James McLurkin‘s talk is titled “Dances with Robots”. He’s a researcher at MIT studying distributed algorithms for multi-robot systems, a former engineer with iRobot. He offers the observation that Hollywood portrayals of robots fall into three basic categories – frankenstein (robot alienated from society), the tin man (robot wants to be human) and terminator (giant killer robots. It’s the last scenario that freaks people out. Isaac Asimov suggested a series of laws that should prevent robots from killing humanity, as in Will Smith vehicle “I Robot” – a robot must not injure a human, must obey orders, and may protect it’s own existence so long as that doesn’t conflict with the first two laws. (Yes, I know the laws are significantly more careful than that, but there’s no connectivity in the room right now…)
But it’s hard in real life to get these laws to apply because no robot in real life can read these laws, none can reliably tell you what’s a human being, and, frankly, can’t prevent themselves from driving off stage and falling to their death. “Your average squirrel can run through trees at eight miles an hour. A honeybee can fly at 20 miles an hour, avoiding obstacles. That’s something no robot can do.”
McLurkin offers three deep philosophical problems associated with robotics:
– What’s intelligence? (He argues that the Turing test doesn’t test intelligence, but demonstrates a failure in intelligence, a test for non-intelligence.)
– Can intelligence emerge from the interaction of unintelligent components? (Sure – atoms aren’t very smart, but pretty smart engineers are built out of atoms.)
– Does intellect need a body, and does a body effect the type in intelligence we have? (It seems likely, but we only know of one type of intelligence and one body. Perhaps we’ve just met only the dumb dolphins, he posits.)
We should contrast the scary Hollywood robots with realworld helpful robots. iRobot’s Roomba is useful to us not because it’s smart, but because it’s “very cleverly stupid”. Rather than calculating an ideal path to vacuum a floor, it simply vacuums enough to probabilisticly cover the entire floor. He shows the Honda ASIMO, NASA’s Spirit and Opportunity Mars Rovers, and iRobot’s Packbot, a $65,000 robot designed to probe dangerous situations to keep humans away from booby traps.
But McLurkin’s research isn’t on single robots, but on swarms. He mentions that “robots are best at tasks that are dangerous, dirty and dull”, and that these tasks are often solved best in swarms. For instance, you’d love to send 20 robots into a forest fire and detect hotspots that could explode into more serious fires. You’d love to have a swarm of robots that could look for survivors of an earthquake. In fact, you might do even better with 20,000 cockroach-sized robots who could search for humans, then rat-sized robots who could analyze the structures the people are stuck in, then huge robots that could move away debris.
But programming lots of robots – 112, in his current research set – or thousands, as he’d like to work with – requires some very different techniques than writing conventional software. He shows off some of the techniques by demonstrating a small fleet of robots. They organize themselves into a line, swarm to a location, orbit a goal, and sort themselves by their unique identity. The communication between the robots is via IR – each robot has four IR transmitters and sensors. Communication throughout the network happens via mesh networking – robots propogate a signal from one to the next, looking for a robot that’s closest to the goal.
(One of the loveliest things about his robots is that they communicate surprisingly well with humans – they blink blue and red to reveal their identity and status, and each features a 1.1 watt audio system, which plays appropriate videogame music to accompany behaviors – the Pac-Man ghost music when the robots swarm, for instance.)
As cool as this is, McLurkin points out that ants do this all the time, and frequently do it better than his robots do. Ants appear to solve extremely complex problems through what seems to be extremely intelligent behavior, dedicating resources to searching for close food sources instead of far ones. But ants are solving this problem through a very simple algorithm – they’re simply following the stinkiest scent trails. Because ants visiting a nearby food source oscilate between the food and the nest more quickly, they end up creating a more stable trail, which other ants end up following.
Programming robots to exhibit this behavior requires you to determine the group behavior you’d like, then figure out how what simple robot behavior would be required to emerge into this group behavior, then figure out how robots real-world interactions require modifications to the programming. It’s really, really hard, but quite impressive when it works. He shows a demo of a swarm of 112 robots searching a room, where robots search for goals, send themselves back to base to recharge when neccesary, and send guides to bring a human to a goal. From his swarm, only one failed to return home, well within his acceptable tolerances for failure.
To give us an example for how these algorithms get written, he invites eight people to the stage, assigns them each a unique number and a calculator. He then asks everyone to find a partner and average their numbers. After three iterations, almost everyone in the group has converged on the average of the entire set of numbers. (McLurkin tells us that, in simulation, it should take about 12 cycles to converge on the average.) He shows us that each pairwise averaging should decrease variation, moving people closer to the mean. He offers the observation that this averaging behavior is roughly how honeybees find food. (He also offers the intriguing insight that honeybees have the highest neural density of any creature – they’re not very smart, but in terms of their size, they’re Einsteins.
McLurkin’s path to robotics is a clear geek path – he walks us through his personal history, from model trains to legos, to BMX biking, to RC cars, to homemade robots. He’s got a real flair for making these machines as magical to an audience as they obviously are to him.