A Kilobot swarm. Credit: Michael Rubenstein, Harvard University.

A Swarm of a Thousand Cooperative, Self-Organising Robots

ByEd Yong
August 14, 2014
9 min read

In a lab at Harvard’s Wyss Institute, the world’s largest swarm of cooperative robots is building a star… out of themselves. There are 1024 of these inch-wide ‘Kilobots’, and they can arrange themselves into different shapes, from a letter to a wrench. They are slow and comically jerky in their movements, but they are also autonomous. Once they’re given a shape, they can recreate it without any further instructions, simply by cooperating with their neighbours and organising themselves.

The Kilobots are the work of Mike Rubenstein, Alejandro Cornejo and Radhika Nagpal, who were inspired by natural swarms, where simple and limited units can cooperate to do great things. Thousands of fire ants can unite into living bridges, rafts and buildings. Billions of unthinking neurons can create the human brain. Trillions of cells can fashion a tree or a tyrannosaur. Scientists have tried to make artificial swarms with similar abilities, but building and programming them is expensive and difficult. Most of these robot herds consist of a few dozen units, and only a few include more than a hundred. The Kilobots smash that record.

They’re still a far cry from the combiner robots of my childhood cartoons: they’re arrange themselves into two-dimensional shapes rather than assembling Voltron-style into actual objects. But they’re already an impressive achievement. “This is not only the largest swarm of robots in the world but also an excellent test bed, allowing us to validate collective algorithms in practice,” says Roderich Gross from the University of Sheffield, who has bought 900 of the robots himself to use in his own experiments.

“This is a staggering work,” adds Iain Couzin, who studies collective animal behaviour at Princeton University. “It offers a vision of the future where robot groups could form structures on demand as, for example, in search-and-rescue in dangerous environments, or even the formation of miniature swarms within the body to detect and treat disease.”

"And I'll form... the wrench!" Credit: Michael Rubenstein, Harvard University.
“And I’ll form… the wrench!” Credit: Michael Rubenstein, Harvard University.

To create their legion, the team had to rethink every aspect of a typical robot. “If you have a power switch, it takes four seconds to push that, so it’ll take over an hour to turn on a thousand robots,” says Rubenstein. “Charging them, turning them on, sending them new instructions… everything you do with a thousand robots has to be at the level of all the robots at once.”

They also have to be cheap. Fancy parts might make each bot more powerful, but would turn a swarm into a budget-breaker. Even wheels were out. Instead, the team used simpler vibration motors. If you leave your phone on a table and it vibrates, it will also slide slightly: that’s how the Kilobots move. They have two motors: if either vibrates individually, the robot rotates; if both vibrate, it goes straight.

LIMITED TIME OFFER

Get a FREE tote featuring 1 of 7 ICONIC PLACES OF THE WORLD

Well, straight-ish, anyway. The tyranny of cost-efficiency meant that the team had to lose any sensors that might tell the robots their bearings or positions. They can’t tell where they are, or if they’re going straight. But each one can shoot infrared beams to the surface below it, and sense the beams reflecting from its neighbours. By measuring how bright the reflections are, it can calculate its distance from other Kilobots.

This combination of stilted motion and dulled senses meant that each robot costs just $20. It also meant that “the robots were even more limited than we expected,” says Rubenstein. “The way they sense distance is noisy and imprecise. You can tell them to move and they won’t, and they’ll have no idea that they’re not moving.”

Fortunately, they have each other. A stuck Kilobot can’t tell if it’s stuck on its own, but it can communicate with its neighbours. If it thinks it’s moving but the distances from its neighbours change, it can deduce that something is wrong. And if neighbours estimate the distances between them and use the average, they can smooth out individual errors.

Using these principles, the team created a simple program that allows the robots to independently assemble into different shapes using just three behaviours. First, they move by skirting along the edges of a group. Second, they create gradients as a crude way of noting their position in the swarm. (A nominated source robot gets a gradient value of 0. Any adjacent robot that can see it sets its gradient value to 1. Any robot that sees 1 but not 0 sets its gradient to 2, and so on.) Finally, although they have no GPS, they can triangulate their position by talking to their neighbours. As long as the team nominates some robots as seeds, effectively turning them into the zero-point on a invisible graph, the rest of the swarm can then work out where they are.

Every Kilobot runs on the same program. The team only has to give them a shape and nominate four of them as seeds. Once that’s done, the rest slowly pour into the right pattern, in an endearingly life-like way. It takes them around 12 hours, but they do it all without any human intervention. And although the final shapes are always a little warped, that’s life-like too. Fire ants don’t have a Platonic ideal of what a bridge or raft should look like; they just work with their neighbours to get the job done.

Stills from movies showing the Kilobots assembling into a K and a star. Credit: Michael Rubenstein, Harvard University.
Stills from movies showing the Kilobots assembling into a K and a star. Credit: Michael Rubenstein, Harvard University.

Scientists have long been able to simulate huge swarms of life-like virtual particles in computers, using very simple rules. But the real world is full of pesky physics, inconvenient noise, and temperamental circuitry. Stuff goes wrong. By building an actual swarm, the team can address these problems and make their programs more robust. They’ve already had to deal with a litany of failed motors, stalled robots, collisions, and traffic jams. “The more times you run it, the more likely some random thing will show up that you don’t expect,” says Rubenstein. “That’s the problem with 1,000 robots: even rare things can happen very frequently.”

The next step will be to build robots that actually self-assemble by attaching to each other, says Marco Dorigo from the Free University of Brussels.. “We did so with tens of robots,” he says. “It will not be easy with one thousand.” Rubenstein agrees: “Physical connection is always difficult. If you have a dock, you tend to design the rest of the robot around that dock. It has a huge impact.”

Eventually, he also wants to get to a position where the robots can sense their environment and react accordingly, rather than just slide into some pre-determined shape. Like fire ants, when they get to a body of water, they wouldn’t have to be fed the image of a bridge; they would just self-assemble into one. “That’s a whole other level of intelligence, and it’s not really understood how to do that in robotics,” says Rubenstein. “But nature does it well.”

Reference: Rubenstein, Cornejo & Nagpal. 2014. Programmable self-assembly in a thousand-robot swarm. http://dx.doi.org/10.1126/science.1254295

Related Topics

Go Further