HAIL: Modular Agent-Based PedestrianImitation Learning

Igor Vozniak

Abstract: In the area of autonomous driving there is a need to flexi-bly configure and simulate more complex individual pedestrian behaviorin critical traffic scenes which goes beyond predefined behavior simula-tion. This paper presents a novel human-oriented, agent-based pedestriansimulation framework, named HAIL, that addresses this challenge. HAILallows to simulate human pedestrian behavior through means of imita-tion learning by virtual agents. For this purpose, HAIL combines the3D traffic simulation environment OpenDS with an integrated imitationlearning environment and hybrid agents with AJAN. For predictive be-havior planning on the tactical and strategical level, AJAN is extendedwith Answer Set Programming. For pedestrian behavior imitation learn-ing on the operational level, HAIL utilizes the module InfoSalGAIL forgeneration of pedestrian paths learned from demonstration by its humancounterpart as expert. Among others, an application example has beendemonstrated that HAIL can be applied to solve a common challengein the Neural Network domain, namely the out-of-distribution (OOD),e.g. never shown scenarios would raise an uncertainty prediction level,by unison work of the two different behavior generation frameworks.