Statistical models of the environment can tell explorer robots where to travel, what samples to collect and what data to return to scientists. Much of my work aims to ground autonomous science decisions by remote explorers in formal principles of information theory, active learning and experimental design.

• D. R. Thompson, R. Castano and D. Wettergreen. Compression Ratio as Indicator of Scientist Preference For Rover Images. LPSC 2010 (PDF).
• D. R. Thompson, et al. Onboard Science Data Analysis: Implications for Future Missions. Planetary Science Decadal Survey, Community Whitepaper 2009 (PDF).
• D. R. Thompson. Domain-Informed Novelty Detection for Autonomous Exploration. IJCAI 2009 (PDF).
• D. Hayden, S. Chien, D. R. Thompson and R. Castano. Onboard Clustering of Aerial Data for Improved Science Return IJCAI Workshop on Artificial Intelligence in Space, 2009 (PDF).
• D. R. Thompson and N. Cabrol. Fast Texture Analysis for Autonomous Exploration. IJCAI Workshop on Artificial Intelligence in Space, 2009 (PDF).
• D. R. Thompson, T. Smith and D. Wettergreen. Information-Optimal Selective Data Return for Autonomous Rover Traverse Science and Survey. ICRA 2008 (PDF).
• D. R. Thompson and D. Wettergreen. A Tale of Two Rovers: Mission Scenarios for Kilometer-Scale Site Survey. ICRA Rover Workshop 2008 (Abstract) (Presentation).
• T. Smith, D. R. Thompson and D. Wettergreen. Generating Exponentially Smaller POMDP Models Using Conditionally Irrelevant Variable Abstraction. ICAPS 2007 (PDF).
• D. R. Thompson, T. Smith, and D. Wettergreen. Data Mining during Rover Traverse: From Images to Geologic Signatures. ISAIRAS 2005 (PDF).

I've been working on a team at JPL whose software will provide automated target detection and response for the Mars Exploration Rovers. Traditionally targets for remote sening (by cameras or spectrometers) are selected manually by operators in advance. This is less feasible for long-distance autonomous drives, where robots may visit terrain that is never seen by scientists. AEGIS allows the rover's flight software to automatically select features of interest and plan targeted observations in an opportunistic fashion. (Press Release). Image courtesy JPL/NASA/Caltech.

• T. Estlin, R. Castano, B. Bornstein, D. Gaines, R. C. Anderson, C. de Granville, D. Thompson, M. Burl, M. Judd, and S. Chien, Automated Targeting for the MER Rovers , Proceedings of the 2009 Infotech, Aerospace AIAA Conference, Seattle, WA, April 2009.
• T. Estlin, R. Castano, R. C. Anderson, D. Gaines, B. Bornstein, C. de Granville, D. Thompson, M. Burl and M. Judd, Automated Targeting for the MER Rovers, Proceedings of the Space Mission Challenges for Information Technology Conference (SMC-IT 2009), Pasadena, CA, July 2009.
• T. Estlin, R. Castano, D. Gaines, R. C. Anderson, D. Gaines, B. Bornstein, C. de Granville, B. Yang, D. Thompson, and M. Judd, “A New Capability for Automated Targeting and Sampling using Remote Sensing Instruments on the MER Rovers,” Proceedings of AGU Fall Meeting, San Francisco December, 2008

New radio astronomy instruments like the Square Kilometer Array (SKA) will generate petabyte data volumes. These investigations would benefit from real time anomaly detection and data mining to identify key features of interest. We're currently investigating adaptive resource allocation strategies and cost-sensitive computing to identify transient radio sources in radio astronomy data. (Image: Supernova remnant W28, courtesy of NRAO/AUI/NSF and Brogan et al.)

• J-P Macquart and the CRAFT collaboration team. "Commensal Real-time ASKAP Fast Transients (CRAFT) Survey," PASA 2010.

I'm working on a JPL project to perform data mining on hyperspectral imagery to automatically draft maps and characterize science phenomena. In older work I've used spatial point processes to generate automatic maps of varying rock distributions in imagery from the Mars HiRise orbiting camera. The statistical models can identify patterns in features' clustering and spread that may not be visible to the naked eye. Image courtesy NASA/JPL.

• L. Mandrake, D. R. Thompson, M. S. Gilmore and R. Castano, Hii-HAT: A Tool for Rapid Hyperspectral Inquiry. LPSC, Mar. 2010 (Poster).
• D. R. Thompson, L. Mandrake, M. S. Gilmore and R. Castano, Superpixel Segmentation for Endmember Detecion in Hyperspectral Images. AGU Fall Meeting, Dec. 2009 (Poster).
• D. R. Thompson, R. Castano and M. S. Gilmore, Sparse Superpixel Unmixing for Exploratory Analysis of CRISM Hyperspectral Images. IEEE Workshop on Hyperspectral Imagery and Signal Processing: Evolution in Remote Sensing, Aug. 2009 (PDF) (Presentation).

I'm working on novel path planning strategies for tasking "underwater gliders:" submersibles designed for ocean sampling missions that last for weeks or months. Part of the challenge is designing an optimal mission that accounts for strong, dynamic ocean currents and tides. Another hurdle is developing the onboard autonomy to keep the glider safe between communications with shore. We're building a path/activity scheduling system as a part of the NSF Ocean Observatory Initiative's Atlantic Sensorweb. Glider image courtesy Rutgers University.

• D. R. Thompson, S. Chien, Y. Chao, P. Li, B. Cahill, J. Levin, O. Schofield, A. Balasuriya, S. Petillo, M. Arrott, M. Meisinger. "Path Planning in Strong, Dynamic, Uncertain Currents," ICRA 2010 (PDF). (to appear).
• D. R. Thompson, S. Chien, M. Arrott, A. Balasuriya, Y. Chao, P. Li, M. Meisinger, S. Petillo, O. Schofield. "Mission Planning in a Dynamic Ocean Sensorweb," ICAPS SPARK 2009. (PDF).
• D. R. Thompson, S. Chien, M. Arrott, A. Balasuriya, Y. Chao, P. Li, M. Meisinger, S. Petillo, O. Schofield. "A Mission Planning System for Underwater Gliders," ICAPS Applications Showcase 2009.

Teaching a compter to find rocks in images is a challenging computer vision problem. However, the ability to produce fast statistical summaries of rock scenes would be a big benefit for geologic analysis. It will also be important for future missions to Mars and outer planets in which an autonomous spacecraft can use rock detection to automatically recognize important science targets (Original Mars image NASA/JPL/Caltech). Here is some previous work on the rock detction problem.

• H. Dunlop, D. R. Thompson and D. Wettergreen. Multiscale Features for Detection and Segmentation of Rocks in Mars Imagery. CVPR 2007 (PDF).
• D. R. Thompson and B. Castano. A Performance Comparison of Rock Detection Algorithms. IEEE Aerospace, 2007 (PDF).
• D. R. Thompson and B. Castano. A Comparison of Rock Detection Algorithms. AGU 2007 (Poster).
• D. R. Thompson and D. Wettergreen. Spatial Point Process Models for the Distribution of Northern Plains Boulders LPSC 2007 (PDF).
• D. R. Thompson, S. Niekum, T. Smith and D. Wettergreen. Automatic Detection and Classification of Geologic Features of Interest. IEEE Aerospace 2005 (PDF).

Several of our more recent tests involved an expedition to the Amboy Crater lava flow in the Mojave Desert, California. This is a relatively young basaltic lava flow, having grown from several eruption events over the past 50,000 years. We focused on building autonomous geologic surficial maps of the older eastern portion of the flow. The image at right is based on a map from A. W. Hatheway's 1971 U. of Arizona thesis work.

• D. R. Thompson, D. Wettergreen and R. Castano. Information-Optimal Subsampling of Temporal Image Sequences for Remote Exploration. IJCAI Workshop on Artificial Intelligence in Space, 2009 (PDF).
• D. R. Thompson. Intelligent Mapping for Autonomous Robotic Survey. Doctoral Dissertation The Robotics Institute, Carnegie Mellon. 2008 (PDF).
• D. R. Thompson and D. Wettergreen. Intelligent Maps for Autonomous Kilometer-Scale Site Survey. ISAIRAS 2008 (PDF).
• D. R. Thompson. Predictive Exploration for Autonomous Science. Proceedings AAAI Doctoral Consortium 2007 (PDF) (Presentation).

We want robots that can track a lot of objects simultaneously, target the interesting ones for analysis, and recognize things they've seen before. This is tough when you're constantly moving around to see features from different perspectives. However, we're making progress with some visual tracking, scene geometry and shennanigans with a wide-baseline stereo rig.

• F. Calderon P., D. R. Thompson and D. Wettergreen. Autonomous Rover Reflectance Spectroscopy with Dozens of Targets. ISAIRAS 2008 (PDF).
• D. R. Thompson and D. Wettergreen. Multi-Object Detection with Multiple-View Expectation Maximization Clustering IROS, 2005 (PDF).

I worked with the Planetary Science Summer School and the Jet Propulsion Lab TEAMX spacecraft design team to develop a prototype Venus lander.

Our concept focused on a short-term mission to investigate atmosphere/surface interatctions and mineralogy. Venus is a fascinating and little-understood planet. The VEIL project's basic premises were that a) surface exploration will eventually be necessary to understand its evolution, and b) this is possible on a relatively strict budget.

• VEIL Team. VEIL: A New Frontiers Class Mission Concept for Exploring Venus. AGU 2007 (Abstract) (Poster).
• VEIL Team. VEIL: A New Frontiers Class Mission Design Concept. DPS 2007 (Poster).

I worked with a CMU project that ran some robot field tests in the driest place on Earth - the Atacama Desert in Northern Chile. There we simulated remote science operations as an analog to a search for life on another planet. During field operations autonomy software automatically identified lichens with the help of a fluorescence imager device.

• D. Wettergreen, M.D. Wagner, D. Jonak, V. Baskaran, M. Deans, S. Heys, D. Pane, T. Smith, J. Teza, D.R. Thompson, P. Tompkins, and C. Williams. Long-Distance Autonomous Survey and Mapping in the Robotic Investigation of Life in the Atacama Desert. ISAIRAS 2008 (PDF).
• T. Smith, D. R. Thompson, D. Wettergreen N. Cabrol, K. Warren-Rhodes and S. Weinstein. Life in the Atacama: Science Autonomy for Improved Data Quality. Journal of Geophysics Researchvol. 112, G04S03, Dec. 2007.
• D. R. Thompson, T. Smith and D. Wettergreen. Automatic Detection of Novel Biologic and Geologic Images in Atacama Desert Rover Traverse Imagery. LPSC 2006 (PDF).
• T. Smith, D. R. Thompson, S. Weinstein and D. Wettergreen. Automatic Chlorophyll Detection and Followup Applied to the Search for Life in the Atacama LPSC 2006 (PDF).