| Behavior-based Robotics | |
| |
|
Almost thirty years from the famous "Darthmouth Conference" of 1956, a six-week
long workshop organized by John McCarthy that laid the foundations of AI, researchers
realized that the idea of computers as intelligent machines was somewhat inappropriate.
The vast majority of AI scientists assumed that some kind of high level reasoning
was essential for intelligent behavior. This approach tried to model and analyze
the logical chain of human thoughts, from problem to symbolic representation,
processing, and solution. This classic paradigm, known as Good Old Fashioned
Artificial Intelligence (GOFAI), failed at making adaptive, mobile and robust,
true real-world robots. It uses functional task decomposition, where perception,
planning, and action are done sequentially. Traditional AI approaches are unfortunately
so tied to the traditional, Von Neumann machine architectures that it seems we've
already reached the top: it is a time to investigate other solutions.
The human or animal brain does not run "programs" or "algorithms": its intelligence manifests itself in various behaviors and interaction with the outside world. In an attempt to understand intelligent behavior, a small group of researchers and visionaries turned the field of AI upside-down in the mid-1980s by introducing the behavior-based approach to robotics. It was pioneered by Rodney Brooks> from MIT and inspired by Braitenberg's "Vehicles: Experiments in Synthetic Psychology". This book demonstrated how seemingly intelligent creatures could emerge from interaction of very simple reactive controllers. Behavioral decomposition decentralizes and divides tasks into different behavior "layers" running in parallel. There are no explicit world models and separation of data and computation - they are both distributed over the same network of simple elements. Layers can be added incrementally, and each behavior is reactive to the sensor input and the states of the "lower" layers. Every layer has the ability to override or subsume the behaviors of the lower layers, thus the term subsumption architecture comes into play.
The subsumption architecture provides a simple, incremental method for building robot control systems linking perception to action. A carefully designed network of finite state machines augmented with internal timers (called Augmented Finite State Machines, or AFSMs) can provide a robot with impressive repertoire of behaviors, without the computing overhead and poor performance so common in GOFAI systems. Technical description of the inner-workings and implementations of subsumption architecture can be found in numerous papers. The book "Cambrian Intelligence: The Early History of the New AI" presents and excellent overview of the key philosophical and technical ideas that made this "bottom-up" approach so attractive to both researchers and hobbyists.
The key concepts behind behavior-based robotics are sometimes controversial,
but almost always very simple and elegant. For example, when discussing intelligence
(and we all know this is a tough term to describe, just like consciousness,
emotions, etc.) Brooks describes Herbert Simon's example of an ant walking along
the beach. The complexity of the ant's behavior is more a reflection of the
complexity of its environment than its own internal complexity. The same may
be true with humans, and it is often hard to draw the line at where is intelligence,
and what is environmental interaction. The key ideas are:
"New AI" has now achieved success around the world. Of course, "pure" behavior based approach has its shortcomings, namely the lack of predictability (how can you determine what will your robot do next?) and exploding complexity when adding new skills and behaviors. These problems can be solved by taking the best of several approaches, so we now have behavior based robots in entertainment, service industries, the military, agriculture, mining, and even the home. An excellent book and a kind of robot builder's Bible, "Mobile Robots: Inspiration to Implementation" by Jones and Flynn, describes all aspects of mobile robot design with an emphasis on the subsumption architecture theory and implementation in C. Brooks has teamed with several colleagues to found iRobot Corporation - aka IS Robotics. They are producing all kinds of all-terrain and indoor robots, but probably the most interesting for the wide audience is their iRobot-LE, the first multi-purpose home robot controllable from a Web browser.