loren@tristan.llnl.gov (Loren Petrich) (03/12/91)
In the current issue of "Discover" magazine, there is an article about attempting to construct robots with insect-level intelligence. It featured the work of Rodney Brooks, who has had great success with such robots. He and his colleagues have built several kinds of such robots, including some six-legged walkers and some wheeled ones. They are programmed with simple heuristics which compete with each other to make responses. For instance, short-range responses like "back off", when triggered by the prospect of a collision, can override long-range goals like "track prey". Is this type of setup an example of "Fuzzy Logic"? The article reports that Pattie Maes has derived a program that can help one of Brooks's six-legged robots learn to walk. At first, the robot flails its legs around uselessly. But it then learns what leg motions both keep the robot up and make it go forward. Before too long, the robot has learned an insect gait -- to move the two outer legs on each side in sync with the middle leg of the other. The article did not describe the algorithm, but it was probably some kind of Neural Net algorithm. Brooks got the idea from considering that insects are awfully dumb animals, but ones which are nevertheless able to walk and fly and see. He had the idea of building up from simple reflexes. He decided to leave behind the traditional AI model of symbolic reasoning about the world, because it had yielded only very limited success in controlling robots. For instance, such models have trouble with unfamiliar circumstances, circumstances that Brooks's robots have no trouble with. And they require a LOT more processing power than Brooks's robots, which only need simple microprocessors. Not surprisingly, he has looked down on symbolic-modeling AI, which has been the mainstream of the field. And he has gotten a lot of flak from those who have worked in the symbolic-AI field, who have accused him of scaling down his goals from human-level intelligence to insect-level intelligence. My response to that would be that insect-level intelligence is better than level-zero intelligence. There are a lot of analogies with the Neural Nets field -- a system not based on traditional, symbol-manipulating, AI that is much simpler than it and easily outperforms it, but which seems to have much less in the way of ultimate potential. Is that fair? And does anyone else know of any details of Brooks's work? $$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$ Loren Petrich, the Master Blaster: loren@sunlight.llnl.gov Since this nodename is not widely known, you may have to try: loren%sunlight.llnl.gov@star.stanford.edu