| Autonomous off-road navigation in unstructured, terrain can encounter mobility difficulties even when the surface that the vehicle is driving over is level and free from obstacles. Mobility challenges result from soil characteristics that prevent the vehicle from developing sufficient traction to overcome resistance. The coupled vehicle-terrain interaction is difficult to predict in advance and can lead to immobilization en route. Furthermore, off-road terrain exhibits non-uniform soil characteristics that impact mobility due to different exposure to the elements, moisture content, temperature, homogeneity, and other factors. This thesis seeks to develop methodologies by which an autonomous vehicle can characterize or classify the terrain using proprioceptive sensors.;Two methods for terrain characterization are developed. The first method estimates vehicle-terrain forces from proprioceptive sensors and uses the well known Bekker terrain model to map terrain parameters from 21 terrains whose parameters are reported in the literature to terrain forces. The method uses a Bayesian multiple-model estimation (MME) approach to compare estimated terrain forces to those predicted for each terrain to identify which terrain from the set of 'known' terrains best matches the terrain that the vehicle is driving over. In doing so, the methodology identifies a set of terrain parameters that, along with the Bekker model, characterize the terrain. Simulation is used to evaluate the method, identifying the 'most likely' terrain by recursive estimation of the probability of each terrain in the hypothesis set. The simulation model accounts for soil stochasticity by allowing physical soil parameters to vary spatially.;The second method uses a machine learning approach based on Hidden Markov Models (HMM) to identify the most probable terrain. Field data from proprioceptive sensors are used to estimate resistive torque and slip at each wheel of the vehicle. Time histories of resistive torque and slip are used as feature vectors to train a HMM for each terrain type. Field data from an 'unknown' terrain is classified using the trained HMMs. The HMM approach is developed using proprioceptive data from a four-wheel drive, all electric robot.;The model-based method aims to characterize a terrain by its parameters, while the machine learning approach aims to classify the terrain into a pre-determined terrain type. Simulation results show that terrain characterization using Bayesian MME works well assuming that the terrain being characterized is well represented among initial hypotheses. The HMM classification approach is able to distinguish between 3 terrain classes for the 22 data sets collected. |
