Wildi, MaelAlahi, AlexandreVisser, Arnoud2023-03-152023-03-152023-03-15202310.1007/978-3-031-22216-0_50https://infoscience.epfl.ch/handle/20.500.14299/196135In this work, we study neural network architectures that will reduce the number of infractions made by autonomous-driving agents. These agents control vehicles by providing future waypoints directly from a forward-facing camera. Building on top of the teacher-student approach of Cheating by Segmentation, we investigate the impact of Pyramid Pooling Module and Feature Pyramid Network with the aim to learn more representative features. We run our experiment with CARLA simulator and show that pyramid perception modules have a positive impact in reducing the number of traffic light infractions and collisions. Détailsconditional imitation learningfeature pyramid networktext::conference output::conference proceedings::conference paper