From: AAAI Technical Report SS-94-04. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. A METHODFOR IMPLEMENTING AI-BASED CONTROLIN A MANUFACTURINGWORKCELLUSING A BIOLOGICAL MODEL by Robert D. Borchelt, Ph.D. Departmentof Industrial and ManufacturingEngineering University of Wisconsin-Milwaukee P.O. Box 784 Milwaukee, WI 53201 (414)229-5898 borchelt@eonvex.csd.uwm.edu 1. Abstract Muchattention has beengiven in recent years to the capabilities of artificial intelligence tools and their possible usefulness in the monitoringand control of manufacturing workcells. This paper presents a technique which allows the use of AI-based control approachesin manufacturing workcells without the necessity of"speeiar’ computersor exotic control systems. The technique advocatedis explained using a biological modelas an analogy. 2. Introduction Control of manufacturingworkcells involves the coordination and sequencingof complextasks (amongother things) in whatis often a less than perfect domain.Strictly algorithmic approachesto control software have required rigid workenvironmentswith limited variability. Recognitionof the limits of strictly algorithmiccontrol methodshas led to a desire to incorporateartificial intelligence into control systems. A major problemfacing industry engineers, wlien they attemptto use AI tools in a "real-world"cell, is that the AI tools are designedto operate in a personal computer(PC), workstation, or mainframeenvironment. The manufacturing systemthat they are designingis often controlled by machine controllers, programmable logic controller’s, and/or perhaps a PCfor data handling. Machinecontrollers, PLC’s,and even standard robot controllers are normallytoo limited to allow implementation of AI-basedcontrol. If the designers chooseto give control of the cell overto a PCthen they often feel that they must sacrifice speed and must "custom-design"interface hardware and softwareto allow the PCto control the cell. Previousresearch in this area has focusedprimarily on a "single processor" model, wherethe software systems developedreside in a cell computeror other controller and utilize various methodologies to interact with the eeUin a somewhat"real-time" manner.Often, AI portions of the control softwareare "off-line" and only called up when 17 needed. The computingpower needed to complete the AI tasks is commonly obtained by placing the cell in a "hold" status until normaloperations can continue. Whilethis is certainly acceptablein somesituations, the necessity of "pausing" the systemis a serious constraint whichwould limit the widespreadapplication of AI in cell control. Even those systems whichallow multi-tasking in the cell computer wouldexperience serious degradationof response times if a significant AI task were taking place and tieing up system resources. Theobvioussolution for this is to incorporate morethan one processor in the control system. Researchon distributed computingand parallel processing has shownthe great powerof multi-processor computingin solving problems requiring quick solutions or massive computation. One of the serious limitations that this researchhas pointedout is that problemsolvers often don’t knowhowto break their problemdowninto pieces which can be solved by separate processors. Even given a problem which can be decomposed,the task of deciding which processor gets what information when,and whereeach processor should send its results whenit is fmishedis daunting. 3. A Model for ManufacturingControl In an attempt to describe howmulti-processor computingcould be implementedto support control of manufacturingworkcells, a biological analogy based on a simplified modelof the humanbrain will be used. Imaginea manufacturingcell with a cell computer(for this discussion PC), a programmable logic controller (PLC), and a robot controller, along with sensors and actuators necessaryto complete a manufacturingtask. The PCcould be consideredas the cerebral cortex of the system, the PLCas the brain stem, and the robot controller as the specialized portions of the brain dealing with motor functions. The sensors wouldbe similar to the sensory systems of the humanbody, and the actuators wouldbe similar to the involuntary musclesand other lower-level systems of the humananatomy(i.e. those systems not requiring free or gross motorskills). Assumingsuch a model, the tasks which should be assigned to each processor can be discerned (part of the reason for using the model)basedon the perceivedlevel at whichthe humanbrain wouldhandle a similar task. Usingthis model, the PCshould be responsible for controlling "higher level" problemssuch as assembly planning, error recovery, and other "knowledgeintensive" tasks. It is a moresophisticated"thinker" than the other processors, and can support AI-basedtools to handle these complexdecisions. Becauseit is handling larger "problems" it will respondslower and should not be "interrupted" by "housekeepingdetails". The PLC,on the other hand, should be responsible for handling"instinctual" or involuntaryreactions like the implementation,and subsequentfeedthrough of, an emergencystop. Anotherexamplemight be the timing of a pallet transfer. The sequenceof operations for a pallet transfer are well defined, simple, and easily accomplishedby a PLC.In addition, the task of transferring the pallet froma mainconveyorinto a transfer station is not one that can be closely controlled by an automatedsystem. The success and/or failure of a pallet transfer is muchmoredependenton the physical dimesionsand tolerances of the material handlingcomponents than on the instructions sent to it buy the controller. Ira problemdevelopsin the transfer, there is not likely to be muchthat an "intelligent" controller can do about it. These "housekeepingdetails" are time consuming but predictable and do not require the higher level "reasoning"ability of the PC.Delegatingthese tasks to the PLCflees up the PCfor its higher level duties. Theparallel nature and speed with whicha PLCcan detect a situation and implementa responsemakeit ideal for time-critical tasks. The "parallel" nature also makesa PLCwell suited for handling the manydetails which must be simultaneously considered during normaloperation. Finally, the robot controller should be used to focus on the tasks for whichit was originally designed. Namely, controlling the actions of the mechanicalarm. Whilerobot controllers often contain user definable inputs and outputs whichcan be used to sequenceoperations within a robot-centeredcell, they are not well suited for the role of cell controller. Sequencingtasks slow downoperation of the robot and the ability of standardrobot languagesto express sequencingconceptsis limited at best. If task planning, gripper control, sensor monitoring,etc., is all givenover to the other processorsin the system,then the robot controller can concentrate on quick and accurate responses to movementcommands.Writing robot programs in such an environment is greatly simplified as well. Anyspecialized controller could be incorporated in a similar fashion. Ira machinevision system is part of the sensor array, then the vision systemcomputershould be focused on processing the video imageand providing a 18 simplified abstraction. The guiding principles are foundin howthe brain handles similar tasks (or at least howwethink it does). The analogyis general enoughto be applicable in most manufacturingcontrol problems, but can be applied specifically enoughto be useful. 4. Conclusions Existing research often focuses on improvingrobot languagesor robot controllers in order to implementAIbased control schemesin robotic workcells. Most robot purchasersare industrial companieswhichare unwilling to pay for the "addedcapabilities" whichwouldbe part of such an "improved"robot system. Whetherthis occurs due to ignoranceof the potential advantagesof such a system, or dueto an actual lack of the needfor these capabilities, is beside the point. The fact is that robot manufacturerswill not offer these "features" until the customersstart demanding them. Havinggeneralized about "most"robot purchasers, it is nownecessary to address the aggressive and forwardthinking minority. Somecompaniesare exploring AI-based control, and they have to do it with equipmentmanufactured today. It is possible to do so, and one approachhas been explained here. The important concepts are not found in the specific discussion of whichtype of processor and/or device is given a certain task. The capabilities of each "robot controller" and "PLC"varies fromone vendorto the next. Rather, the importantconceptsare found in the division of tasks amongcooperating processors, with information being available whereneeded. The tradeoffs betweenspeed of decision-makingand complexityof the decision are important. It is not necessaryfor the "Ar’ portion of the control systemto consider all decisions. Onthe other hand, it is necessaryfor the AI portion to understandthe effects of those decisions and be able to notice whenthey cause errors. TheAI portion of a controller should be able to over-ride the "lower-level" decision makerif it deemsthat to be necessary, just like the brain can makeyou"hold your breath" even though you need air. The technologies involved will vary fromsystemto system, but the idea of structuring the levels of decision-makingaccording to a biological modelseems both feasible and easily understandable. The resulting control systemshould be both powerful and fast, and wouldtake advantageof the strengths of each of the controller platforms involved. Theresponse times of the systemwouldbe appropriate to the tasks encounteredrather than "uniformlyfast" but required to be simple, or "uniformly slow" but "thoroughlyconsidered". Overall, this model should provide a guideline for implementing multiple-processor computingin the manufacturingcontrol environment.