A Statistical Approach for Planetary Surface Exploration Rovers in Uncertain Terrain

I n this article, a statistical method for mobility prediction that incorporates terrain uncertainty is presented. Mobile robotics has been performing a significant role in scientific lunar/planetary surface exploration missions [1]. In such missions, mobile robots are required to predict their mobility to avoid hazards such as immobilizing wheel slip on loose sand or collision with obstacles. This mobility prediction problem is thus important to the successful exploration on challenging terrain. Of particular interest is mobility prediction on sloped terrain, since travel on slopes can cause extreme longitudinal and lateral slips. There have been significant works dealing with mobility predictions and analyses in the military community [2], [3]. These works have primarily focused on empirical analysis of large (i.e., several ton gross vehicle weight) vehicles. Other works have been performed to predict the mobility of small mobile robots while considering interaction mechanics of a slipping wheel on deformable terrain. Jain et al. have developed the rover analysis, modeling, and simulation software (ROAMS) simulator, which can be used for deterministic mobility prediction and includes models of terrain/vehicle interactions [4]. A multibody system for deterministic simulation of rover tire–soil interaction has also been demonstrated [5]. A terramechanics-based dynamic model for exploration rovers that considers slip and traction forces of a rigid wheel on deformable terrain has been developed [6]. Theseworks have employedwell-known dynamic and terramechanics models to calculate vehicle motion and wheel forces. However, these models assume prior knowledge of wheel– terrain interaction physical parameters (i.e., soil cohesion, internal friction angle, and others). In practical situations, mobile robots often traverse environments composed of terrain with unknown properties. These parameters can be estimated by onboard robotic sensor systems [7]–[9]; however, these estimated parameters remain subjected to uncertainty. Some recent work has attempted to predict rover mobility on slopes via a learning-based approach [10]; however, this work does not explicitly consider uncertainty in terrain physical parameters. Based on these observations, it can be asserted that practical approaches to mobility prediction should explicitly consider uncertainty in terrain physical parameters. A conventional technique for estimating probability density function of a system’s output response from uncertain input distributions is the Monte Carlo method [11], [12]. This approach generally requires a large number of analytical or numerical simulation trials to obtain a probability distribution of an output metric(s) associated with ranges of uncertain input parameters. Monte Carlo methods are typically computationally expensive, with computational cost increasing as the simulation model complexity increases. Structured sampling techniques such as Latin hypercube sampling, importance sampling, and others can be used to improve computational efficiency; however, these gains may be modest for complex problems [13], [14]. Alternatively, extended Kalman filters (EKFs) or particle filers have been well used for the prediction of a robot’s position Digital Object Identifier 10.1109/MRA.2009.934823 © RAND X PTURES