Simulation-based reinforcement learning for real-world autonomous driving

TL;DR

This paper demonstrates how reinforcement learning in simulation, combined with synthetic data, can effectively transfer a driving policy to a real-world vehicle, reducing costs and engineering effort.

Contribution

It introduces a simulation-based reinforcement learning approach for autonomous driving that leverages synthetic data and minimal real-world data for training and transfer.

Findings

Successful sim-to-real policy transfer confirmed in real-world tests.

Design choices in perception and control significantly affect real-world performance.

Synthetic data can effectively replace extensive real-world data in training.

Abstract

We use reinforcement learning in simulation to obtain a driving system controlling a full-size real-world vehicle. The driving policy takes RGB images from a single camera and their semantic segmentation as input. We use mostly synthetic data, with labelled real-world data appearing only in the training of the segmentation network. Using reinforcement learning in simulation and synthetic data is motivated by lowering costs and engineering effort. In real-world experiments we confirm that we achieved successful sim-to-real policy transfer. Based on the extensive evaluation, we analyze how design decisions about perception, control, and training impact the real-world performance.