ebook img

Fast Anytime Motion Planning in Point Clouds by Interleaving Sampling and Interior Point ... PDF

16 Pages·2017·2.04 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Fast Anytime Motion Planning in Point Clouds by Interleaving Sampling and Interior Point ...

Fast Anytime Motion Planning in Point Clouds by Interleaving Sampling and Interior Point Optimization AlanKuntz,ChrisBowen,andRonAlterovitz UniversityofNorthCarolinaatChapelHill ChapelHill,NC27599-3175,USA {adkuntz,cbbowen,ron}@cs.unc.edu Abstract. Roboticmanipulatorsoperatinginunstructuredenvironmentssuchas homesandofficesneedtoplantheirmotionsquicklywhilerelyingonreal-world sensors,whichtypicallyproducepointclouds.Toenableintuitive,interactive,and reactiveuserinterfaces,themotionplancomputationshouldprovidehigh-quality solutionsquicklyandinananytimemanner,meaningthealgorithmprogressively improvesitssolutionandcanbeinterruptedatanytimeandreturnavalidsolu- tion.Toaddressthesechallenges,wecombinetwoparadigms:(1)asymptotically- optimalsampling-basedmotionplanning,whichiseffectiveatprovidinganytime solutionsbutcanstruggletoquicklyconvergetohighqualitysolutionsinhighdi- mensionalconfigurationspaces,and(2)optimization,whichlocallyrefinespaths quickly.Weproposetheuseofinteriorpointoptimizationforitsabilitytoperform inananytimemannerthatguaranteesobstacleavoidanceineachiteration,andwe provideanovellazyformulationthatefficientlyoperatesdirectlyonpointcloud data.Ourmethoditerativelyalternatesbetweenanytimesampling-basedmotion planning andanytime, lazyinterior pointoptimization tocompute highquality motionplansquickly,convergingtoagloballyoptimalsolution. 1 Introduction Robotic manipulators are entering unstructured environments, such as homes, offices, hospitals, and restaurants, where robots need to plan motions quickly while ensuring safetyviaobstacleavoidance.Motionplanninginsuchsettingsischallenginginpartbe- causetherobotmustrelyonreal-worldsensorssuchaslaserscanners,RGBDsensors, orstereoreconstruction,whichtypicallyproducepointclouds.Inaddition,enablingin- tuitive,interactive,andreactiveuserexperiencesrequiresthattherobotgenerateplans ofhighqualityasquicklyaspossible,withoutnecessarilyknowinginadvancethemax- imum time allocatable to motion planning. Hence, motion planning in such settings shouldbeimplementedasananytimealgorithm,meaningthealgorithmprogressively improvesitssolutionandcanbeinterruptedatanytimeandreturnavalidsolution. Toenablefast,anytime,asymptotically-optimalmotionplanninginpointclouds,we proposetoblendtwopopularmotionplanningparadigms:(1)sampling-basedmotion planningand(2)optimization-basedmotionplanning.Sampling-basedmotionplanners candirectlyusepointclouddatabyincorporatingappropriatecollisiondetectionalgo- rithms [19] and can optimize a cost metric in an asymptotically optimal manner [12]. 2 AlanKuntz,ChrisBowen,andRonAlterovitz However,sampling-basedmotionplannersinpracticecanbeslowtoconvergeandfre- quentlyreturnpathsinfinitetimethatarefarfromoptimal,especiallyforproblemswith higher dimensional configuration spaces [10]. In contrast, optimization-based motion planners are often very fast [21,23,20,11], but typically do not converge to globally optimal solutions, do not operate in an anytime manner (since, even when initialized withafeasiblesolution,theirintermediateiterationsmayconsiderrobotconfigurations thatareincollisionwithobstacles),andareinefficientwhenusinglargepointclouds. Inthispaper,weintroduceanewoptimizationapproachtomotionplanningbasedon interior point optimization [25,24], which has unique properties that enable anytime motion planning and efficient handling of large point clouds. By integrating our new interior point optimization formulation with a sampling-based method, our approach quickly computes locally optimized plans in an anytime fashion, and continues to re- finethesolutiontowardglobaloptimalityforaslongastimeallows. Ournewmethod,InterleavedSamplingandInteriorpointoptimizationMotionPlan- ning (ISIMP), interleaves global exploration with local optimization (see Fig. 1). The methodstartswithglobalexplorationbybuildingagraphusinganasymptoticallyopti- malmotionplanner,suchask-nearestProbabilisticRoadmap(PRM*)[12],untilitfinds acollision-freepath.Itthenusesourlazyinteriorpointoptimizationformulationtore- finethepathfoundbythesampling-basedmethod.Thealgorithmtheniteratesbetween (1) resuming the sampling-based motion planner until a new better path is found and (2) running interior point optimization on this new path. The sampling-based motion planningphaseofeachiterationexploresglobally,discoveringotherhomotopyclasses in configuration space, as well as escaping local minima in the path cost landscape. Theinteriorpointoptimizationphaseinturnallowsthemethodtoprovideahighqual- itylocally-optimizedmotionplanbasedonthebestpathfoundbythesampling-based method. Either phase can be interrupted at any time and still return a valid, collision- freeresult.Interleavingthesephasesprovideshigherqualitymotionplansearlierthan asymptoticallyoptimalsampling-basedmotionplanningalgorithmsalone.Inthisway, our method preserves guarantees which are not provided by most optimization-based motionplanningalgorithms—namelycompletenessandasymptoticoptimality—allin ananytimefashionanddesignedspecificallytoworkefficientlyonpointclouddata. ISIMPisbasedoninteriorpointoptimization,atypeoflocaloptimizationthathas propertiescriticaltoachievingfast,anytimemotionplanningforroboticmanipulators. Interiorpointmethodssolveoptimizationproblemsbyonlyconsideringfeasiblepoints in each iteration of the method. These methods typically incorporate constraints into the objective function by adding barrier functions, continuous functions whose value on a point increases to infinity as the point approaches the boundary of the feasible region. These methods adjust weights on the barrier functions over time so the so- lution converges to a locally optimal solution of the original objective function. We applyinteriorpointoptimizationtorefiningmotionplansandshowthatitisfast(like othermotionplanoptimizers,e.g.,[21,23,20,11,13,18])whileprovidingthreeaddi- tionalimportantproperties.First,ourinteriorpointoptimizationimplementationisan anytimealgorithm:eachintermediateiterationoftheoptimizationalgorithmconsiders onlycollision-freerobotconfigurations,sotheoptimizeralwaysreturnsafeasiblemo- tionplanevenwheninterrupted.Second,interiorpointoptimizationcanbeformulated FastAnytimeMotionPlanninginPointClouds 3 S G S G S G (a)SampledFirstPath (b)OptimizedFirstPath (c)SampledSecondPath S G (d)OptimizedSecondPath (e)SampledPathonBaxter (f)OptimizedPathonBaxter Fig.1. ISIMP interleaves sampling-based and optimization-based motion planning. (a) ISIMP firstperformssampling-basedmotionplanninguntilafeasiblemotionplanisfoundfromstartto goal.(b)Itthenusesinteriorpointoptimizationtolocallyoptimizetheplan.(c)Sampling-based motion planning resumes until a shorter plan is found. (d) The shorter plan is then optimized. Themethoditeratesinthisfashionandcanbeinterruptedatanytimeafterstep(a)andreturn afeasiblesolutionthatgetsbetterovertime.(e)Anexamplesampling-basedplanfoundfora BaxterrobotinapointcloudsensedbyMicrosoftKinect.(f)Thatplanlocallyoptimized. toefficientlyanddirectlyhandlelargepointclouds,whichweaccomplishusingalazy evaluation of constraints. This eliminates the time-consuming process of transform- ingthepointcloudintomorecomplexgeometricprimitivesormeshes,whichisoften required by other optimization-based motion planners to be efficient. Third, our inte- riorpointoptimizationisdesignedtooptimizepathlength,acommonlydesiredmetric which is often used by sampling-based motion planners. Many existing optimization- basedmotionplanners[21,23,20,11]aredesignedtooptimizemetricssuchassmooth- ness,whichdonotsatisfythetriangleinequalityandhenceareincompatiblewithcost requirementsofasymptotically-optimalsampling-basedmotionplanners. We evaluate the efficacy of ISIMP for a Baxter robot’s manipulator arm using a MicrosoftKinecttosensepointclouddatainaclutteredenvironment.Wedemonstrate ISIMP’sfast,anytime,andasymptoticallyoptimalperformanceincomparisontoother motionplannersthatonlyuseeithersamplingorlocaloptimization. 2 RelatedWork Insampling-basedmotionplanning,agraphdatastructureisconstructedincrementally viarandomsamplingprovidingacollision-freetreeorroadmapintherobot’sconfigu- rationspace.Thesealgorithmsprovideprobabilisticcompleteness,i.e.,theprobability offindingapath,ifoneexists,approachesoneasthenumberofsamplesapproachesin- finity.ExamplesincludetheRapidly-exploringRandomTree(RRT)[16]andtheProb- 4 AlanKuntz,ChrisBowen,andRonAlterovitz abilistic Roadmap (PRM) [13] methods. These methods have been adapted in many ways,includingtotakeadvantageofstructureintherobot’sconfigurationspaceandby modifyingthesamplingstrategy[6,15]. Adaptations of these algorithms can provide asymptotic optimality guarantees in which the path cost (e.g., path length) will approach the global optimum as the num- berofalgorithmiterationsincreases.ExamplesincludeRRT*andPRM*[12]wherethe underlyingmotionplanninggraphiseitherrewiredorhasasymptoticallychangingcon- nectionstrategies.Otherasymptoticallyoptimalalgorithmsgrowatreeincost-to-arrive space[10],identifyverticeslikelytobeapartofanoptimalpath[1],orinvestigatethe distributionsfromwhichsamplesandtrajectoriesaretaken[14].Lazycollisioncheck- inghasbeenshowntosubstantiallyimprovethespeedofthesealgorithms[9],andin somecases,nearoptimalitycanbeachievedwhileimprovingspeed[22,7]. Optimization-basedmotionplanningalgorithmsperformnumericaloptimizationin a high dimensional trajectory space. Each trajectory is typically encoded as a vector of parameters representing a sequence of robot configurations or controls. A cost can becomputedforeachtrajectory,andthemotionplanner’sgoalistocomputeatrajec- tory that minimizes cost. In the presence of obstacles and other constraints, the prob- lemcanbeformulatedandsolvedasanumericaloptimizationproblem.Examplesin- cludetakinganinitialtrajectoryandperforminggradientdescent[21],usingsequential quadraticprogrammingwithinequalityconstraintstolocallyoptimizetrajectories[23], and combining optimization with re-planning to account for dynamic obstacles [20]. Thesemethodstypicallyproducehighqualitypathsbutarefrequentlyunabletoescape localminima,andassucharesubjecttoinitializationconcerns.Toavoidlocalminima, somemethodsinjectrandomness[11,3].Thesolutionsoftheseexistingoptimization- basedmotionplannersmayconvergetolocallyoptimalmotionplansthatincluderobot configurations that are in collision with obstacles. This limitation can be partially cir- cumventedthroughtechniquessuchasrestartingtheoptimizationwithmultiplediffer- entinitialpaths(e.g.,[23]),butsuchapproachesareheuristicandprovidenoguarantee thatacollision-freetrajectorywillbefoundingeneral. Severalmethodsaimtobridgethegapbetweensampling-basedmethodsandopti- mization.Somemethodsusepathsgeneratedbyglobalplannersandrefinethemusing shortcuttingorsmoothingmethodsadaptivelyorinpostprocessing[18,8].GradienT- RRTmovesverticestolowercostregionsusinggradientdescentduringtheconstruction of an RRT [2]. More recently, local optimization has been used on in-collision edges duringsampling-basedplanningtobringthemoutofcollisionandeffectivelyfindnar- rowpassages[4].Incontrast,ourmethodisusingthelocaloptimizernottofindnarrow passages,buttoimprovetheoverallqualityofthepathsfoundbythesampling-based planner, while relying on the sampling-based planner’s completeness property to dis- cover narrow passages. BiRRT-Opt [17] utilizes a bi-directional RRT to generate an initial trajectory for trajectory optimization, demonstrating the efficacy of a collision- freeinitialsolutionforlocaloptimization.Ourmethoddiffersinthatinteriorpointopti- mizationprovidescollision-freeiteratesallowingittoworkinananytimefashion,and ourmethodcontinuesbeyondasinglelocaloptimizationtoprovideglobalasymptotic optimality. FastAnytimeMotionPlanninginPointClouds 5 3 ProblemDefinition LetC be theconfiguration spaceof therobot. Let q∈C represent asingle robotcon- figurationofdimensionalityd andΠ ={q ,q ,...,q }representacontinuouspath 0 1 n−1 inconfigurationspaceinapiecewiselinearmannerbyasequenceofnconfigurations. Suchpathsmayneedtosatisfygeneralized,user-definedinequalityconstraints.These constraintscouldincludejointlimits,endeffectororientationrequirements,etc,where each constraint may be represented by the inequality g(Π)≥0 for some constraint specificfunctiong.Letthesetofallsuchuser-definedconstraintsbeJ. In the robot’s workspace are obstacles that must be avoided, and which are being represented by a point cloud. We consider each point in the point cloud an obstacle andletthesetofallsuchpointsbeO.Wealsodefinetherobot’sgeometryasasetof geometricprimitivesS.Anindividualgeometricprimitives∈Scouldberepresentedas amesh,boundingsphere,capsule,etc. A path is collision-free if the robot’s geometry S over each edge (q,q ),i = i i+1 0,...,n−2, does not intersect an obstacle o ∈ O. Formally, we require a function clearance(q,q ,s,o)whichisafunctionparameterizedbyapathedge(q,q ), i i+1 i i+1 a geometric primitive s, and an obstacle o. This function, formally defined in Sec. 4, is continuous and monotonically increasing, has positive value when o is not in col- lision with s over the edge, negative value when o penetrates into s on the edge, and 0 at the boundary. A collision-free path then becomes a path for which each edge has non-negative clearance for all o∈O and s∈S. This requirement can be repre- sented as an additional set of inequality constraints, which we define as K, wherein clearance(q,q ,s,o)≥0,forallo∈Oands∈S. i i+1 Our objective is to find a collision-free path for the robot from the robot’s start configurationq toauserspecifiedgoalconfigurationq thatsatisfiesallthecon- start goal straintsandminimizespathlength.Wedefinepathlengthbylength(Π),thesumof Euclideandistancesinconfigurationspacealongthepath.WeusethesumofEuclidean distances because it is a commonly used metric for path length and because it satis- fiesthetriangleinequalitypropertyrequiredbyasymptoticallyoptimalsampling-based motionplanners(unlikeothermetricssuchassumofsquareddistances).Thisoptimal motion planning problem can be formulated as a nonlinear, constrained optimization problem: Π∗=argminlength(Π) Π Subjectto: clearance(qi,qi+1,s,o)≥0, 0≤i<n−1, ∀o∈O, ∀s∈S (1) g(Π)≥0, ∀g∈J q =q 0 start q =q n−1 goal whereΠ∗isanoptimalmotionplan.Tosolvethisoptimizationproblem,wepresentan efficient,anytime,iterativealgorithminwhichthesolutionasymptoticallyapproaches Π∗. 6 AlanKuntz,ChrisBowen,andRonAlterovitz 4 Method ISIMPintegratesideasfrombothsampling-basedandoptimization-basedmotionplan- ning.Themethodinterleavesinteriorpointoptimizationstepswithsampling-basedmo- tionplanningstepstoachievefastconvergence.Thetoplevelalgorithm,Alg.1,runsin ananytimemanner,iteratingastimeallowsandstoringthebestpathfoundasitruns. Algorithm1:ISIMP Input:qstart,qgoal,O,optimizationconvergencethreshold∆Π Output:motionplanΠ∗ 1 Sampling-basedmotionplannerpathcostcsamp←∞ 2 Π∗←0/ 3 whiletimeremainingdo 4 Πsamp←GlobalExplorationStep(O,csamp,qstart,qgoal) 5 csamp←cost(Πsamp) 6 Π∗←minCostPlan(Πsamp,Π∗) 7 Π∗←LazyAnytimeInteriorPointOptimization(Πsamp,O,∆Π,Π∗) 8 end The first step of each iteration is the global exploration step, wherein ISIMP exe- cutesasampling-basedmotionplanneruntilanewpathisfoundthathascostlowerthan c ,thebestcostfoundbythesampling-basedmotionplannerupuntilthatpoint.The samp sampling-basedmotionplannerreturnsthenewpathΠ andupdatesc .Inthesec- samp ondstepofeachiteration,theinteriorpointoptimizationstep,ourmethodexecutesan interiorpointlocaloptimizationmethodtolocallyoptimizeΠ.Thelocaloptimizercon- stantlycomparesagainstthebestpathfound,andupdatesitifabetterpathisfounddur- ingtheoptimization.Attheendoftheinteriorpointoptimizationitreturnsthebestpath found. The algorithm then iterates, returning to global exploration with the sampling- basedmotionplanner,andoptimizingbetterpathsastheyarefound.Ifthealgorithmis interrupted,itreturnsthebestpathfoundupuntilthattime. 4.1 GlobalExplorationusingSampling-basedMotionPlanning The global exploration step uses a sampling-based motion planner to expand a graph untilanewpathisfoundthatisoflowercostthananypathithaspreviouslyfound.The sampling-basedmotionplannermaintainsagraphG=(V,E),whereV isasetofver- ticeswhichrepresentcollision-freeconfigurationsoftherobotandE isasetofedges, whereanedgerepresentsacollision-freetransitionbetweentworobotconfigurations. To expand the graph G, our method is designed to use an asymptotically optimal sampling-based motion planner. Although many methods would work, we implement ISIMPwithbothk-nearestProbabilisticRoadmap(PRM*)[12]andRRT# [1].PRM* samplesrandomconfigurationsintherobot’sconfigurationspace,locatestheirknear- est neighbors (where k changes as a function of the number of vertices in the graph), FastAnytimeMotionPlanninginPointClouds 7 and attempts to connect the configurations to each of its neighbors. RRT# is a state- of-the-artasymptoticallyoptimalsampling-basedmotionplanner,whichincrementally buildsatreeidentifyingwhichverticesarelikelytobelongtoanoptimalpath.Ineach globalexplorationstep,theasymptoticallyoptimalsamplingbasedmotionplannerex- ecutesuntilitfindsacollision-freepathbetterthanthecurrentbest,atwhichpointthe optimizationstepbegins. 4.2 LazyInteriorPointOptimization Local optimization for motion planning is typically performed by constructing a high dimensional vector out of the state variables of the problem to be optimized. In our case, this would be a vector representing the motion plan we are optimizing, i.e. if we have a path with n configurations of dimensionality d, then each motion plan can be represented as an n×d dimensional vector. Optimization can then be viewed as iterativelymovingthisvectorthroughitshighdimensionalspacetominimizethecost. Inthecaseofconstrainedoptimization,thereareregionsofthisspacewhichrepresent areaswheretheconstraintsaresatisfied(i.e.,feasibleorunconstrainedspace)andareas wheretheconstraintsarenotsatisfied(i.e.,infeasibleorconstrainedspace). Manyotherclassesofoptimizationmethods(suchaspenaltymethodsandsequen- tialquadraticprogramming)allowtheirintermediatesolutionstomovethroughinfeasi- blespace,i.e.collidewithobstaclesorviolatejointlimits,withthehopethattheeven- tuallocallyoptimalsolutionisfeasible.Interiorpointmethods,bycontrast,requiretheir intermediate solutions to always be feasible [25,24]. This is typically accomplished throughtheuseofbarrierfunctions.Abarrierfunctionworksbyintroducingacontinu- ousfunctionwhosevalueapproachesinfinityattheedgeoftheconstrainedspace.The intermediate iterations of optimization are then influenced by the barrier functions to avoid the constrained regions. As the optimization iterates, the width of these barrier functionsisreducedsuchthatthesolutionisallowedtoapproachtheconstrainedspace but not to enter it. This is the property that we leverage to create an anytime solution inside the optimization framework. Each intermediate solution of the optimization is alwaysacollision-freemotionplanandassuchcanbereturnedearlyifnecessary. Webuildaninteriorpointoptimizationframeworkaroundablackboxinteriorpoint optimizer.Theoptimizerisresponsibleforgeneratingintermediate,feasiblesolutions, andourframeworkupdatesthestatefortheoptimizerasafunctionofthoseintermediate solutionstoallowforfastercomputation.Detailsofourlazyinteriorpointoptimization frameworkcanbeseeninAlg.2. We optimize our paths with respect to path length. The collision avoidance con- straintsareformulatedovereachedgeinthepath,i.e.thelinearinterpolationthrough configurationspacebetweenthetwoconfigurations.Wedefineourclearancefunction (as in equation (1)) as a function of the minimum distance between the point in the point cloud and the geometric primitive interpolated through the workspace as deter- minedbytherobot’sforwardkinematics.Thisformulationgeneratesalargenumberof constraints(#ofrobotgeometricprimitives×#ofedgesinthepath×#ofpointsinthe pointcloud).Becausetheoptimizationmustconsidereachconstraint,fewerconstraints resultsinquickeroptimizationtimes.Inthefollowingparagraphs,wediscussournovel 8 AlanKuntz,ChrisBowen,andRonAlterovitz Algorithm2:LazyAnytimeInteriorPointPathOptimization Input:initialpathΠ ,obstaclesetO,convergencethreshold∆ ,bestfoundpathso init Π farΠ∗ Output:bestpathfoundΠ∗ 1 Kenab←0/,Πˆ ←Πinit,Π←Πinit 2 whiletimeremainingdo 3 Πnext←takeInteriorPointOptimizationStep(Kenab,Π) 4 Ocollide←inCollisionPoints(Πnext,O) 5 ifOcollide(cid:54)=0/ then 6 Kenab←Kenab(cid:83)Ocollide 7 Π←Πinit 8 else 9 ifdiscontinuousConstraintDetected(Kenab,Πnext)then 10 {k,k+1}=discontinuousJacobianEdge(Kenab,Πnext) 11 Πinit←bisectEdge(Πinit,k,k+1) 12 Π←Πinit 13 else 14 ifconverged(Πnext,Πˆ,∆π)then 15 returnminCostPlan(Πnext,Πˆ,Π∗) 16 end 17 Πˆ ←minCostPlan(Πnext,Πˆ) 18 Π∗←minCostPlan(Πˆ,Π∗) 19 Π←Πnext 20 end 21 end 22 end 23 returnΠ∗ approachtoreducingthenumberofconstraintsthatareconsideredbytheoptimization byaddingconstraintsinalazyfashion. Improving Performance through Lazy Constraint Addition Instead of using the entiresetofobstacle-basedinequalityconstraintsK,whichcontainsconstraintsforeach pointinthepointcloud,weinsteaddefineasubsetK astheenabledconstraintset. enab Thisset,whichstartsoutempty,isaddedtoaspointsinthecloudbecomerelevantto theoptimization(seeFig2andAlg.2lines5-7). Westarttheoptimizationwithanemptyenabledconstraintset(Fig.2(a)).Wethen take an optimization step, modifying the motion plan (Fig. 2(b)). At each iteration of suchastep,wechecktherobot’spathforcollisionwiththewholepointcloud(anoper- ationwhichiscomputationallyinexpensivecomparedwithoptimizingwitheachpoint asaconstraint).Ifthepathisfoundtobeincollision,weidentifywhichpointsarein collision, and add those points to the enabled constraint set, and restart the optimiza- tion(Fig.2(c)).Inthenextiteration,theoptimizeravoidscollisionwiththepreviously addedpoints(Fig.2(d)).Theprocessthenrepeatsuntilacollision-freeconvergenceis achieved. In this way, only points which prove to be relevant over the course of the FastAnytimeMotionPlanninginPointClouds 9 S G S G (a)PRM*PathEmpyConstraintSet (b)InCollisionOptimizationIteration S G S G (c)PRM*PathwithAddedConstraints (d)OptimizedCollision-FreePath Fig.2.Anexampleofthewayourmethodaddsconstraintsfromthepointcloudinalazyfashion, usinga2Ddiskrobotinapointcloud(greypointsareun-enabledconstraintpoints)(a)Example unoptimizedpathreturnedbythePRM*andusedtoinitializetheoptimizer.Theoptimizerstarts outwithanemptyenabledconstraintset.(b)Duringoptimization,therobot’spathisfoundto beincollisionwiththepointcloud(orangepoints).(c)Thein-collisionpointsofthecloudare immediatelyaddedtotheenabledconstraintset(blackpoints)andoptimizationisrestarted.(d) Theoptimizerconvergestoacollision-freepathusingtheenabledconstraintset,whichissmaller thanthesetofallpoints. optimizationaretakenintoaccountbytheoptimizer.Thisimprovesthecomputational speedoftheoptimizationalgorithmdrastically,asinpracticerelativelyfewpointsturn outtoberelevanttocollisiondetectioninanotherwiselargepointcloud. Enabled Constraint Cutoff Function One of the most time consuming parts of the optimizationiscomputingtheconstraintJacobianfortheconstraintsinK ,whichhas enab arowforeachconstraintandacolumnforeachstatevariable,i.e.d×(n−1)columns. Evenifaconstraintisrelevanttotheoptimizationasawhole,andassuchincludedin K , sometimes the point represented by the constraint is far away from the specific enab robot geometric primitive or edge representing an entry in the Jacobian. To leverage this intuition and further speed up the evaluation of the constraints in K , we use enab a cutoff function (3) as our clearance function (see Fig. 3(a)) rather than the distance itself,allowingtheusertosetaradius,r>0,asa“look-ahead”distance.Let d=min dist(s,q,q ,o), (2) i i+1 2(cid:0)d(cid:1) (cid:0)d(cid:1)<0  r r clearance(d,r)= (cid:0)d(cid:1)4−2(cid:0)d(cid:1)3+2(cid:0)d(cid:1) 0≤(cid:0)d(cid:1)≤1 (3) r r r r 1 (cid:0)d(cid:1)>1, r 10 AlanKuntz,ChrisBowen,andRonAlterovitz wheremin dististheminimumdistancebetweenthepointoandthegeometricprim- itiveswhensissweptalongthespacecurvedefinedbythetransitionbetweenconfig- urations q and q , i.e. as the robot arm is moving from configuration q to q . i i+1 i i+1 Equation 3 sends the clearance function’s derivative to zero at r, allowing us to spar- sifytheconstraintJacobianandonlyconsiderpointsinK whicharewithindistance enab r from the robot’s geometry at a given time. This allows even greater speedup to the optimization,astheconstraintsetK ,whichisalreadygreatlyreducedinsizefrom enab the full point cloud, is still larger than needs to be considered in many entries of the constraintJacobian. 1 ecnarae 0 s1 qk qk+1 s1 qk qk+1 qk+2 lc d -1-0.5 0 0.5 1 1.5 d2 d1 d2 o 1 d o min_dist(q, q , s,o) = d, k k+1 1 2 r min_dist(qk, qk+1, s1,o) = d1= d2 min_dist(qk+1, qk+2, s1,o) = d1 (a)CutoffFunction (b) Non-Differentiable Mini- (c)Bisection mumDistance Fig.3. (a) The cutoff function used as the clearance function for our obstacle avoidance con- straints.Thefunctionflattensoutafteracertaindistancesoastoallowustoonlyconsiderpoints inK whicharewithinaspecifiedradiusr.(b-c)Anexamplewithamanipulator(shownin enab red).(b)Ifasphereboundingtherobotgeometryissweptbetweentwoconfigurationsandpro- ducesacurvesuchasthebluecurvehere,thenwhend andd areequal,theminimumdistance 1 2 betweenspheres sweptfromq toq andpointcloudpointoisnotnecessarilydifferentiable. 1 k k+1 (c)Whensuchacaseisdetected,anintermediateconfigurationisaddedintothepathwithits ownconstraint.Inthisway,bothd andd areusedindifferentconstraintsandarebothlocally 1 2 differentiable. Anotherchallengeofourconstraintformulationisthattheminimumclearancebe- tweenagivenrobotgeometricprimitivealonganedgeandapointinthepointcloudis notguaranteedtobedifferentiable.Thisisbecausetherobot’sarmistracinganonlinear, complexcurvethroughtheworkspace,eventhoughtheedgeislinearinconfiguration space. This means that in some places, the closest position on that curve to an obsta- cle point can be discontinuous as the edge is perturbed (see Fig. 3(b)), i.e., there are multiple closest positions on the curve which are equidistant to the point cloud point. Usingdistanceoverthenonlinearedgeallowsustoensureobstacleavoidance,butcan causenumericalproblemsduringtheoptimization.Toaddressthisissue,whenweare evaluatingthederivativeofaconstraintandfindittobediscontinuousoveranedge,we bisectthatedgeandstarttheoptimizationanew(seeFig.3(c),andAlg.2lines9-12).In thisway,afterenoughbisectionsthediscontinuityisavoided.Inpractice,weobserved relativelyfewbisections.

Description:
Interleaving Sampling and Interior Point Optimization. Alan Kuntz, Chris Bowen, and Ron Alterovitz. University of North Carolina at Chapel Hill. Chapel Hill, NC 27599-3175, USA. {adkuntz,cbbowen,ron}@cs.unc.edu. Abstract. Robotic manipulators operating in unstructured environments such as.
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.