Goal-Directed Scientific Exploration Using Multiple Rovers Tara Estlin, Rebecca Castafio, Ashley Davies, Darren Mutz, Gregg Rabideau, Steve Chien and Eric Mjolsness Jet Propulsion Laboratory California Institute of Technology 4800 Oak Grove Drive Pasadena, CA 91109-8099 {firstname.lastname}@jpl.nasa.gov Abstract Tt~is paper describes an integrated system for co- ordinating multiple rover behavior with the overall goal of collecting planetary surface data. The Multi- Rover Integrated Science Understanding System com- bines concepts from machine learning with planning and scheduling to perform autonomous scientific ex- ploration by cooperating rovers. The integrated system utilizes a novel machine-learningclustering component to analyze science data and direct newscience activ- ities. A distributed planning and scheduling system is employed to generate rover plans for achieving sci- ence goals, to coordinate activities among rovers, and to replan whennecessary. We describe each of these components and describe howthey are integrated with a planetary environment simulation. Introduction NASA has recently outlined a new Mars program which will have us visit the red planet over six times in the next two decades. At least four of these missions will involve rovers or other robotic craft that will be used to explore the surface of the planet and’ perform numer- ous geological, atmospheric, and other scientific exper- iments. In order to increase science return and enable certain types of science activities, future missions such as these will utilize large sets of rovers to gather the desired data. These rovers will need to behave in a coordinated fashion where each rover accomplishes a subset of the overall mission goals and shares any ac- quircd information. In addition, it is desirable to have highly autonomous rovers that require little communi- cation with scientists and engineers on Earth to per- form their tasks. An autonomous rover will be able to makedecisions on its own as to what exact science data should be returned and how to go about the data gathering process. This paper discusses the Multi-Rover Integrated Sci- ence Understanding System (MISUS) (Estlin et al. 1999) which provides a framework for autonomously generating and achieving planetary science goals. This system integrates techniques from machine learning with planning and scheduling to enable autonomous multi-rover behavior for analyzing science data, eval- Copyright © 2001, AAAI. All rights reserved. uating what new science observations to perform, and deciding what steps should be taken to perform them. These techniques are also integrated with a simulation environment that can model different planetary terrains and science data within a terrain. Science data analysis in MISUS is performed using machine-learning clustering methods, which use image and spectral mineralogical features to help classify dif- ferent planetary rock types. These methods look for similarity classes of visible, rock image regions within individual spectral images and across multiple images. Output clusters are used to help evaluate scientific hy- potheses and also to prioritize visible surfaces for fur- ther observation based on their "scientific interest." As the system builds a modelof the rock type distribution, it continuously assembles a newset of observation goals for a team of rovers to collect from different terrain loca- tions. Thus, the clusterer drives the science process by analyzing the current data set and then deciding what new and interesting observations should be made. A distributed planning and scheduling component is used to determine the rover activities required to achieve requested science goals. Based on an input set of goals and each rover’s initial conditions, the planner generates a sequence of activities that satisfy the goals while obeying each of the rover’s resource constraints and operation rules. Furthermore, as information is acquired regarding command-execution status and ac- tual resource utilization, the planner updates future- plan projections. Planning is distributed among the in- dividual rovers where each rover is responsible for plan- ning for its own activities. A central planning system is responsible for dividing up the goals among the in- dividual rovers in a fashion that minimizes the total traversing time of all rovers. The components described above are also integrated with a simulation environment that models multiple- rover science operations in a Mars-like terrain. Different Martian rockscapes are created for use in the simulator by using distributions over rock types, sizes and loca- tions. Whenscience measurements are requested from a terrain during execution, rock and mineral spectral models are used to generate sample spectra based on the type of rock being observed. AI IN AEROSPACE 33 From: FLAIRS-01 Proceedings. Copyright © 2001, AAAI (www.aaai.org). All rights reserved.