testbed. This is what I intend to do with my master's thesis. It will be a mixed sim-real digital twin environment supporting a swarm of drones that initially can be controlled by a human directly.

I plan to produce a simulation amenable to wide-area, large-unit-count drone control by a human operator, using interfacing paradigms inspired by real-time-strategy (RTS) video games (notably StarCraft II, Beyond All Reason), which have proven effective for this actuation mode. I intend to demonstrate mixed sim-real operation in a digital twin model of a test space, enabling a small number of physical drones (five to ten) to act as a representative subset of a complete virtual swarm of up to hundreds of members, validating behaviors in the real world, helping the simulation converge towards real behavior, and acting as demonstration aids. Taking additional inspiration from the RTS game genre, individual drones will be of limited local intelligence: capable of locomotion, basic obstacle and neighbor avoidance, and a small command library, deferring to the human controller for all higher-level planning and decision-making.

I hope to demonstrate that under an appropriate interfacing paradigm, this limited drone intelligence can be leveraged by a single human operator to rapidly complete wide-area tasks such as object cleanup, delivery logistics, surveillance, interdiction, and sensor deployment / recovery. The goal is to show that this approach provides a substantial productivity multiplier compared with manual human operation, without requiring complicated, training-intensive, motivationally-opaque behavior implemented via ML or RL. I intend to demonstrate that this approach is also amenable to remote control, even with significant actuation and sensing latency.

This approach is exciting because RTS games supply a proven, robust interfacing paradigm that supports sustainable human control of hundreds of independent agents at an actuation rate of hundreds of discrete inputs per minute – actions which are leveraged in a context of evolving, goal-oriented human tactical and strategic decisionmaking, which directly mirrors many forms of real-world productive work. The RTS genre provides a clear model for separation of agent and human responsibilities, making the agent relatively dumb (and therefore easy to implement), but proves that this extremely minimal form of autonomy is sufficient to enable a human operator to produce wide-area effects rapidly and effectively.

To tie this back to the conceptual motivation, the point of this work is to demonstrate that highly-scalable swarms are possible now, implementable by leveraging human intelligence as the strategic and tactical control and keeping individual drone behavior simple and understandable. The abstraction boundary provided by the agent-human interface provides a natural modularization surface that simplifies implementation, enabling me to demonstrate results quickly, but is readily suited to incremental improvements in capability as the work progresses.