自动驾驶论文1976 Decision Making for Autonomous Driving Via Augmented Adversarial Inverse Reinforcement.pdfVIP

2021IEEEInternationalConferenceonRoboticsandAutomation(ICRA2021)

May31-June4,2021,Xian,China

DecisionMakingforAutonomousDrivingviaAugmentedAdversarial

InverseReinforcementLearning

PinWangDapengLiuJiayuChenHanhanLiChing-YaoChan

Abstract—Makingdecisionsincomplexdrivingenviron-[2],andAdversarialInverseReinforcementLearning(AIRL)

mentsisachallengingtaskforautonomousagents.Imitation[8].Amongthem,AIRLiscapableoflearningbotha

learningmethodshavegreatpotentialsforachievingsuchapolicyandarewardfunctionwithbetterperformance.In

goal.AdversarialInverseReinforcementLearning(AIRL)is

oneofthestate-of-artimitationlearningmethodsthatcanlearntheoriginalwork,AIRLisverifiedwithonlysimplestatic

bothabehavioralpolicyandarewardfunctionsimultaneously,environmentssuchasmazenavigationofa2Dpointmass

yetitisonlydemonstratedinsimpleandstaticenvironmentsandaquadrupedalantrunningtaskinOpenAIGyms.There

wherenointeractionsareintroduced.Inthispaper,weim-isnointeractionwithdynamicobjectsinthetasks.

proveandstabilizeAIRL’sperformancebyaugmentingitwith

semanticrewardsinthelearningframework.Additionally,weHowever,drivingenvironmentsofautonomousvehicles

adapttheaugmentedAIRLtoamorepracticalandchallengingaremuchmorecomplicatedwithdiversesituations.The

decision-makingtaskinahighlyinteractiveenvironmentin

autonomousdriving.Theproposedmethodiscomparedwithinteractivebehaviors