Slashdot Mirror
Cockroach-Controlled Robot
robotsrule writes "The latest issue of Make Magazine volume 2 from O'Reilly publishing has an article on a cockroach controlled robot. Roboticist Garnet Hertz has mounted a Giant Madagascan Hissing Cockroach that drives a small mobile robot around by walking on top of a Kensington trackball. There is a row of proximity sensor triggered LEDs that shine light in the roach's eyes, making him steer the robot since roaches instinctively avoid light. Garnet's web page 'Control and Communication in the Animal and the Machine' details the project with several images of the roach in action. Debugging the project is inherently impossible."
So let me get this straight; humans are controlling a cockroach that is controlling a robot? If they're going to tell the cockroach to turn left, why not just do it directly? It seems the cooler thing to do would be to let the cockroach bump around on his own will.
Hmm indeed.
These folks have constructed a fly flight simulator, where they tether a fly in front of a set of lights that form a low-resolution view of a simulated environment. They then measure the torques produced by the fly and use that data to manipulate the environment simulation, so that the fly sees its environment moving about it even though it's tethered in place. They can use this to study the behavioral responses of flies to various stimuli (like a rapidly-approaching light or dark spot simulating a fly-swatter, to examine escape response).
I guess I would be more impressed if instead of having the cockroach walk on a trackball, they used cockroach EMG to control the robot.
It was a Guinea Pig, and it controled the movements of the machines limbs directly with its limbs. It was intened to reverse the psycological power relationship between people and small rodents, and the rodent seemed to enjoy it. Never shown publicly because animal rights sorts got in a tizzy about it.
Plato seems wrong to me today
What's interesting about this is the way it combines biological and machine intelligence. Cockroaches (at least the kind we get here; I've never seen a Giant Madagascar Hissing Cockroach -for which, I heartily thank whatever power there may be) have a fairly complex pattern of behavior; they "understand" finding food, running away from predators, locating good places to hide and live, etc, etc. These behaviors are not easy to model using machine intelligence; at least, not in a way that works in the real world.
Imagine if you could program a cockroach. You have chemical and optical receptors, and the ability to move individual legs and appendages. You have to be able to do things like "go forward" and "eat this object," both of which are quite complex in practice. You also have to answer questions like "is it food" or "is it scary." And you need an overriding program that prioritizes the various items in your environment and decides, at each moment, what you should do next. How many lines of C code do you think this would take? How do you suppose any given program might fare in the real world, competing for resources (and trying to avoid being eaten or stepped on) alongside real cockroaches?
So what if it turns out that real cockroaches are way ahead of the state of the art in machine intelligence - but you find cockroach behavior useful in some potential system? Instead of waiting for AI research to catch up to the cockroach, why not just put a cockroach in the driver's seat and let it run the robot. Treat it as a "black box" heuristic machine. Use machine intelligence only to constrain its behavior.
As to shining a light on the cockroach to get it to run a particular direction, you have to ask: Why not just turn refuse to obey the cockroach's commands if you are going "the wrong way" according to the machine intelligence? I don't know the answer, but I didn't have to face the task of actually building the thing. However, I would imagine it is much simpler to turn LEDs on and off than to get involved in the mechanics of how the legs work.
If you can figure out how cockroaches (or some other control animal) identifies friend and foe, how they focus on a particular target, and what their attack behaviors are, it's entirely plausible that you could use this in a military application. Imagine armored, landscape-destroying robots with giant laser weapons; piloted, perhaps, by kittens. I'm sure this is where the research will inevitably lead.
-Graham