High-Level Sensor Models for the Reinforcement Learning Driving Policy Training

Abstract

Performance limitations of automotive sensors and the resulting perception errors are one of the most critical limitations in the design of Advanced Driver Assistance Systems and Autonomous Driving Systems. Ability to efficiently recreate realistic error patterns in a traffic simulation setup not only helps to ensure that such systems operate correctly in presence of perception errors, but also fulfills a key role in the training of Machine-Learning-based algorithms often utilized in them. This paper proposes a set of efficient sensor models for detecting road users and static road features. Applicability of the models is presented on an example of Reinforcement-Learning-based driving policy training. Experimental results demonstrate a significant increase in the policy’s robustness to perception errors, alleviating issues caused by the differences between the virtual traffic environment used in the policy’s training and the realistic conditions.