Using Physics-Based M&S for Training and Testing Machine Learning Algorithms

Abstract

Machine learning algorithms have been used to successfully solve many complex and diverse problems especially in the domain of unmanned vehicle systems. However, machine learning algorithms require training data that contain extensive variations in specimens. These variations include variations of the sensor settings, terrain conditions, environmental conditions, and variants of the objects of interest themselves. Capturing training specimens that span these variants is time consuming, expensive, and in some cases impossible. Training data in a narrow range of variations lead to decreases in performance such as overfitting. Therefore, collecting training data is often the limiting factor in developing robust machine learning algorithms for applications such as object detection and classification. Another time-consuming task is labeling training specimens with metadata needed in some training approaches. In this paper, we demonstrate using a physics-based modeling and simulation (M&S) capability to generate simulated training data spanning variations in sensor settings, terrain conditions, and environmental conditions that include a versatile automated labeling process. The product of prior efforts, the Virtual Autonomous Navigation Environment (VANE), is a high-fidelity physics-based M&S tool for simulating sensors commonly used on unmanned ground vehicles (UGVs) and unmanned aerial vehicles (UAVs).