Overview

This group focuses on developing sensing and control algorithms that enable new paradigms in data gathering, processing and understanding in order to advance NASA's science and exploration objectives.

In the area of sensing, we maintain that in understanding data generated by known physical phenomena, our greatest asset is the availability of physical laws that govern the behavior of the observed systems. Accordingly, our group specializes in building from first principles parametric models of physical systems, and subsequently using these models in a Bayesian framework to infer from sensor data the system's static and dynamic properties. This approach thus leads, for both natural and man-made systems, to a deep understanding of the structure and the evolution of the observed system.

In the area of distributed control, we focus on how to coordinate large collectives (i.e., a large collection of agents with a system level objective function to optimize). Accordingly, our group specializes in deriving general solutions to the agent objective assignment problem and devising adaptive methods (e.g., reinforcement learning, evolutionary algorithms) that allow agents to optimize those objective functions. These approaches have been successfully applied to many domains, including rover coordination, data routing, job scheduling over heterogeneous servers, congestion problems and NP-hard optimization problems (bin packing and faulty device selection).

Project List (Active)

Bayesian super-resolved 3D vision
Project Lead: Frank Kuehnel This project is engaged in the Bayesian inference of surface shape and reflectance models, from remote imagery, by modeling all relevant aspects of the image formation process. In one sense, this is an extension of visual stereo modeling, from discrete point inference and interpolation to continuous full image inference. In another, it is an exercise in principled data fusion, since the image formation models can account for wide disparities in camera position and surface illumination, and thus potentially for different instrument types. Super-resolution, of surface details smaller than the pixel footprints, is achieved when highly redundant imagery is available.