Learning to Drive from Simulation without Real World Labels
This addresses the sim-to-real transfer challenge for autonomous driving, enabling training without costly real-world data, though it appears incremental as it builds on existing image-to-image translation methods.
The paper tackles the problem of transferring a vision-based lane following driving policy from simulation to real-world operation without real-world labels, achieving successful autonomous driving on rural and urban roads.
Simulation can be a powerful tool for understanding machine learning systems and designing methods to solve real-world problems. Training and evaluating methods purely in simulation is often "doomed to succeed" at the desired task in a simulated environment, but the resulting models are incapable of operation in the real world. Here we present and evaluate a method for transferring a vision-based lane following driving policy from simulation to operation on a rural road without any real-world labels. Our approach leverages recent advances in image-to-image translation to achieve domain transfer while jointly learning a single-camera control policy from simulation control labels. We assess the driving performance of this method using both open-loop regression metrics, and closed-loop performance operating an autonomous vehicle on rural and urban roads.