BUT IT'S OK, THEY'RE HERE TO HELP YOU

In his second dispatch from the Idea Festival in Louisville, Evgeny Morozov watches a podium-full of robots buzz around like bees, ask each other questions, find an orange, leave the room, form an orchestra, and prepare one day to save your life ...

Special to MORE INTELLIGENT LIFE

Surrounded by buzzing robots that end the session by performing in an orchestra, James McLurkin, a PhD student at MIT Computer Science and Artificial Intelligence Laboratory, talks about distributed robotics and swarm behavior to a packed house. His work has its roots in "swarm intelligence"—the study of collective behavior in decentralised, self-organised systems. Think of ant colonies, bird flocking, animal herding, fish schooling, and many other examples in nature.

During the last few hundred million years, nature has perfected such interactions. Now, scientists such as McLurkin want to get a better understanding of how these biological processes work and apply this knowledge to programming robots for doing complex tasks in groups. Perhaps, this is the ultimate interpretation of the Wisdom of Crowds thesis: individuals don’t have to be smart to produce very smart group outcomes. Did somebody mention Wikipedia?

Early on, McLurkin pulls up a slide of Isaac Asimov's famed three laws of robotics, intended to forestall a robot revolt against humanity. “Well, robots don’t know how to read, so those laws are not particularly useful”, he smiles. Robots are not even smart enough to travel from the stage to the audience: they would get trapped in wires or collapse to the floor. For all the talk about robotics, today an average squirrel can still do more than any robot, he says.

He points to a number of philosophical, not just engineering problems, in his field. Problem number one is that we don’t know what intelligence is, nor how to define it. Should we subject the robots to some upgraded version of the Turing test (which says that if a judge can’t tell whether he is talking to a machine or a person, the machine passes the intelligence test)?

Can intelligence emerge from interactions of unintelligent components? That is a second philosophical question. As we are all built from molecules, continues McLurkin, either intelligence is something that results from interactions, or molecules are intelligent.

The third and final question is whether an intellect needs a body. Can a brain in a vat understand and experience the world without anything to relate to? Can we build such an intellect?

That slide with the three philosophical questions is subtitled ”things that make you go “hmmm”, and one can hear half of the audience “hmming”.

Having finished with the philosophy, McLurkin gives a brief overview of earlier efforts to mass-build robots, presenting quite a few models, from iRobot Roomba to Honda Asimo to iRobot Packbot, all of them having different looks and different functionality. And, of course, NASA’s successful launch of two robots on to Mars.

Quite naturally, he makes a transition to his own work. He has 112 robots in his arsenal and he is trying hard to make them work together. In his view, robots are best at jobs that are dangerous, dirty, or dull: “What if we sent 20 robots to work in hot spots around the world? What if we sent 200 robots to look for surivors after an earthquake? What if sent 2,000 robots to explore Mars?”.

It’s this last question he wants to address with his on-stage demonstration. McLurkin turns to a few dozen robots that he has on stage (he controls them with a remote). As a starter, he asks the robots to form a line; surely enough, they do. Next, he orders the robots to spread out. They do this too. The demonstration proceeds quite smoothly.

One thing that the robots don’t know yet is how to define boundaries of the network, so they often spread out from the center and then get disconnected. The robots can communicate via one another (they know the neighbors, but don’t know about everybody else) but not with everybody at once. So if they need to find a robot that is not in their neighborhood, they must relay the info via their neighbors.

To find the answer, they go around and query one another to find the result. The robot that is searching just goes around and asks a robot next to him. The network reconfigures in real-time and the robot is going to move around the network until it finds the robot in question. They can also form protective areas/fences. And, of course, they can also leave the planet in orderly fashion, so McLurkin has his robots leave the stage by ID. Two special robots know they are special and the rest know that they are ordinary. So they query all neighbors about their ID and then place themselves between the two neighbors—one that has a greater id than them and one that has a lower id than them—until the whole “squad” is arranged.

After the demonstration is over, McLurkin wonders aloud how exactly to program 2,000 robots. Natural systems—and in particular in swarm intelligence—can provide insight into these extremely complex programming problems.

Thus, one of the areas that he examines very closely is nectar collection in bees—how foraging bees communicate with workers beers in the hive (Thomas Seeley, a biologist at Cornell University, has been doing some very interesting work on this topic). With thousands of workers in a hive, honeybees have learned to bypass their individual judgements and do what’s best for the colony.

Getting enough information about the processes that drive this decision-making in bees would allow one to present the bee algorithm in a software-like way. McLurkin’s question is whether he can run this software on his robots. This is what he dubs “beeware”—the one that doesn’t need to get debugged, as the nature has already done its job of “debugging” over the last 120 million years.

Communication among ants is not as complicated as it seems. They leave pheromone trails whenever they find food. Other ants follow the trails when they find them, instead of searching for food randomly. The ants are foraging in a globally optimal fashion, by exploring the closest food source—"ant colony optimisation".

The magic of complexity, according to McLurkin, is how simple local interactions can form complex group behaviors. This is pretty much how insect communities work—"distributed systems". This is a very interesting research area, but it still makes a robot swarm difficult to program, he says. So he wants to do something different. Instead of writing software for individual robots, he wants to write it for groups of them. (This wasn't the easiest part of his talk to understand.)

The key to understanding the thought process behind McLurkin’s robotics is what is called “distributed averaging”. Think about software than runs on multiple computers and interacts to form a group result. To illustrate his point, McLurkin asks for eight volunteers from the audience, gives each of them a piece of paper with a number, splits them in pairs, handed them calculators and asks them to calculate the average and then change the partners and repeat the calculation. After a few rounds, almost everybody arrives at the same number—although they never talk to the whole group, just to their partners in each round.

Bees constantly engage in similar processes. Honeybee workers share food all the time, precisely thanks to their computation of a global average. This lets an individual worker know when the hive is hungry by measuring when she is herself is hungry. The assumption is simple: “When I am hungry, the rest of the hive is hungry”.

McLurkin shows a short video of a few dozen bots exploring a room and searching for an orange ball hidden there. There are four types of robots, each playing a special role in the experiment, operated and commanded via SwarmOS, a special operating system written by McLurkin. They robots succeed in their mission and leave the room. Only one robot gets stuck. Imagine now that orange is a land-mine or a bomb, and you understand why government agencies as well as idea festivals, are so excited about this kind of work.