自动驾驶论文3466 Imitation Learning of Hierarchical Driving Model From Continuous Intention to Continuous Trajectory.pdfVIP

2021IEEEInternationalConferenceonRoboticsandAutomation(ICRA2021)

May31-June4,2021,Xian,China

ImitationLearningofHierarchicalDrivingModel:FromContinuous

IntentiontoContinuousTrajectory

YunkaiWang,DongkunZhang,JingkeWang,ZexiChen,YuehuaLi,YueWang,RongXiong

Abstract—Oneofthechallengestoreducethegapbetween

themachineandthehumanleveldrivingishowtoendowthe

systemwiththelearningcapacitytodealwiththecoupledcom-

plexityofenvironments,intentions,anddynamics.Inthispaper,

weproposeahierarchicaldrivingmodelwithexplicitmodelsof

continuousintentionandcontinuousdynamics,whichdecouples

thecomplexityintheobservation-to-actionreasoninginthe

humandrivingdata.Specifically,thecontinuousintentionmod-

uletakesperceptiontogenerateapotentialmapencodedwith

obstaclesandintentions.Then,thepotentialmapisregarded

asacondition,togetherwiththecurrentdynamics,togenerate

acontinuoustrajectoryasoutputbyacontinuousfunction

approximatornetwork,whosederivativescanbeusedfor

supervisionwithoutadditionalparameters.Finally,ourmethod

isvalidatedbybothdatasetsandstimulation,demonstratingFig.1:Differenttrajectoryplanningapproachesfor

thatourmethodhashigherpredictionaccuracyofdisplacementautonomousvehicles.(1)Classicalmodulardesign;(2.1)

andvelocityandgeneratessmoothertrajectories.OurmethodisEnd-to-endplanningmethod;(2.2)Simiend-to-end

alsodeployedontherealvehiclewithlooplatency,validatingitsplanningmethod;(3)Ourhierarchicalmod