Model Validity in Observers: When to Increase the Complexity of Your Model?

TL;DR

This paper evaluates the accuracy of common vehicle models in autonomous vehicle systems, identifying their valid operational domains and comparing model-based and learning-based observers for safety and performance.

Contribution

It provides a systematic analysis of vehicle model validity, establishes specific lateral acceleration domains for accurate modeling, and compares traditional and learned observers.

Findings

Serious accuracy issues occur beyond certain lateral acceleration thresholds.

Defined a lateral acceleration domain where models are valid.

Learning-based observers can outperform traditional model-based ones within valid domains.

Abstract

Model validity is key to the accurate and safe behavior of autonomous vehicles. Using invalid vehicle models in the different plan and control vehicle frameworks puts the stability of the vehicle, and thus its safety at stake. In this work, we analyze the validity of several popular vehicle models used in the literature with respect to a real vehicle and we prove that serious accuracy issues are encountered beyond a specific lateral acceleration point. We set a clear lateral acceleration domain in which the used models are an accurate representation of the behavior of the vehicle. We then target the necessity of using learned methods to model the vehicle's behavior. The effects of model validity on state observers are investigated. The performance of model-based observers is compared to learning-based ones. Overall, the presented work emphasizes the validity of vehicle models and presents clear operational domains in which models could be used safely.