UAS in Swarm Robotics and Multi-Vehicle Systems

ost things called a drone swarm are not swarms at all. The real dividing line is not how many aircraft are in the air, it is who, or what, is making the decisions.

In October 2016, the United States Department of Defense launched 103 Perdix micro-drones from three F/A-18 Super Hornet fighters over China Lake, California. Perdix is a palm-sized, expendable micro-drone, originally developed by MIT students and built out by the Pentagon's Strategic Capabilities Office, designed to be carried in a fighter jet's flare dispensers and launched in bulk mid-flight. None of them had a central controller telling them what to do. Each one communicated with every other one, the group had no leader, and when drones were added or lost the rest adapted without anyone reprogramming anything. The program's director described it afterward as a single collective organism sharing one distributed brain.

That same year, and in most years since, a very different kind of event has been marketed under the same word. A few hundred drones rise over a stadium, trace a logo in the sky, and the broadcast calls it a drone swarm. It is an impressive show. It is not what the Perdix demonstration was. One of these systems is a single computer and operator choreographing individually pre-planned flight paths. The other is a distributed network of aircraft making their own decisions in real time. Both get called swarms, and the gap between them is the actual subject of this post.

A useful definition of a swarm has nothing to do with how many aircraft are involved. It has to do with where the decision-making lives. A formation of drones following a pre-planned route, however large, however visually complex, is still a single plan executed by many airframes, with one computer or one operator holding the complete picture. A true swarm pushes that decision-making out to the edge: each aircraft senses its immediate surroundings, communicates with nearby neighbors, and decides its own next move, with no single point that holds, or needs, the complete picture at all.

This is the same dividing line this series drew in post 11 between VLOS and BVLOS operations, just scaled up from one aircraft to many. There, the question was whether the control loop lived in a human's head or in the aircraft's own systems. Here, the question is whether the control loop lives in one central planner or in the distributed interactions of every aircraft in the group. A light show answers that question one way. A true swarm answers it the other way, and that answer changes what the system can survive, what it can scale to, and how hard it is to certify.

A hundred aircraft flying a single pre-planned route is one decision executed many times over. A swarm is many decisions, made locally, that add up to one coherent behavior with nobody holding the complete picture.

The clearest model for how a real swarm produces coordinated behavior without central control comes from outside aerospace entirely. In 1986, computer graphics researcher Craig Reynolds built a flocking simulation called Boids, and showed that convincing, lifelike flocking emerges from just three local rules, each one followed by every individual with no awareness of the group as a whole. Separation: steer away from neighbors that are too close. Alignment: steer toward the average heading of nearby neighbors. Cohesion: steer toward the average position of nearby neighbors. No boid knows the flock's overall shape or destination. The flock-level behavior, the wheeling, tightening, splitting pattern you see in a starling murmuration, a dense flock of thousands of starlings wheeling through the sky in one continuously shifting shape, is an emergent result of simple local rules, not a plan anyone is executing.

A real UAS swarm runs the same logic with sensors and radios standing in for a bird's eyes. Each aircraft maintains awareness of its nearest neighbors through peer-to-peer communication, applies local rules for spacing and heading, and the coordinated group behavior, maintaining formation, closing a gap left by a lost unit, converging on a target, emerges from those local interactions rather than being issued from above. Nobody is drawing the flock's shape on a map. The shape is what falls out when every unit follows its own local rules well.

The practical difference between a formation and a swarm starts with the communication topology. A drone light show runs on something close to a hub and spoke network: a single ground computer plans every aircraft's individual, pre-rendered flight path in advance, and each drone mostly just executes its own script, holding position relative to GPS time rather than reacting to its neighbors in real time. Intel's Shooting Star, a purpose-built, LED-equipped quadcopter Intel designed specifically for synchronized light shows rather than any other mission, used in Super Bowl halftime shows and similar large-scale displays, is explicit about this: a single computer and operator controls large numbers of drones through pre-planned choreography. That is a legitimate and genuinely difficult engineering achievement, precise GPS-timed formation flying at scale, but it is formation flying, not swarm intelligence, and the aircraft are not making decisions about each other.

A true swarm runs on something closer to a mesh: each aircraft talks to its nearby neighbors directly, with no hub that has to stay up for the system to keep functioning. This is what made the Perdix demonstration notable rather than just large. Because every Perdix communicated with every other Perdix, the group had no leader, and the swarm could adapt gracefully when drones entered or exited the formation, properties a hub and spoke system does not have, since losing the hub there means losing the whole show. The U.S. military has kept building on that foundation since. DARPA's OFFSET program ran live field experiments through the early 2020s putting hundreds of aerial and ground robots through coordinated urban tactics, operated by a single human supervisor issuing high-level intent rather than individual flight commands, with the swarm itself handling the moment-to-moment coordination.

That distributed decision-making is what buys a swarm its most valuable property: graceful degradation. A formation show that loses its central computer loses the entire show. A swarm that loses a third of its members, as the Perdix demonstration was specifically designed to show, keeps functioning, because no single node was ever load-bearing for the whole system. This is the same redundancy argument this series made about individual aircraft back in post 6, just moved up a level: instead of redundant sensors inside one airframe, it is redundant decision-making authority spread across an entire fleet.

Intel's own description of its Shooting Star platform is direct about this: a single computer and operator controls large numbers of drones through pre-planned choreography. That is formation flying executed beautifully. It is not the aircraft deciding anything.

The uncomfortable part is that true swarm behavior is genuinely harder to certify and harder to trust than formation flying, for exactly the reason that makes it valuable. A pre-planned formation's entire behavior can be simulated, reviewed, and approved in advance, because the path is fixed before the aircraft ever take off. A true swarm's behavior is emergent: nobody, including the engineers who built it, can state in advance the exact trajectory any single aircraft will fly, only the local rules that govern it and the range of group behaviors those rules tend to produce. That is a difficult thing to hand a regulator. The design freeze and traceability discipline this series covered in the previous post assumes a fixed, known design. A swarm's value comes specifically from not having a single fixed answer for what any one unit will do next.

This is also why most commercial and civilian applications of multi-drone systems today are closer to formation flying, or to independently operating fleets that merely share airspace, than to true decentralized swarms. A warehouse running several inventory-scanning drones at once, or an agricultural operation flying multiple spraying aircraft over the same field, benefits enormously from drones operating simultaneously, but most of these systems still rely on centralized mission planning and collision avoidance rather than the kind of leaderless, self-healing coordination Perdix demonstrated. The hard problem that remains open is not flying many aircraft at once. It is giving up the single point of planning and still trusting the result enough to certify it, insure it, and fly it around people.

Ant colonies solved this same coordination problem millions of years before anyone built a drone. No ant holds a map of the colony's foraging routes or issues instructions to the rest. Each ant follows a simple local rule: lay down a pheromone trail when carrying food, and preferentially follow the strongest nearby trail. Colony-level behavior that looks deliberate, efficient foraging routes, rapid convergence on a new food source, is an emergent result of thousands of individuals following that one simple rule, a phenomenon biologists call stigmergy, coordination through traces left in a shared environment rather than through direct communication or central planning. Nobody designed the colony's foraging pattern. It fell out of the rule, the same way a starling murmuration's shape falls out of three simple flocking rules, and the same way a true drone swarm's group behavior falls out of whatever local rules its engineers gave each aircraft.

The word swarm gets applied to two genuinely different systems, and the difference is not scale, it is where the decision-making lives.

• Formation flying runs on a hub and spoke network with one central plan; a true swarm runs on a mesh with decisions made locally by each aircraft
• The Boids model shows that complex, coordinated group behavior can emerge from three simple local rules, with no individual aware of the group's overall shape
• A true swarm's key advantage is graceful degradation: losing a unit does not break the system, because no single node was ever load-bearing
• That same emergent, unpredictable-in-detail behavior is exactly what makes a true swarm harder to certify than a pre-planned formation
• Most commercial multi-drone systems flying today are closer to coordinated formations than to true decentralized swarms

The next post in this chapter turns from what is technically possible to what regulation, cost, and material limitations actually allow, the practical bottlenecks that decide which of these systems make it out of a demonstration and into daily use.

Curious to exchange some ideas? Reach out via the contact form or connect on Linkedin!