人工智能核心550篇论文 机器人 运动规划1802.08705v5.pdfVIP

PDDLStream:IntegratingSymbolicPlannersandBlackboxSamplersvia

OptimisticAdaptivePlanning

´´

CaelanReedGarrett,TomasLozano-Perez,andLesliePackKaelbling

ComputerScienceandArtificialIntelligenceLaboratory

MassachusettsInstituteofTechnology

caelan,tlp,lpk@

0

2

0Abstract

2

rManyplanningapplicationsinvolvecomplexrelationships

adefinedonhigh-dimensional,continuousvariables.Forex-

ample,roboticmanipulationrequiresplanningwithkine-

Mmatic,collision,visibility,andmotionconstraintsinvolv-

3ingrobotconfigurations,objectposes,androbottrajecto-

2ries.Theseconstraintstypicallyrequirespecializedproce-

durestosamplesatisfyingvalues.WeextendPDDLtosup-

]portageneric,declarativespecificationfortheseprocedures

Ithattreatstheirimplementationasblackboxes.Weprovide

Adomain-independentalgorithmsthatreducePDDLStream

s.problemstoasequenceoffinitePDDLproblems.WealsoFigure1:Left:Domain1(with5blocks).Right:Areal-

introduceanalgorithmthatdynamicallybalancesexploring

cworldrobotplanningto“serveameal”onthebrowntray.

[newcandidateplansandexploitingexistingones.Thisen-

ablesthealgorithmtogreedilysea