Ground Robot Energy Prediction and Reachability in Off-Road Environments Through Spatial Terrain Mapping

For robotic applications, energy is a key resource that can both enable and limit the tasks that a robot can perform in an environment. In off-road environments, ground robots may traverse numerous different terrains with significantly and spatially varying energy costs. The cost of a particular robot moving through such an environment is likely to be uncertain, making mission planning and decision-making challenging. In this dissertation, we develop methods that use information on terrain traversal energy costs, collected during robot operation, so that future energy costs for the robot can be more accurately and confidently predicted. The foundation of these methods is to build a spatial map of the energy costs in an environment, while characterizing the uncertainty in those costs, using a technique known as Gaussian process regression (GPR). This map can be used to improve performance in important robotic applications, including path and mission planning. First, we present a 2-dimensional energy mapping formulation, based on GPR, that properly considers the correlation in path energy costs for computing the uncertainty in the predicted energy cost of a path through the environment. With this formulation, we define a robot's chance constrained reachability as the set of locations that the robot can reach, under a user-defined confidence level, without depleting its energy budget. Simulation results show that as a robot collects more data on the environment, the reachable set becomes more accurately known, making it a useful tool for mission planning applications. Next, we extend the spatial mapping formulation to 3-dimensional environments by considering both data-driven and vehicle modeling strategies. Experimental testing is performed on ground robot platforms in an environment with varied terrains. The results show that the predictive accuracy of the spatial mapping methodology is significantly improved over baseline approaches. Finally, we explore information sharing between heterogeneous robot platforms. Two different robots are likely to have different spatial maps, however, useful information may still be shared between the robots. We present a framework, based multi-task Gaussian process regression (MTGP), for learning the scaling and correlation in costs between different robots, and provide simulation and experimental results demonstrating its effectiveness. Using the framework, robot heterogeneity can be leveraged to improve performance in planning applications.PHDMechanical EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttps://deepblue.lib.umich.edu/bitstream/2027.42/153361/1/maquann_1.pd