07 Introduction to Swarm Robotics · #28 of 46

What Does a Robot Know?

Why a robot that sees almost nothing can run a swarm of thousands

A quadruped robot with a sensor-laden head pausing on rough ground, its perception reaching only a few metres out. — A real robot's whole universe: a few metres of sensed ground around its own body, fading into ignorance at the rim. No robot ever gets the wide shot, and a swarm is built from thousands of these short-sighted machines. · Airman Rhea Beil, Public domain

A swarm of a thousand drones can sweep a stadium in formation. Not one of them has the faintest idea where the other 999 are.

Each sees a handful of neighbours, a few metres of air, and nothing else. The miracle isn't that they coordinate despite this. It's that they coordinate because of it.

You probably imagine a swarm robot as a tiny god with a god's-eye view: a glowing tactical map of every agent, updated live. Delete that picture. The real thing is closer to you at a crowded party with the lights off: you hear the two people beside you, you feel the elbow that just hit your ribs, and the room beyond that is a rumour. Partial observability is not a bug to be engineered away. It is the whole design.

The blindfolded party

Put yourself back at that dark party. Your senses are brutally short-range:

You can hear voices within a few metres.
You feel anyone who bumps into you.
You know roughly where you are in the room.

What you cannot do is see the whole party, locate everyone, or spot the exit. What you can do is drift toward the nearby voices, dodge the collisions, and stumble your way out. And here is the thing: that is enough. Crowds of humans evacuate buildings, form queues, and flow through doorways with exactly this much information. Nobody has the map. The map is an illusion you assemble locally, one neighbour at a time.

A robot's "knowledge," then, is almost embarrassingly small: it is its sensor reach, plus whatever its neighbours bother to relay. That's it. There is no hidden global state it can consult. If it wants to know something far away, the news has to physically travel to it, hop by hop, through the swarm.

Why the textbooks bake 'no global view' straight into the definition

This is not a vibe. It is written into how the field defines its own systems. The standard textbook list of what makes a "multi-agent system" includes three blunt requirements, and two of them are about ignorance on purpose:

Autonomy. Each agent runs itself.
Local views. No agent has a full global view, or the system is simply too complex for any one agent to use such knowledge even if it had it.
Decentralization. No single agent is in charge. The moment one is, you no longer have a swarm; you have a regular machine with extra steps.

Swarm robotics goes further and turns this into a design rule you are not allowed to break. One widely cited set of swarm principles states it flatly: robots possess only local perceiving and communicating ability, and robots do not exploit centralized swarm control or global knowledge. So the short-sightedness in this lesson is not a limitation the engineers are stuck with. It is a constraint they deliberately impose, because a system that secretly relies on a god's-eye view stops scaling the instant the crowd gets big, and it dies the instant the all-knowing controller does.

The sensing radius, and the trap inside it

The textbook model gives each robot a sensing radius $r$ : a circle of attention. Anything inside the circle is a neighbour it can perceive and talk to; anything outside is grey, unknown, effectively nonexistent. Picture one robot with that circle drawn around it, and you can already feel the trade-off in your gut.

Shrink the radius and the swarm fragments. Each robot sees fewer neighbours, links snap, and the group shatters into islands that can no longer pass information to one another. Grow the radius and coordination tightens, but every robot now pays more, in sensing, in bandwidth, in battery, to track a crowd it mostly doesn't need.

A PR2 mobile manipulator robot at Maker Faire. — Willow Garage's PR2: a mobile manipulator whose every move depends on what it knows, its map, its joint angles, and the objects around it. · Timothy Vollmer, CC BY 2.0

So a fixed radius forces a miserable choice: connected-but-expensive, or cheap-but-fragmented. The slick fix is to stop thinking in radii at all.

The trick that scales to thousands

Here is the move that makes massive swarms possible, and it is delightfully counterintuitive. Don't track everyone within a fixed distance. Track a fixed number of your nearest neighbours, your closest six or seven, whoever they happen to be.

This single change rewrites the economics. In a sparse patch, your "six nearest" might be metres away; in a dense crush, centimetres. Either way, the work per robot stays flat. Drop a thousand new agents into the swarm and each robot still reasons about roughly the same handful of neighbours it always did. The cost of being one robot does not grow with the size of the crowd.

One thousand and twenty-four

"Thousands" stopped being a thought experiment in 2014. A team at Harvard switched on a swarm of 1,024 tiny robots called Kilobots and watched them shuffle, sense their handful of neighbours, and slowly assemble themselves into shapes like a starfish or the letter K, with no robot in charge. It was the largest robot swarm ever run at the time, and it worked for exactly the reason this lesson keeps hammering: each Kilobot only ever paid attention to the few machines right next to it. Compare that to the field's first big project, the European SWARM-BOTS effort of 2001-2005, which got its robots to physically clip together and haul objects as a team but topped out at around 20 of them. In thirteen years the headcount jumped fiftyfold, because the per-robot recipe never had to change.

The math: neighbour sets, noise, and why connectivity is everything

The radius model defines robot $i$ 's neighbour set as everyone within distance $r$ :

$N_i = \{\, j : \lVert p_i - p_j \rVert \le r \,\}$

where $p_i, p_j$ are the positions of robots $i$ and $j$ . The $k$ -nearest variant instead keeps $N_i$ as the $k$ closest agents to $i$ , fixing the cardinality of the set rather than the region it covers, which is precisely what bounds the per-robot cost as the swarm grows.

Real sensors don't return the true distance. They return a noisy estimate:

$\tilde{d}_{ij} = \lVert p_i - p_j \rVert + \epsilon, \qquad \epsilon \sim \mathcal{N}(0, \sigma^2)$

That Gaussian $\epsilon$ is the flicker from the gotcha above: a neighbour near the boundary jumps in and out as the noise pushes $\tilde{d}_{ij}$ across the threshold.

Stack every neighbour set together and you get the swarm's interaction graph. Coordination (agreement, consensus, shared heading) is only possible if that graph stays connected:

$\text{Connected} \iff \forall\, i, j \;\; \exists \text{ a path } i \to j$

Let the graph split into disconnected components and consensus becomes mathematically impossible: two islands that share no path can never reconcile their views. Connectivity is not a nicety. It is the precondition for the swarm being a swarm at all.

Making the search cheap. A brute-force neighbour search compares every pair, which is $O(n^2)$ and dies on large swarms. The standard cure is a uniform-grid spatial hash: divide the world into cells the size of the sensing radius, and let each agent only check its own cell plus the eight adjacent ones. Because near-neighbours can only live in those nine cells, the search collapses to roughly $O(n)$ .

Key takeaways

Partial observability is the default. A robot's knowledge is its sensor reach plus whatever neighbours relay; there is no global map to consult.
A fixed radius forces a bad trade: too small and the swarm fragments into islands; too large and every robot pays to track a crowd it doesn't need.
Tracking a fixed number of nearest neighbours, not a fixed radius, is what lets swarms scale to thousands. The work per robot stays flat as the crowd grows.
Noise lives at the edge. Far neighbours flicker in and out; robust swarms smooth and gate their sensing instead of reacting to every jitter.

We tend to assume intelligence needs a wide view: the more you see, the smarter you are. The swarm quietly disagrees. A starling, a drone, a fish in a school: each one knows almost nothing, sees almost nothing, and trusts a handful of neighbours to carry the rest.

Out of all that local ignorance, a single coherent thing turns in the sky. The whole was never in any one of them. It was always in the spaces between.