With the influx of artificial intelligence (AI) models aiding the development of
autonomous driving (AD), it has become increasingly important to analyze and
categorize aspects of their operation. In conjunction with the high predictive
power innate to AI solutions, due to the safety requirements inherent to
automotive systems and the demands for transparency imposed by legislature,
there is a natural demand for explainable and predictable models. In this work,
we explore the various strategies that reveal the inner workings of these models
at various component levels, focusing on those adapted at the modeling stage.
Specifically, we highlight and review the use of explainability in
state-of-the-art AI-based scenario understanding and motion prediction methods,
which represent an integral part of any AD system. We break the discussion down
across three key axes that are inherent to any AI solution: the data, the model
architecture, and the loss optimization. For each of the axes, we outline the
general methodologies for introducing explainability, and reference and review
some practical realizations for each methodology. We conclude the article by
identifying several strategies that we believe are yet to be fully explored,
such as physics-inspired machine learning methods, neural network pretraining,
graph neural networks designed using domain-specific priors, and end-to-end
trainable networks based on differentiable kinematic models.
