robotics//fleet//swarm robotics

Swarm robotics is the design of multi-robot systems that deliberately exploit emergence: many simple, nearly identical robots with local sensing, local communication and no coordinator, whose useful collective behaviour arises from rules each robot applies to its neighbours. Its promises are scale (each robot talks to a few neighbours however large the swarm grows), tolerance to losing members and independence from infrastructure; hundreds of cheap robots combing a disaster zone with no network is the picture it is built for.

Swarm robotics is the design of multi-robot systems that deliberately exploit emergence: many simple, nearly identical robots with local sensing, local communication and no coordinator, whose useful collective behaviour arises from rules each robot applies to its neighbours. Its promises are scale (each robot talks to a few neighbours however large the swarm grows), tolerance to losing members and independence from infrastructure; hundreds of cheap robots combing a disaster zone with no network is the picture it is built for.

The building blocks are local rules such as those of flocking and the averaging of the consensus protocol, and the behaviour of the whole is emergence, written in no individual rule. Reference demonstrations exist, such as Harvard's thousand Kilobots assembling into shapes in 2014 and outdoor flocks of tens of drones. They are serious research; in production almost nothing meets the full definition.

Emergence scales, but it is hard to design and hard to verify.

Going from local rules to global behaviour is done by simulating; going back from the desired behaviour to the rules (the inverse problem) has no general method, so parameters are searched by optimization, evolution or reinforcement learning in simulation, with the sim-to-real gap that brings. Proving that an emergent behaviour will never do something costs far more than proving it of an explicit plan.

Failures are collective too. Agreement amplifies whatever value wins, so a consensus on a wrong value is an error with quorum: one bad reading that spreads becomes the swarm's belief. The swarm also gives up the optimality a central plan would reach.

A drone light show with thousands of drones looks like the ultimate swarm and is nearly the opposite. Each drone flies a trajectory computed and checked in advance so that it never crosses another, follows it with high-precision GNSS (usually RTK) and a common clock, and a ground station supervises and can abort. Nothing is negotiated in flight: if one drone fails it lands and the rest carry on. It is an orchestra playing from a printed score, which is why it works and can be authorised.

Before accepting that a system is a swarm, ask what happens when the radio goes down and who signs the safety analysis. Most deployed multi-robot systems, defence included, are fleets with a coordinator (fleet), for the reasons set out in fleet architecture.