The NIST Real - time Control System ( RCS )

The Real-time Control System (RCS) architecture developed at NIST and elsewhere over the past two decades [1] defines a canonical form for a nested intelligent control system. The RCS architecture consists of a hierarchically layered set of processing modules connected together by a network of communications pathways. The primary distinguishing feature of the layers is the bandwidth of the control loops. The characteristic bandwidth of each level is determined by the spatial and temporal integration window of filters, the temporal frequency of signals and events, the spatial frequency of patterns, and the planning horizon and granularity of the planners that operate at each level. At each level, tasks are decomposed into sequential subtasks, to be performed by cooperating sets of subordinate agents. Signals from sensors are filtered and correlated with spatial and temporal features that are relevant to the control function being implemented at that level. The four basic types of processing modules from which the RCS architecture is built are: 1) Behavior Generating (BG) modules BG modules contain job assignment, planning, and control algorithms. These embody knowledge about how to perform tasks -i.e., how to decompose tasks into subtasks that subordinate agents know how to execute. BG modules can accommodate a variety of planning algorithms, from simple table look-up of pre-computed plans, to real-time search of configuration space, or game theoretic algorithms for multi-agent cooperating and competitive groups. Planning horizons at high levels may span months or years, while planning horizons at the bottom level typically are less than 50 milliseconds. Control loop bandwidth at each level is typically at least ten times the reciprocal of the planning horizon at that level. 2) World Modeling (WM) modules The WM modules model the state space of the problem domain. They contain information storage and retrieval mechanisms, as well as algorithms for transforming information from one coordinate system to another. WM modules use dynamic models to generate expectations, and predict the results of current and future actions. WM modules may contain recursive estimation algorithms and processes that compute lists of attributes from images, graphics engines that generate images from symbolic lists, and storage and retrieval algorithms that perform and maintain both short term and long term memory about features, surfaces, objects, and groups. The WM module maintains a knowledge database (KD), acts as a question answering system, and uses information from the knowledge database to predict or simulate the future. 3) Sensory Processing (SP) modules SP modules process data from visual, auditory, tactile, proprioceptive, taste, or smell sensors. SP modules contain filtering, masking, differencing, correlation, matching, and recursive estimation algorithms, as well as feature detection and pattern recognition algorithms. Interactions between WM and SP modules can generate a variety of filtering and detection processes such as Kalman filtering and recursive estimation, Fourier transforms, and phase lock loops. Vision system SP modules process images to detect brightness, color, and range discontinuities, optical flow, stereo disparity, and utilize a variety of signal detection and pattern recognition algorithms to analyze scenes and compute information needed for manipulation, locomotion, and spatial-temporal reasoning. 4) Value Judgment (VJ) modules VJ modules contain algorithms for computing cost, risk, and benefit, for evaluating states and situations, and alternatives courses of action for estimating the reliability of state estimations, and for assigning cost-benefit values to objects and events. VJ modules may compute Beysian and Dempster-Schafer statistics on information about the world based on the correlation and variance between observations and predictions. The world modeling module maintains a set of: Knowledge Database (KD) modules KD modules consist of data structures that contain state variables, iconic images, and symbolic frames containing lists of attributes. Information in the KD includes knowledge about entities and events, and about how the world behaves, both logically and dynamically. The KD contains both short term and long term memory elements. The KD is typically implemented in a distributed fashion, suitable for real-time data retrieval and update. The entire system is interconnected by: A communication system that conveys messages between the various modules The communication system provides a network of pathways that transmits messages between the various processing and database modules. The communications system richly, but not completely, interconnects the modules, i.e. every module is not connected to every other module. The various modules in the RCS architecture act as a collection of intelligent agents (or software objects), sending and receiving messages to and from each other. These messages convey commands and requests, and return status. The RCS architecture has evolved over the past two decades from a rather simple robot control schema to a reference model architecture for intelligent system design. From the beginning, RCS has represented a conscious attempt to emulate the function and structure of the neurological machinery in the brain. Each RCS module has properties that are known, or hypothesized, to exist in the brain. For example, RCS modules may be constructed from neural nets such as CMAC [2] that compute arithmetic and/or logical functions on a set of inputs to produce a set of output state variables. These can be carried over communications pathways to other functional modules that may use them to perform further functional computations, or to generate addresses, or to store information in memory for latter use. RCS functional modules may add, subtract, multiply, differentiate, integrate, compute correlation functions, recognize patterns, generate names or addresses of symbolic representations, or perform planning functions at a hierarchy of levels. In its most complete theoretical form, the RCS reference model architecture provides a framework for integrating concepts from artificial intelligence, machine vision, robotics, computer science, control theory, operations research, game theory, signal processing, filtering, and communications theory. The RCS architecture has been used in the implementation of a number of experimental projects.