Robotics Thesis Proposal
- Remote Access - Zoom
- Virtual Presentation
- ALBERTO CANDELA GARZA
- Ph.D. Student
- Robotics Institute
- Carnegie Mellon University
Bayesian Models for Science-Driven Robotic Exploration
Planetary rovers have traversed many kilometers and made major scientific discoveries. However, they spend a considerable amount of time awaiting instructions from ground operators. The reason is that they are designed for automated science data collection, not for autonomous exploration. The exploration of more distant worlds with stronger communication constraints will require a new approach.
This work advocates Bayesian models as powerful tools for a new paradigm for robotic explorers. In this approach, the explorer’s description of where to go is not prescribed by a path, but instead by a model of what the explorer believes. This formulation has several benefits. It allows a robot to gain a deeper understanding of the evolving scientific goals guiding the missions. Moreover, it can empower scientists by providing explainable results, as well as more intuitive commands. In this research we develop a means for learning and exploiting structure in the environment. We show how these models enable robotic explorers to take adaptive actions based on instantaneous information. Ultimately, we demonstrate how science productivity is improved by formulating the exploration problem in terms of well-known principles of information theory and Bayesian experimental design.
We start by laying out the foundation for a simple, yet complete Bayesian framework for robotic exploration. We establish a spatial probabilistic structure in which scientists initially describe their abstract beliefs and hypotheses, and then this belief evolves as the robot makes raw measurements. We also efficiently compute and maximize scientific information gain. Afterwards, we show improvements to specific components of our framework. We present a deep generative model that enhances low-resolution remote data, but also allows scientists to quantify and interpret the learned statistical dependencies. Additionally, we introduce an active model that uses contextual information from remote data to efficiently extrapolate features from in situ measurements. We also derive a related strategy for improving science productivity. All of these models are evaluated in either simulations or field experiments using real spectroscopic data of Cuprite Hills, Nevada.
Finally, the proposed work for this thesis will advance the state of the art in science-driven robotic exploration in two important ways. The first research direction will develop and incorporate powerful algorithms for spectroscopic composition analysis. The second research direction will address the holistic integration of both science productivity and risk awareness during rover path planning. The proposed work will be evaluated in different experiments involving the automated mapping of coral reefs, a Mars surface exploration simulation study, and robotic surveys of lunar analog environments with the autonomous rover Zoë.
David Wettergreen (Chair)
David R. Thompson (Jet Propulsion Laboratory/California Institute of Technology)
Zoom Participation. See announcement.