My current and primary focus is on finding and analyzing optimal policies for POMDPS. However, some of my early research was directed towards determining traversable regions in of natural, outdoor terrain.

Control and Analysis of Stochastic Systems

Evolution of the hyperbelief

How does uncertainty affect a robot when attempting to generate a control policy to achieve some objective? How sensitive is the obtained control policy to changes in the description of the system? These are the central questions we wish to address. For most practical robotic systems, the state of the system is observed only indirectly through various sensors Since the actual state of the robot is not fully observable, the partially observable information is all that is available to the robot to infer its state and to use to make decisions. Further complicating matters, the system may be subject to disturbances that not only perturb the evolution of the system but also the observations. In such cases determining policies to effectively and efficiently govern the behavior of the system relative to a stated objective becomes computationally burdensome and impractical for the exact case. Thus, much research has been devoted to determining approximately optimal solutions for these partially observed Markov decision processes (POMDPs). We have developed a technique that is versatile: the majority of the computational effort is performed offline and the cost function and/or the initial hyperbelief can be changed with a minimal additional computational cost for any given system. In the future, we will use this framework to explore the sensitivity of the optimal policies to variation in the parameters describing the system. Ideally, this will endow us with an understanding of interplay between process model, measurement model, and the objective being optimized and the uncertainty therein as well as provide us with a means to derive a compact representation for the set of optimal policies over the space of of system parameters.

Travserability Analysis for Outdoor Navigation

Traversability analysis of an outdoor terrain

Determining traversable regions in outdoor, richly vegetated environments is critical when attempting to navigate through the environment. Indoor environments are typically highly structured: flat floors with perpendicular walls and objects in the environment are easily recognized and are almost always solid obstacles that vary only a small amount in appearance.

Outdoor environments are nearly the opposite in this regard. They are often highly unstructured with unlevel ground and boundaries that are not clearly defined. Moreover, objects in the workspace are not only difficult to recognize but whether they are obstacles or not is dependent on the capabilities of the robot. When does a shrub become an inpassable obstacle that the robot is incapable of traversing over? One of the major reasons why objects are difficult to recognize as obstacles is the high level of interclass variance as well as overlap within different classes. For instance a yucca plant may be an obstacle while a sage brush plant is not.

We consider the problem of automatically determining whether regions in an outdoor environment can be traversed by a mobile robot. To deal with the large interclass variability as well as the potential overlap of descriptors within classes, we attempt to generate features that maximize the effectiveness of making the distinction between obstacles and non-obstacles. We propose a two-level classifier that uses data from a single color image to make this determination. At the low level, we have implemented three classifiers based on color histograms, directional filters and local binary patterns. The outputs of these low level classifiers are combined using a voting scheme that weights the results of each classifier using an estimate of its error probability.

Further unpublished results segmented the image into a set of grids and determined the traversability of each grid. Various methods for classifying each of these regions were explored including nearest neighbors, neural networks, SVM, and various combinations of these methods.

This research was funded and sponsored by Sandia National Laboratories