Extending Operational Design Domain for Perception Systems Through Robust Learning

Abstract

Perception modules in Autonomous Driving Systems (ADS) provide excellent performance in simple, constrained environments but generally have precarious performance in real driving situations with various weather and lighting conditions. Therefore, robust perception performance in new and unanticipated domains is a crucial factor for the large-scale application of autonomous driving, especially when considering unexpected scenarios outside the predefined Operational Design Domain (ODD). This paper proposes a novel approach to extending the ODD for perception modules in ADS through robust learning. The model's robustness is characterized by the anchor data and corresponding perturbation model. The robust learning task is then formulated as a min-max optimization problem conjugated to the perturbation model and a semantically parameter-defined constrained exploration space. The proposed robustify procedures solve the optimization problem in terms of robustness-related batch loss and worst-case loss, which improves the model resilience in multiple domain shift experiments, including virtual-real and weather changes. This paper presents experimental results that demonstrate the efficacy of robust learning approaches in extending the ODD for perception modules and provides insights into future research directions in this field.