Robot Navigation by Conditional Sequencing (s~lbmiile(i to the 199 IEEE Confcrcncc on Robo

Robot Navigation to theon Robotics and Automation)of Michiganapplication of conditionalsequencingby layering a conditional behaviors. The approach is uniquely flexible, permitting verycomplex tasks toperiods of time. The tcchniquc was used in the mobile robot contest. All of thecontest-specific code was written in by a single programmer.The robot turned in the bestoverall performance of any entry. ofcircumstances.In earlier work wc formobiletoongoingArtificialtheinvestigations in this direction, asTo date wc have performedover ninety formalfor getting a robot to follow instructions.Conditional scqucnccs arc distinguished from the more classical notion of a plan in that theinstructions are noton the situation may includeconstraints on situations which do notis an To assemble a widget, locate PartA and PartB and a l/2-inch screw. A into the hole in Part R and the screw.Ifslide into sandpaper to remove any protruding bumps.This is a set of instructions, but it is not alinearis a constraint to bc applied simultaneously with the fastening operation. The last instruction is an example of a contingency procedure.It isFor example, locating for searching for an object, which carries with it its own constraints about where to search, and contingencyprocedures about what to do if the search2.2 softwareAllis a customincludes an assembler, anoperatingprogramming language [Gat91 ] which isuseddcscribcd in this paperconditional sequencing was done usingThe firstinconstruct. The secondusingbeenandescribed in more detail inOurbehaviors which arcused a building blocks for more complex activities.In this case wc arc concerned with navigation in anindoor environment where walls and doorways canThe robotobstacles, but must not relytoo heavily on dead reckoning.fastersquare ofFor indoor environments wc usc three basicbehaviors: dead-reckoning to a position whileavoiding obstacles, following walls, and aligning towalls. Thisslows down and veersaway from the obstacle. If the robotmove forward by a smallamount. This “directionalempirically.Wall-following is performed by simplyfollowing algorithm also stop for obstacles.(Reliably moving around obstacles while followinga wall is an unsolvedshown mathematically that the naivealgorithm oughtunstable.reliability.Wall alignment is done byread the samedistance, Experimentally this simple algorithmaligns the robotWall alignment is usedmisalignments in the robot’s wheels, The result isdramatically improved dead-reckoningHowever,to show thatconditional sequencing works. In fact, to testconditional sequencing the primitivebehaviors not work perfectlycanfrom failures, and thus produce reliable performance even when the primitives arc unreliable.In order tofor the robot tofailures whencanTo produce cognizant failure wcy.These monitor routines vary fromsituation to situation, but arc is outside of a bounding region.It also imposes a time limit on reaching thedestination, The timechanged by the to the goal,The wall-following monitor checks the robot’sheading it is not turning faster than also makessure that the distancetime limit on the alprogrammer ifThe dead-reckoning and wall-alignmentprimitives can more robust dead-reckoning and wall-a simplyby retrying the primitive if it fails. For an for circumnavigatingthe obstacle, the robot chooses a thean inefficient path. At some point the robot monitor indicates a cognizant failure when the robotface thegoal, at which obstacle is on thecourse,procedures.I basic sequence.A second example of a basic is no longer a wall tofollow. Ifsimply invokesrequires its own monitorto the building up successively more complex wc have beenable toin 1: The AAAI Robot ContestThe hardware and computational infrastructurecode was written in three days by a single programmer.The robot turned in the bestoverall performance of any entry in the eventsThetwo werethird boxes whichour smalltook place in a large arena made ofThe sequencing infrastructure used inwas ausing the catch/throw facility ofFor this simplified when a cognizant failure occurred. Wctherefore procedure )(f a i l u r e -c o d eof failure. This failure code was propagated upwards through nested with-recovery-procedureformsproccdurc forhad a robot’s motors off.On lop of this basic sequencing facility wcafrom the OfficeThe required the robot 10 explore amock-up office and find its waydoors, chosen was opened. The robot had to find the open door, go through it, and then navigate an obstacle field to afinish line some fifteen within two meters of the finish line.The robot was starred at a known orientation but at an unknown randomly chosen location within the Oft-lee.The strategy used was the following. The robot wandered randomly around the office for oncminute (the time that the doors were guaranteed to of x-y positions,Using this data it computed x-y positions that would guarantee that it was outside of the office. It then began to try each door in turnuntil it rcckoning primitive.The robot was tested about a dozen times during the preparation for the contest. It never failed, During the contest the judges consistently chose to open the door that happened to result in theworst-case performance.proceduresfor entering and exiting the arena autonomously.all wc had to do was write a singletop-levelruns) and worked every time. cardboard boxes were placed in the arena to serve as obstacles. The robots were allowedto have changed from run to run),Thewhich was in the arena and deliver it to a designated location, To accommodate robots without any manipulation hardware the robots were not required to actuallypick up theand indicate that it knewwas in, and thelocation of the destination.The robot was not told its initial location nor had tobegin by self-localizingThe robot began by searching for a wall (in amannerbegan following the wall while making and right turns it was making. Itturnsthisassemblies inits position after locating andcircumnavigating a wall plan awhich was located in roomC. where the found thesince the robotwas in the same room.) The robot then moved directly tothe deliveryrecorded at was only performed once.However, our robot was the only robot in thecontest to hours 10 program therobot for this event.4. Experiment 2: Trek to the Lab Theonly anecdotal data, and is thus of limited utility for drawing general conclusions about conditional sequencing. In order to provide somemore rigorous data wc set out to perform an wc used was the hallways of the building where wc work, a typical modern office building consisting of a maze of orthogonal hallways with doorways and random obstacles. We wanted to show that the performance of the robotimproved in aas theywere being tested.For thesewc used a slightly Action Package (RAP) systcm. RAPs arca sophisticated conditional sequencingis dcscribcd in detail inbridge the gapexecutionWc chose as our initial benchmark task meters away on theopposite side of the hallway. This turned out nottrials version of the software; all butonc worked room for improvement, wcchose a more difficult task.This time wcprogrammed the robot to navigate to ourWe defined a based on thedistance traveled before an procedures asproblems occurred. By the tenth trial the robot was Figure 4:Fifteen consecutive runs from the office to the lab.Wc once andprogrammedlab of the run was now nearly150 meters, Wc ran forly-sevenare shown inbut a clear upward trend in performancecantimeof(including the Work Wc have of the efficacy of conditional sequencing for controlling real-world autonomous mobile robots inindoor environments. Wc runs in a variety of situations using two different conditional sequencing infrastructures.In all cases, robust effectiveperformance waspreliminary despite the large number of trials relative to other published studied of this kind. Wc arc currently working on adesign for a rigorousconditional sequencing to run on small processors such as those available for planetary rovers. Some preliminarywork in this direction has already been doneof Technology undera contract withRonald C. Arkin,Navigation,” Robotics and Autonomous Systems, vol.6,“Using Parallel Program Specifications For Journal of Applied, MA, June 1992.for a Mobile Robot”,and Automation,Jonathan Applied to Robot Navigation,”unpublished manuscript,Executionthesis, Yale UniversityRoboticConference on Robotics and Automation, 1991,Erann Gat, “Low-compilationControl fora n d Automation, )991.AsynchronousNational Conference on Artificial IntelligenceTransactions on Robotics and Automation,Michael“A864, Yale UniversityReid Simmons, “AnGuy。