Teleo-Reactive Programs and the Triple-Tower Architecture.

摘要: I describe an architecture for linking perception and action in a robot. It consists of three “towers” layered components. The “perception tower” contains rules that create increasingly abstract descriptions the current environmental situation starting with primitive predicates produced by robot’s sensory apparatus. These are deposited “model which is continuously kept faithful to “truthmaintenance” system. model tower, turn, evoke appropriate action-producing programs “action tower.” proposed actions be written as “teleo-reactive” programs---ones react dynamically changing situations ways lead inexorably toward their goals. Programs tower organized more-or-less hierarchically---bottoming out cause robot take its environment. effects sensed mechanism, completing sense-model-act cycle quiescent only at those times when goal perceived satisfied. illustrate operation using simple block-stacking task. I. Agent Architectures Can anything general said about intelligent agent architectures? Just there millions species animals, occupying different niches, expect will many artificial agents---each specialist one countless number tasks. exact forms architectures depend on tasks environments. For example, some work time-stressed reactions unpredictable states must fast unequivocal. Others have time knowledge predict future courses so more rational choices can made. Even though probably never single, all-purpose architecture, think might play prominent role systems. viewed elaboration 1 Parts this section adapted from Chapter 25 my book, Artificial Intelligence: A New Synthesis, San Francisco: Morgan Kaufmann, 1998. 2 first two levels popular three-level been robotics research. A. Three-Level One integrated systems was collection computer hardware known “Shakey Robot” (Nilsson, 1984). Shakey's design early example what has come called architecture. correspond paths signals motor commands. At lowest level such use short path effectors. Important “reflexes” handled pathway---such “stop” touch sensors detect close object ahead. Servo control motors achieving set-point targets shaft angles also these low-level mechanisms. intermediate combines low into complex behaviors---ones whose realization depends (as modeled) execution. This uses (or “coarse”) perceptual than do lower ones. Whereas reflex typically evoked signals, coordination intermediate-level requires elaborate processing. third usually involves generate plans consisting sequence programs. used variety As typical see (Connell, 1992). B. Triple-Tower Architecture generalization Albus colleagues (Albus, 1991; Albus, McCain, & Lumia, 1989). They envision hierarchies or perceptual, modeling, We propose here particular instantiation triple-tower novel features our proposal are: 1. teleo-reactive 2. tower. simpler ones 3. truth-maintenance system (TMS) keep changes environment My version illustrated Figure would proceed follows: Aspects relevant agent’s roles converted values. stored Their presence may immediately bottom actions, affect environment, sensed---creating loop itself important computational role. 3 convert then higher processes continue until even highest populated. Fig. loose hierarchy routines triggered contents reflexes---evoked corresponding percepts. More actions. High-level “call” other process bottoms actually allow possibility themselves directly (in addition through environment) writing additional and/or altered content. With ability both read write memory, structure perfectly Sensors Model Tower (Predicates + TMS) Perception (Rules) Action (Action Routines)