Until very recently, most AI work applied to robotics has been of the symbolic variety. Symbolic AI has had most of its success dealing with problems at the ``cognitive level'': e.g., theorem proving, medical diagnosis, mineral prospecting, X-ray crystallography, game playing, etc. Autonomous systems designers and other roboticists have tried to exploit these techniques, but have often suffered from the mismatch between what AI systems are good at and what is needed for controlling a robot in real-world environments. Symbolic AI works best for well-defined, representable problems. On the other hand, what practical autonomy needs are systems which exhibit robust behaviors in ill-defined, poorly-constrained, and often unfriendly environments.
In an effort to solve these very difficult problems facing autonomous systems, a new approach has recently emerged and is still being defined. It has been given many names by different researchers, including the ``animat'' approach [\protect\citenameWilson, 1985], reactive planning [\protect\citenameGeorgeff &Lansky, 1987][\protect\citenameFirby, 1987][\protect\citenameAgre &Chapman, 1987], computational neuroethology [\protect\citenameCliff, 1991], and the task-oriented subsumption architecture [\protect\citenameBrooks, 1986]. Maes ##1[][]maes93 refers to these various approaches by a common underlying feature, calling it behavior-based AI.
Behavior-based AI is a bottom-up approach to autonomy that most often avoids declarative representation of domain knowledge. In place of knowledge structures and an ``inference engine'' are behavior-producing modules which link specific environmental inputs to specific reactions, or behaviors. This keeps the system in intimate contact with its environment: it is ``situated'' in that environment spatiotemporally. The behavior modules operate in parallel, with their outputs contributing to the overall system's ``emergent'' behavior. The behavior-based approach stresses the system's autonomy as a basic fact of the system's existence. It has already produced some remarkable results, both in terms of robots carrying out complex behaviors while coexisting with humans in a laboratory environment [e.g.,]brooks86 as well as in our own domain of controlling AUVs [\protect\citenameBonasso, 1992][\protect\citenameBellingham, 1992][\protect\citenameBellingham et al., 1990].
A problem with existing behavior-based control architectures is that their behaviors are almost exclusively low-level perception-action linkages that are little more than simple to moderately complex reflexes. While behaviors at this level are crucial for a real-world autonomous agent, they are not the sum total of behaviors needed. A useful agent must also be able to create and follow both near-term and long-range plans of action, assess the current situation in terms of how it is likely to impact the vehicle's health and mission, coordinate its activities with other agents with whom they are cooperating, and interact with human users. Any generic behavior-based approach to autonomous control must also pay attention to these high-level behaviors that are just as much a part of an agent's behavioral repertoire as the low-level ones. In addition, one of the very characteristics that make behavior-based AI interesting can also get in the way of creating useful autonomous systems: emergent behavior can be quite opaque to human observers, which can both make human users uncomfortable (or even mistrustful of the system) and make a principled approach to system design difficult. Although a growing number of researchers are beginning to relax their stance on the representation of knowledge and goal-directed behavior [e.g.,]maes90,mataric92, none have yet completely bridged the gap between what behavior-based AI can do and what is needed for useful autonomous systems.
The generic behavior (GB) approach attempts to unify symbolic and behavior-based AI for autonomous vehicle control. From behavior-based AI, we draw the idea of using discrete behaviors operating in parallel to achieve robust, situated control. From symbolic AI, we draw goal-directed behavior, explicit representation (including such representation of the behaviors themselves), and deliberative reasoning.