Efficient generation of accurate mobility maps using machine learning algorithms

U.S Army’s mission is to develop, integrate, and sustain the right technology solutions for all manned and unmanned ground vehicles, and mobility is a key requirement for all ground vehicles. Mobility focuses on ground vehicles’ capabilities that enable them to be deployable worldwide, operationally mobile in all environments, and protected from symmetrical and asymmetrical threats. In order for military ground vehicles to operate in any combat zone, the planners require a mobility map that gives the maximum predicted speeds on these off-road terrains. In the past, empirical and semi-empirical techniques (Ahlvin and Haley, 1992; Haley et al., 1979) were used to predict vehicle mobility on off-road terrains such as the NATO Reference Mobility Model (NRMM). Because of its empirical nature, the NRMM method cannot be extrapolated to new vehicle designs containing advanced technologies, nor can it be applied to lightweight robotic vehicles.The mobility map is a function of different parameters such as terrain topology and profile, soil type (mud, snow, sand, etc.), vegetation, obstacles, weather conditions, and vehicle type and characteristics.A physics-based method such as the discrete element method (DEM) (Dasch et al., 2016) was identified by the NATO Next Generation NRMM Team as a potential high fidelity method to model the soil. This method allows the capture of the soil deformation as well as its non-linear behavior. Hence it allows the simulation of the vehicle on any off-road terrain and have an accurate mobility map generated. The drawback of the DEM method is the required simulation time. It takes several weeks to generate the mobility map because of the large number of soil particles (millions) even while utilizing high performance computing.One approach to reduce the computational time is to use machine learning algorithms to predict the mobility map. Machine learning (Boutell et al., 2004; Burges, 1998; Barber et al., 1997) can lead to very accurate mobility predictions over a wide range of terrains. Machine learning is divided into two categories: the supervised and the unsupervised learning. Supervised learning requires the training data to be labeled into predetermined classes, while the unsupervised learning does not require the training data to be labeled. Machine learning can help generate mobility maps using trained models created from a minimum number of simulation runs. In this study different supervised machine learning algorithms such as the support vector machine (SVM), the nearest neighbor classifier (k-NN), decision trees, and boosting methods were used to create trained models labeled as 2 classes for the ‘go/no-go’ map, 5 classes for the 5-speed map, and 7 classes for the 7-speed map. The trained models were created from the physics-based simulation runs of a nominal wheeled vehicle traversing on a cohesive soil.