The idea of building mechanical devices that act like human beings is at least as old as the Queen’s Gardens in seventeenth century France, where a vast network of pipes carried pressurized water that made the statues move and even seem to speak. But it wasn’t until the development of modern electronic computers that the idea could be more than a fantasy. Early electronic automata or robots behaved in very stereotypic ways [IMAGES?], and had no ability to respond adaptively to changes in the environment. As processing power increased, researchers began to look for ways to give their creations genuine intelligence. One obvious strategy for building artificial intelligence or AI is to isolate the procedures that make humans so smart, and build them into machines. Since the most sophisticated of our abilities seem to depend on our capacity to reason, plan, and communicate with other thinkers, early roboticists tried to program the brains of robots to have just these same abilities.
The world-modeling approach to robotics
How do you give a robot human-like abilities? We are smart because we reason and plan, the argument goes, but we have to have something to reason about. That information comes from various sensors – eyes, ears, and so forth – that allow us to build up an internal map or representation of the world. Various logical routines or algorithms can then operate on that internal world model to plan actions and reason about how to best achieve the robot’s goals. Based on this analysis of the world model, actions are chosen and executed. This changes the world, and thus the input to the sensors, which requires that the world model be updated and the algorithms applied again to plan the next round of actions. This cycle can continue indefinitely, or until some particular goal is satisfied. [GRAPHIC THAT GIVES AN EXAMPLE]
Many famous robots were built and tested using this basic approach. The first of these to achieve any kind of success operated from 1966 to 1972 at SRI (the Stanford Research Institute). Named Shakey [IMAGE and/or VIDEO], for its tendency to wobble badly when stopping and starting, the robot could only operate in a carefully designed laboratory environment where its world was kept very simple. With a lot of work, it was eventually able to move blocks around and navigate it’s simplified world. The next generation of mobile robots, exemplified by the famous Stanford Cart [IMAGE and/or VIDEO– see www.frc.ri.cmu.edu/~hpm/book97/ch2/ -- we’ll need permission], were able to move outside of the laboratory and into the wider world. Unlike the ponderously slow Shakey, the Cart could achieve very high speeds, sometimes moving far faster than its minimal intelligence warranted, to the dismay of anyone who got in its way!
Problems with the world-modeling approach
Although in retrospect these researchers accomplished a great deal with very limited processing power, their successes were far fewer, and far less dramatic, than the early promises of the theoreticians, and some people began to question the entire approach. Fusing the input from several different sensors into a coherent model of the world is an enormously complicated computational task, and that’s before any reasoning or planning has begun. Each action changes the world, which means that the robot has to compute an updated model before it can take further action. This creates a serious information bottleneck, and makes it very difficult to respond quickly to events in the world. Since the point of having sensory systems is to allow actions to be directed and timely, this sort of bottleneck looked to some researchers like evidence that the whole approach was wrong. But if so, what is the alternative? Some researchers in the 1990s began to look back to much simpler robots, and simpler organisms, for inspiration.
The behavior-based approach to robotics
Human beings are arguably the most complicated, intelligent agents on the planet. Complicated systems are always harder to understand than simple ones, so why start by trying to copy human intelligence? Why not start with much simpler organisms for inspiration? That idea occurred to J. Grey Walter in the 1950s, and led to the creation of a very simple robot that he called, “Machina speculatrix.” [IMAGE ] With only a photocell for detecting levels of brightness and bump sensors around its shell, Grey Walter’s “tortoises,” as they were called, managed to do some surprisingly complicated things. But with the vacuum tube technology of the its day, it was very difficult to go much beyond these simple devices, and the approach vanished from the scene until revived in the 1990s.
A number of people contributed to the return to this simpler approach to robotics, but no one is more associated with the move than Rodney Brooks of MIT and his students. Brooks offered a new architecture for organizing information and guiding robotic behavior, one that did not require building internal world models at all. As Brooks put it, “the world is its own best representation,” so why bother to reconstruct it inside the robot? Instead, build the robot to respond in distinctive ways to the information as it comes in. Different categories of sensory inputs are defined, and each category, when it occurs, triggers a specific behavior. All that is needed is a rule that ranks the behaviors in importance, so that if more than one is triggered, the most important one always gets executed. Figure 2 shows how this alternative approach works.
The results were quite dramatic. Surprisingly simple robot programs were suddenly doing very complicated things, without having to solve the hard problem of keeping track of everything in the world. The behavior-based approach took off, and continues to inspire research in laboratories around the world. It is also the approach that you will learn about at this web site.
Problems with the behavior-based approach
The behavior-based approach to robot programming is not without its critics. Probably the most frequent complaint is that it cannot possibly scale up to more complex levels of intelligence. Just adding more and more behaviors to a list won’t be able to achieve the flexibility and ability to handle novel situations that makes human intelligence so powerful. Brooks disagrees, and has embarked on new projects designed to show how it might work [link to Cog project], but for the time being, the issue remains unsettled. There is, however, a new compromise position that is rapidly gaining new adherents.
Toward hybrid systems
A number of people have noticed that behavior-based systems are ideal for solving problems like avoiding collisions, tracking motion, responding quickly to sudden changes in the environment so forth, but not so good at long-range thinking and planning. World modeling systems, on the other hand, respond too slowly to rapid change but are excellent at thinking about how to achieve long-term goals. Why not find a way to combine the two, using each approach to solve the problems for which it is best suited? This hybrid approach is increasingly common in robotics, and many see it as the wave of the future. Indeed, many in cognitive science see it as the most promising approach for eventually understanding human intelligence. They argue that we are built in just this way ourselves. We handle many daily tasks – walking, reaching and grasping, noticing loud noises, and so forth – automatically, leaving precious resources like memory and attention for the more complicated problems we have to solve. The growing success of this approach suggests that knowing something about behavior-based robotics will continue to be important for understanding both robotics and cognitive science, even if it turns out not to be the whole story.
Learning more
Now that you know a little about the history of robotics, you’re ready to start learning about how behavior-based architectures actually work. Whether or not you have or plan to get or build a robot of your own, we have resources to help you learn more. If you want to read more about robots and robotics, you can go to any of the references below. Or, if you want to try a little exercise in robot ethology, just follow the link to the getting started page.
References Arkin Brooks stuff Hans Moravek ????