Learning human like driving policies from real interactive driving scenes - Archive ouverte HAL
Pré-Publication, Document De Travail Année : 2022

Learning human like driving policies from real interactive driving scenes

Résumé

Traffic simulation has gained a lot of interest for massive safety evaluation of self-driving systems in a risk free setting but the reality gap remains a big challenge. Adversarial Imitation learning (AIL) already proved that it is possible to learn driving policies from real demonstrations and more specifically on Highways (NGSIM dataset). However traffic interactions remains very restricted on those scenarios and it is necessary to cope with various and multiple real traffic participants to get real insight of human driving style. INTERACTION dataset was specifically designed for those concerns with complex and rich interactions on a variety of scenes like intersections, roundabouts, ramp merging in different countries. In this paper we introduce our training pipeline that is built upon the Lanelet2 road map format for learning human like driving policies based on most recent implementations of Adversarial Imitation Learning (AIL) algorithms. We compare different AIL algorithms and Behavioural Cloning (BC) baseline on various driving scenes and investigate how realistic driving policies can become as well as their ability to generalise on new scenes. We show that driving policies not only follow expert trajectories but also get safer with less offroad driving and collisions than BC baseline. This work opens new possibilities for multi agent traffic learning based on AIL techniques with real and highly interactive traffic data.
Fichier principal
Vignette du fichier
ICINCO_2022_Camera_ready.pdf (497.97 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03427989 , version 1 (14-11-2021)
hal-03427989 , version 2 (20-06-2022)

Identifiants

  • HAL Id : hal-03427989 , version 2

Citer

Yann Koeberle, Stefano Sabatini, Dzmitry Tsishkou, Christophe Sabourin. Learning human like driving policies from real interactive driving scenes. 2022. ⟨hal-03427989v2⟩

Collections

LISSI UPEC
364 Consultations
219 Téléchargements

Partager

More