Autonomous Ground Vehicle Guidance and Mapping in Challenging Terrains by using On-Board Vibration Measurements
How can measurements of a vehicle's ground-induced vibrations be used to improve its performance and longevity? This is the central question explored in this thesis. It is known that high levels of vehicle vibrations can lead to heightened mechanical wear and rider discomfort. Guidance algorithms should therefore strive to minimize these vibrations. This thesis starts by proposing a novel metric that quantifies vibrations using only a vehicle's gyroscopic data. This metric, which is adaptive to the vehicle's speed and is invariant to the sensor's mounting location, inspires three subsequent navigational algorithms. The first algorithm is a heuristic optimizer based on finite-difference approximations that finds a solution to a novel formulation of the autonomous tramming problem. On each tramming iteration, the vehicle evaluates the cost of a driven path by using the vibration metric. The optimizer then iteratively searches for a curvature-constrained path with locally minimal cost. This optimization method is validated on a Clearpath Husky A200 mobile robot that finds a path with minimal vibrations through an evolving terrain. The second algorithm is a method for laterally localizing terrain anomalies along a vehicle's path by using gyroscopic data. Terrain anomalies are detected by adaptively thresholding the vibration metric and located by adding a lateral offset to the vehicle's position when the anomaly is detected. This offset corresponds to half of the vehicle's width towards the side affected by the anomaly. The affected side is determined by using features extracted from the continuous wavelet transform of, and the ratio between, the gyroscopic roll and pitch rate signals. The anomaly detector and localizer are demonstrated on a Husky robot that learns to circumnavigate anomalies after hitting them once. Finally, this thesis presents an algorithm that can estimate the current state of an evolving environment by using a sparse set of samples. This is accomplished by introducing spatiotemporal forgetting into the Gaussian Process Regression framework. The resulting regressor is demonstrated with a Husky that makes a vibration map of a piecewise evolving indoor terrain. Such a map could be used to plan paths of minimal vibration through the terrain.