End-to-end Methods for Autonomous Driving in Simulation

Abstract

Fully autonomous driving is considered one of the grand challenges of modern technology and a variety of approaches have emerged for creating and evaluating autonomous driving agents. The self-driving industry typically adopts a modular software architecture and uses large fleets of autonomous vehicles for data collection and evaluation. Simulation generally occupies a minor role in these workflows as a tool for basic unit tests and debugging. However, recent end-to-end methods in academia have elevated simulation’s role in the training of machine-learned policies and demonstrate state-of-the-art performance on simulated driving benchmarks as a result.

In this thesis, we consider two questions. First, what are the state-of-the-art approaches for end-to-end driving in simulation and how can we improve them? We use the open-source CARLA simulator as our testbed and work with the recent Leaderboard benchmark for evaluation. In our first project, we give a detailed failure analysis of “Learning by Cheating” (LBC) which has been considered the state-of-the-art in imitation learning. We then propose “Offline Imitative Reinforcement Learning” (OIRL) as an alternative approach and investigate its performance. In our second project, we give a failure analysis of the “World on Rails” (WOR) offline reinforcement learning method and propose “Slightly Online Reinforcement Learning” (SORL) as an improvement.

Second, what value does simulation provide for the training and evaluation of driving algorithms? We draw attention to the crucial role that privileged information plays in both LBC and WOR’s algorithms. We emphasize that these methods are impossible to replicate without simulation and advocate for privileged information as an important tool for developing planning algorithms. Finally, we give a critical analysis of CARLA by pointing out issues in its perceptual and behavioral realism. We also critique the Leaderboard benchmark and provide discussion on why its primary metric cannot be reliably used as a proxy for good driving behavior.