Core Concepts
We use reinforcement learning in simulation to obtain a driving system controlling a full-size real-world vehicle, leveraging mostly synthetic data and achieving successful sim-to-real policy transfer.
Abstract
The authors present a series of experiments to train an end-to-end driving policy using the CARLA simulator and deploy it on a full-size car in real-world scenarios.
Key highlights:
They use reinforcement learning in simulation, with mostly synthetic data, to train a driving policy that takes RGB images and semantic segmentation as input.
The real-world experiments confirm successful sim-to-real policy transfer, with the policy achieving a substantial level of autonomy in various driving scenarios.
The authors analyze how design decisions about perception, control, and training impact the real-world performance.
Promising directions include using more regularization, control via waypoints, and leveraging offline proxy metrics for evaluation.
The authors also discuss challenges such as the sim-to-real gap and the need for more robust training algorithms.