Quantifying the Effects of Environmental Conditions on Autonomy Algorithms for Unmanned Ground Vehicles

Phillip J. Durst and Justin Carrillo

Abstract. Autonomy for commercial applications is developing at a rapid pace; however, autonomous navigation of unmanned ground vehicles (UGVs) for military applications has been deployed to a limited extent. Delaying the use of autonomy for military applications is the environment in which military UGVs must operate. Military operations takeplace in unstructured environments under adverse environmental conditions. Military UGVs are infrequently tested harsh conditions; therefore, there exists a lack of understanding in how autonomy reacts tochallenging environmental conditions. Using high-fidelity modeling andsimulation (M&S), autonomy algorithms can be exercised quickly andinexpensively in realistic operational conditions. The presented research introduces the M&S tools available for simulating adverse environmental conditions. Simulated camera images generated using these M&S toolsare run through two typical autonomy algorithms, road lane detectionand object classification, to assess the impact environmental conditions have on autonomous operations. Furthermore, the presented researchproposes a methodology for quantifying these environmental effects.

Key words: Autonomy, Autonomous Ground Vehicles, Environment Effects, Perception

1 Introduction

Autonomy for commercial ground vehicle applications is enjoying a great deal of success in the field, such as the Google car or Tesla [1]. However, autonomy for military ground vehicles has lagged far behind industry. While several factors have led to this lag, one primary obstacle to military autonomy applications is the operational environment. While commercial applications have the luxury of operating in stable, benign environments, military applications take place in harsh and unpredictable conditions. Autonomy algorithms, which are by nature often shown to be susceptible to the operational environment [2], often fail inharsh conditions, such as rain, snow, and fog. Moreover, little research has been given over to understanding exactly how such environmental conditions impact autonomous operations. Rather, autonomous ground vehicles remain tested and fielded primarily in the case of predictable, known, on-road conditions.

The effects of adverse weather on autonomy are not understood or, to date, measured in a quantitative fashion. Human driver response to rain, dust, softsoil, etc., can be empirically measured and modeled. However, sensor / autonomy responses to these conditions require complex data acquisition that is difficult and expensive to recreate in the field. The goal of this paper is to demonstrate how high-fidelity simulations can be used in lieu of field testing and that the controlled and repeatable conditions achieved in simulation can enable the development of quantitative metrics for algorithm performance.

The paper is laid out as follows. Section 2 gives a brief introduction to the simulation software used in this study, Section 3 provides example evaluations of algorithm performance as a function of environment. Specifically, performanceis assessed for both a lane detection algorithm operating in ideal and heavy rain situations and an object classification computational neural network (CNN) algorithm at varying times of day. Lastly, Section 4 provides concluding thoughts and recommendations for future work.