An Integrated System for Multi-Rover Scientific Exploration
Tara Estlin, Alexander Gray, Tobias Mann, Gregg Rabideau, Rebecca Castafio, 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 techniques from machine learning with planning and
This paper describes an integrated system for co- scheduling to enable autonomous multi-rover behavior ordinating multiple rover behavior with the overall for analyzing science data, evaluating what new sci- goal of collecting planetary surface data. The Multi- ence observations to perform, and deciding what steps P~over Integrated Science Understanding System com- should be taken to perform them. These techniques bines concepts from machine learning with planning are also integrated with a simulation environment that and scheduling to perform autonomousscientific ex- can model different planetary terrains and science data ploration by cooperating rovers. Theintegrated system within a terrain. utilizes a novel machinelearning clustering component Science data analysis in MISUSis performed using to analyze science data and direct newscience activi- ties. A planning and scheduling system is employedto machine-learning clustering methods, which use image generate rover plans for achieving science goals and to and spectral mineralogical features to help classify dif- coordinate activities amongrovers. Wedescribe each ferent planetary rock types. These methods look for of these componentsand discuss someof the key inte- similarity classes of visible, rock image regions within gration issues that arose during developmentand in- individual spectral images and across multiple images. fluenced both system design and performance. Specifically, clustering is performedby a distributed al- gorithm where each rover alternates between indepen- dently performing learning computations using its local Introduction data and updating the system-wide model through com- Landmarkevents have recently taken place in the ar- munication among rovers. Output clusters are used to eas of space exploration and planetary rovers. The help evaluate scientific hypotheses and also to prioritize Mars Pathfinder mission was a major success, not only visible surfaces for further observation based on their demonstrating the feasibility of sending rovers to other "scientific interest." As the system builds a model of planets, but displaying the significance of such missions the rock type distribution, it continuously assembles a to the scientific community. Future missions are be- new set of observation goals for a team of rovers to col- ing planned to send additional robotic vehicles to Mars lect from different terrain locations. Thus, the clusterer as well as to the outer planets and an asteroid (JPL drives the science process by analyzing the current data. 1999). In order to increase science return and enable set and then deciding what new and interesting obser- certain types of science activities, future missions will vations should be made. require larger sets of rovers to gather the desired data. A planning and scheduling component is used to These rovers will need to behave in a coordinated fash- determine the necessary rover activities required to ion where each rover accomplishes a subset of the over- achieve science goals requested by the learning system all mission goals and shares any acquired information. (Rabideau, Estlin, & Chien 1999). Based on an input In addition, it is desirable to have highly autonomous set of goals and each rover’s initial conditions, the plan- rovers that require little communicationwith scientists ner generates a sequence of activities that satisfy the and engineers on Earth to perform their tasks. An au= goals while obeying each of the rover’s resource con- tonomous rover will be able to make decisions on its straints and operation rules. Plans are produced by own as to what exact science data should be returned using an "iterative repair" algorithm which classifies and how to go about the data gathering process. conflicts and resolves them individually by performing This paper presents the Multi-Rover Integrated Sci- one or more plan modifications. Planning is distributed ence Understanding System (MISUS) which provides among the individual rovers where each rover is re- a framework for autonomously generating and achiev- sponsible for planning for its own activities. A central ing planetary science goals. This system integrates planning system is responsible for dividing up the goals Copyright (~)1999, American Association for Artificial amongthe individual rovers in a fashion that minimizes Intelligence (www.aaai.org). All rights reserved. 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. When science measurements are requested from a terrain during execution, rock and mineral spectral models are used to generate sample spectra based on ~J I sc!~n~¢o~I the type of rock being observed. The remainder of this paper is organized in the fol- lowing manner. Webegin by characterizing the cooper- 111111 ating rovers application domain and describing our sci- ence scenario. Next, we present the MISUSintegrated system framework and describe each of its components. Env,ronment We then discuss design decisions and system require- ments that arose during integration and any general lessons learned. In the final sections, we discuss related work, planned future work, and present our conclusions.
Cooperating Rovers for Science
Utilizing multiple rovers on planetary science missions Figure 1: MISUSArchitecture Diagram has many advantages. First, multiple rovers can col- lect more data than a single rover. A team of rovers For the framework presented in this paper, we have can cover a larger area in a shorter time where sci- initially chosen the configuration of a team of three ence gathering tasks are allocated over the team. Sec- rovers where each rover has a planning and learning ond, multiple rovers can perform tasks that otherwise tool on-board. Each rover can thus plan for its assigned would not be possible using a single rover. For in- goals, collect the required data, and perform science stance, rovers landed at different locations can cover analysis on-board which will direct its future goals. In areas with impassable boundaries. Also, with several addition, a central planner and learner are assumed to rovers, one rover can afford to take more risk and be located on either a lander or one of the rovers, which thus attempt tasks that usually might be be avoided. are used to coordinate science data and goals. Third, more complicated cooperative tasks can be ac- We evaluate our framework by testing its ability to complished, such as taking a wide baseline stereo image build a model of the distribution of terrain rocks, clas- (which requires two cameras separated by a certain dis- sifted according to composition as measured by a bore- tance). Finally, multiple rovers can enhance mission sighted spectrometer. To perform testing for differ- success through increased system redundancy. If one ent planetary terrain models, different rock fields are rover fails, then its tasks could be quickly taken over generated by using distributions over rock types, sizes, by another rover. In all cases, the rovers should behave and locations. Science goals consist of requests to take in a coordinated fashion, dividing goals appropriately spectral measurements at certain locations or regions. among the team and sharing acquired information. These goals can be prioritized so that if necessary low Coordinating multiple distributed agents for a mis- priority goals can be preempted (e.g. due to resource sion to Mars or other planet introduces some interest- constraints such as low battery power). ing new challenges for the supporting technology. Issues Science goals are divided among the three rovers. arise concerning communication, control and individual Each rover is identical and is assumed to have a spec- on-board capabilities. Manyof these design decisions trometer on-board as well as other resources including a are related, and all of them have an impact on any on- drive motor, a solar panel that provides power for rover board technologies used for the mission. For example, activities, and a battery that provides backup power for an on-board science analysis system, the amount when solar power is not available. The battery can also of communication available will determine how much be recharged using the solar panel when possible. Col- science data can be easily shared. This factor will also lected science data is immediately transmitted to the affect a planning system by determining how mucheach lander where it is stored in memory. The lander can rover can coordinate with other rovers to perform tasks. only receive transmissions from one rover at a time. The control scheme will determine which rovers execute what science gathering tasks which affects the on-board Multi-Rover Science Architecture components. For instance, some rovers may be utilized only for science data gathering, while other may be used The overall MISUSarchitecture is shown in Figure 1. for planning and/or science analysis. Decisions on the The system is comprised of three major components: on-board capabilities of each rover can also determine ¯ Data Analysis: A distributed machine-learning sys- the independence of a rover. tem which performs unsupervised clustering to model the distribution of rock types observed by the rovers. This system is designed to direct rover sensing to con- tinually improve this model of the scientific content of the planetary scene. ¯ Planning: A distributed-planning system that pro- duces rover-operation plans to achieve input rover science goals. Planning is divided between a central planner, which efficiently divides up science goals be- tween rovers, and a distributed set of planners which each plan for operations upon an individual rover. ¯ Environment simulator: A multiple rover simula- tor that models different geological environments and rover-science operations within them. The simulator o"~o manages science data for each environment, tracks rover operations within the terrain, and reflects read- reqI~ ings by rover science instruments. MISUSoperates in a closed-loop fashion where the Figure 2: Examplespectra feature space data analysis system can be seen to take the role of the scientist driving the exploration process. Spectra Distributed Clustering At any given time, each data are received by individual rover clustering algo- rover has a different location on the planetary surface rithms, which attempt to locally model the distribution and is sensing different targets. So each rover has its of rocks according to broad classifications of rock com- owndistinct segment of the overall dataset, stored lo- positions. This information is then sent to a central cally in its data buffer. Over time, each rover collects clusterer which integrates all gathered data into an up- a new set of data points, or 14-dimensional spectrum dated global model and broadcasts the new model back readings, adding it to its existing store of data points. to the distributed clusterers. A prioritization algorithm Clustering is initiated after each rover has obtained new uses the clustering output to generate a newset of ob- observations. A sample cluster model (shown for 2 of servation goals that will further improve the accuracy 14 dimensions) is shown in Figure 2. of the model. These goals are passed to a central plan- Clustering based on the EM (Expectation-Max- ner which assigns individual rovers to goals in a fashion imization) algorithm, an iterative optimization proce- that will most efficiently serve the requests. Then each dure, normally requires several passes over the entire rover planner produces a set of actions for that rover data. Since rovers must share information through a which will achieve as manyof its assigned goals as pos- power-expensive communication channel. Rather than sible. These action sequences are sent to the simulator send its local dataset to one or more other rovers, the where they are executed and any gathered data is sent distributed clustering algorithm allows a rover to send back to the rover clusterers. This cycle continues until only a small set of parameters which summarizes its enough data is gathered to produce distinct clusters for local data. Each rover’s model parameters are com- any observed rock types. puted locally, then sent to a central clusterer which In the next few sections, we discuss each of the integrates them into an updated global model (which MISUSsystem components in more detail. is also a small set of parameters) and broadcasts that model to all rovers in the system. Each rover takes Data Analysis System this global model into account when making its local To perform science analysis, we use a machine-learning estimate. This process continues iteratively until con- system which performs unsupervised clustering to vergence. This scheme trades off some accuracy in the model the distribution of rock types in the observed global model in order to minimize communication. In terrain. A primary feature of the MISUS is that the sep- the limit of large datasets, this scheme approximates arate rovers cooperate to form a joint consensus for the the equivalent non-distributed clustering model (where observed distribution of rock types. Througha learning one processor may examine all the data at once) more process, the global distribution model keeps improving and more closely. as more data is observed over time. For this demonstra- The algorithm is very homogeneous,i.e. each proces- tion prototype, the model used for this distribution is sor performs the same computation with the exception a simple K-means-like unsupervised clustering model, of the central clusterer, which performs a few additional where each cluster represents a different rock type in computations to compute the global model and broad- the sensor space. In the present simulation, each sen- cast it to the other processors. It also tolerant to pro- sor reading is a spectral measurementreturning values cessor dropouts, i.e. a circumstance in which one or at 14 wavelengths; learning takes place in the full 14- more rovers contributes zero data to the clustering, for dimensional continuous space. any reason such as a rover malfunction. RN.K Goal Selection The clustering model in this initial prototype system may be viewed as the scientific end- F_, - + - - product of the exploration. The overall purpose of r n k the system is to increase the accuracy of the cluster- R Nr K ing model by obtaining sensor readings in regions that +TEE E Murk (l°g M,~ - 11 are likely to improve the model. An update of the clus- r n k tering model determines new planetary locations to be R N~ K explored by the rovers. These locations are sent as for- mal goals by the learner to the planner. r n k Recall that clusters are defined in a high-dimensional Figure 3: Clustering objective function spectra space in which unsupervised learning will iden- tify different rock types. Every datum also has an asso- ciated position in physical space, on the planetary sur- Model and Optimization Our clustering model is a face. Assumingthere is some (perhaps very noisy) cor- special case of the general mixture of Gaussians studied respondence between rock type and spatial location, the extensively in statistics (Redner & Walker 1984). The purpose of goal selection is to direct exploration toward K-meansclustering model also corresponds to a special certain rock types by specifying newspatial targets (co- form of Gaussian mixture where a hard class member- ordinates in 3-space at which to take sensor readings) ship restriction is made (Bishop 1995). Our model according to the observed rock type distribution. similar but lifts the restriction. A very simple heuristic for goal selection is used in Hathaway (Hathaway 1986) shows an equivalence be- the current system. A constant number G of new spa- tween probabilistic and statistical mechanics-style ob- tial targets will be specified for each cluster. For each jective functions for mixture distributions. This trans- cluster, two of the G spatial targets are chosen by first formation allows us to generalize the probabilistic ob- finding the two mutually most distant points (in phys- jective function to include a temperature variable, al- ical space) of that rock type, then selecting a point in lowing us to use deterministic annealing to perform space stochastically from within a neighborhood of each global optimization of the model parameters. of those 2 points. These goals are given high priority. The distributed version of the clustering model fol- The rest of the G targets are chosen from neighbor- lows a development similar to that in (Tsioutsias hoods of randomly selected rocks in the cluster, and Mjolsness 1994) for partitioned neural networks. The are given lower priority. The idea of this heuristic is entire dataset (across all rovers) X contains N vectors to bias the system toward exploration in extremal di- x_ = (xt,..., xD) indexed by n. Denoting each of the R rections, as well as to explore the rock distribution in rovers with an index r, each rover has a subset Xr of the a way which balances effort equally between rock types data, containing Nr sensor readings. The global, shared (thus avoiding, say, spending undue energy on a very clustering model consists of K centroids, #k" Each rover commonrock type at the expense of rare rock types). r stores its ownlocal estimates of the centroids,/~k’ based Planning System on its subset of the data. Mnk denotes the membership g To produce individual rover plans, we used a distributed of datum x__, in cluster k, where ~k=l M,a = 1. Cluster version of the ASPEN(Automated Scheduling and membership is determined by a softmax over the dis- Planning Environment) system (Fukanaga et al. 1997). tance of a datum to each cluster mean. Wewill write ASPENis a configurable, generic planning/scheduling Mrk since memberships are only computed and stored application framework that can be tailored to specific locally on each rover. domains to create conflict-free plans or schedules. Its The clustering algorithm adjusts the values of the components include: centroids in order to minimize the objective function shownin Figure 3, by alternating between a centroid re- ¯ An expressive modeling language to allow the user to laxation s~ep, where the cluster means are re-estimated naturally define the application domain based on the current membership weights, and a mem- ¯ A constraint management system for representing bership relaxation step, where the membershipsare re- and maintaining domain operability and resource estimated based on the means, a is the reward for constraints, as well as activity requirements committing to cluster k and each An corresponds to a ¯ A set of search strategies and repair heuristics Lagrange multiplier enforcing non-negativity of mem- berships. T is the temperature parameter. ¯ A temporal reasoning system for expressing and maintaining temporal constraints Thefirst term in this objective function can be identi- fied with minimizing the distance between centroids and ¯ A graphical interface for visualizing plans/schedules the data associated with them and keeping the R esti- ASPEN employs techniques from planning and mates of the centroids close to each other. The second scheduling to automatically generate the necessary serves to prevent negative memberships. The third en- rover activity sequence to achieve the input goals. This forces that membershipssumto unity across the classes. sequence is produced by utilizing an iterative repair to designated waypoints. Decisions must be made not only to satisfy the requested goals, but also to provide more optimal schedules. ASPENcan consider optimiza- tion goals during the repair process. As certain types of conflicts are resolved, heuristics are used to guide the search towards making decisions that will produce higher quality schedules. In other words, when several options are available for repairing a conflict, these op- tions are ordered based on predictions on how favorable the resulting schedule will be. For this application, we have implemented heuristics based on techniques from the Multi-Traveling Sales- men Problem (MTSP), which is similar to the Trav- eling Salesman Problem (TSP) (Johnson & McGeoch Figure 4: Example ASPENplan 1997). For MTSP,at least one memberof a sales team must visit each city such that total traveling time is algorithm (Zweben et al. 1994) which classifies con- minimized. Both the central and rover planners utilize flicts and attacks them each individually. Conflicts oc- the MTSPheuristics. These heuristics are used to se- cur when a plan constraint has been violated where lect what rover should be assigned a particular science this constraint could be temporal or involve a resource, goal and a temporal location for the science activity. In state or activity parameter. Conflicts are resolved by previously reported results, they were shown to make performing one or more schedule modifications such as a significant impact in reducing overall traversal dis- moving, adding, or deleting activities. tance and expected execution time (Rabideau, Estlin, A rover that is at the incorrect location for a sched- & Chien 1999). uled science activity is one type of conflict. Resolving this particular conflict involves adding a traverse com- Environment Simulator mandto send the rover to the designated site. Other The environment simulator is designed to provide a conflicts may include having more than one rover com- source of data for the science analysis system by sim- municating with the lander at a time or having too ulating the science gathering activities of the rover. manyactivities scheduled for one rover, which over sub- Giventhe current science scenario, this entails the gen- scribed its power resources. The iterative repair algo- eration of an environment and the simulation of rover rithm continues until no conflicts remain in the sched- data gathering activities within the environment. ule, or a timeout has expired. Figure 4 shows an exam- Generation of the environment requires producing a ple rover-plan displayed in the ASPENGUIinterface. field of rocks for the rovers to traverse. The rock field Distributed Planning To support missions with is generated as a plane with rocks of various sizes em- bedded at various depths. The simulator maintains in- multiple rovers, we developed a distributed version of formation about the mineral composition of each rock, ASPENwhere it is assumed each rover has an on- and the spectrum that would correspond to its mineral board planner. This allows rovers to plan for them- composition. The size and spatial distributions of the selves and/or for other rovers. If communication is rockfield were developed by examining distributions of slow, it may be useful to have rovers construct their rocks observed by the Viking Landers, Mars Lander and own plans (and to plan dynamically when necessary, which is discussed in future work). Also, by balancing Mars Pathfinder. The distribution of minerals that can the workload, distributed planning can be helpful when occur in rocks was developed in collaboration with plan- individual computing resources are limited. etary geologists at JPL, and the spectra associated with rocks are generated from the spectra of the component The approach to distributed planning utilized in minerals via a linear-mixing model. MISUSis to include a planner for each rover, in addi- The simulation of the rover activities was done at tion to a central planner. The central planner develops a coarse level. Such considerations as kinematics and an abstract plan for all rovers, while each rover planner obstacle avoidance were not modeled in this simula- develops a detailed, executable plan for its ownactivi- tion. Other considerations, such as power consumption ties. The central planner also acts as a router, taking a and memory management were modeled by the plan- global set of goals and dividing it up amongthe rovers. ner for plan generation but not simulated by the sim- For example, a science goal may request an image of ulator. The rovers were essentially modeled as roving a particular rock without concern for which rover ac- spectrometers by the simulator. Figure 5 shows sev- quires the image. The central planner could assign this eral rovers and their spectrometer reaches modeled in goal to the rover that is closest to the rock in order to a sample rockscape. The simulation of rover activi- minimizethe traversals of all rovers. ties was accomplished by executing the plan generated Plan Optimization One of the dominating charac- by the planner, consisting of a list of movement,rota- teristics of the multi-rover application is rover traversals tion, and instrument commands. The simulator would representations. Within the planner, constraints may deal with sets of goals, resources and/or states which are primarily scientific, primarily engineering, or which form part of the interface between these two layers. A separate design consideration was that the interac- tions between the modules of the integrated system be asynchronous. In other words, each module needed to signal the next module when appropriate, rather than designating one process a control process, which would then control the actions of the others. For instance, the planner would be not begin planning until receiving a new set of science requests from the science analysis module. To that end, we designed a synchronization architecture that would facilitate interprocess signaling and also communicationof data. Essentially, each mod- Figure 5: Overhead view of simulated rockscape. Wedges ule acted as both a server process and a client process. denotedifferent rovers’ spectrometers’fields of view. A process would wait in server mode until the client initiated contact, do its processing, and then initiate then, from the location and direction specified by the contact with the server process of the next system as a movement and rotation commands, determine whether client. After the process finished communicatingto the or not a rock was visible by the boresighted spectrom- next process, it would go back to server modeuntil it eter. If so, the simulator would perturb the spectra had new data to process. in an amount proportional to the distance of the rover One important decision is the design of the overall from the rock in order to simulate instrument noise, planner, execution, and learner feedback. How often and store the spectrum for later communication to the the system loop is run is one issue. Increasing the fre- relevant clusterer. After all of the activities in a plan quency of feedback improves the responsiveness of the were executed by the simulator (i.e. moves, turns, and overall system to changes in the inputs (e.g. changes data gathering activities), the data was communicated in the observed science data) but increases the compu- to each clusterer via synchronization agents. The sim- tation cost of running the constituent algorithms (e.g. ulator would then wait for the next plan. planner, learner). Additionally, due to the design of the algorithms, one may know how much change in in- Integration Issues formation is needed to likely change the results of the computation (e.g. for the learner how muchnew data is The integration of two AI problem solvers and a sim- ulated environment involved a number of decisions. In likely to change the collection goals, for the planner how this section, we review someof the interesting and chal- muchof a change in execution state or goals is likely to lenging issues that arose in performing this integration; require another plan). While not critically sensitive to we particularly focus on the areas of system and inter- the amount of new data it receives, the more data ob- face design and system performance. tained by the science analysis moduleon a given system One major integration issue is interfacing between cycle, the more its model of the rock type distribution the different components. For instance, the planner will improve, resulting in useful newexploration goals. If, say, the learner obtains little new information, the was required to produce plans in a format compatible targets it decides upon will not be muchmore useful with the action representation required by the simula- tor. Also, the learner required the ability to ingest any than those it produced on the last iteration. science data returned from the simulator. A more com- A second issue related to system feedback is the plicated interface arose between the learning and plan- length of the horizon (i.e. the allowed plan execution ning components. Issues such as shared representation time period) that is considered by each cycle. If this of goals and objectives had to be resolved. horizon is short, it imposes constraints on how long the When specifying a new science goal, the learning cycle must be run (e.g. if the horizon is two hours, the component usually requested additional measurements cycle must be run at least every two hours). If the hori- be taken from a particular rock. However, this gen- zon is long, the individual modules may take longer to eral request had to be grounded in the form of terrain run (e.g. a planner takes longer to plan for a longer coordinates in order to represent the goal in the plan- horizon). The number of goals that are requested per ner’s modeling language. In addition, the planning and iteration also (to some degree) drives the size of the learning components had to agree on a priority repre- planning horizon since only a certain number of goals sentation that was expressive enough to represent the can be solved in any set length of time. information required by the learner but that could also The frequency and horizon of each cycle is not con- be easily utilized during planning to removegoals if nec- strained by our architecture. However,for our scenario, essary due to resource constraints. Another important we chose to have the cycle invoked once per local day issue was interfacing between science and engineering and to include a horizon of one day. This time scale is reasonable because science activities are not possible software from the NASAJPL Rocky 7 rover. MISUSis during the night period (as the rover is mainly solar intended to provide the science layer for this architec- powered) but computation is possible during such peri- ture, which will allow for more realistic testing of the ods (using the battery). Thus possible execution time MISUSframework. In the rest of this section we de- not expended during planning. However, other choices scribe extensions planned for each MISUScomponent. for cycle frequency and horizon are possible, and may make sense for different mission parameters. Future Work for the Data Analysis System The learning component described here represents an Related Work initial model intended primarily to bring out system is- The idea of having a scientific discovery system direct sues. The most straightforward of the improvements future experiments is present in a numberof other sys- under consideration are concerned with strengthening tems. Work on learning by experimentation, such as the clustering model to include outlier handling, covari- IDS (Nordhausen & Langley 1993) and ADEPT(Raja- ance modeling, incremental updating, model size deter- money1990), varied certain quantitative and qualita- mination, robustness to failure and missing data, and tive values in the domain and then measured the effects multiple parent representation. of these changes. MISUSdiffers from these systems in Animportant area for future work is in goal selection. that it interacts with an environment simulator to per- Combining spectral with spatial distribution models form experimentation and it is specialized to particular may allow for better-informed targeting. A combined problems and scenarios in planetary science. MISUSis clustering objective function would be more sensitive to also integrated with a planning system which constructs novel data in some locations than others. The system the detailed activity sequence needed to perform each could attempt to select target data which maximizes experiment based on a domain model. improvement of the data model. This is similar to the Other work has used experimentation to learn from saliency measure for an attentive, relaxation-based neu- the environment but experiments have not been scien- ral network (Tsioutsias & Mjolsness 1996) To improve tifically driven. EXPO(Gil 1993) integrates planning the system’s applicability to planetary science, we are and learning methods to acquire new information by in- working to select targets which maximally aid discrimi- teracting with an external environment. However, while nation between two competing hypotheses of geological MISUSlearns classification models of new geological processes, in collaboration with JPL geologists (Davies features, EXPOtries to improve its planning-related et al. 1999). One approach is to use stochastic parame- domain knowledge. terized grammars(Mjolsness 1997) to create a more de- The distributed clustering presented in this paper tailed spatial-spectral model of rock distributions than bears some similarity to other distributed learning a mixture of Ganssians. methods, such as (Provost & Hennessy 1999), which are constrained by low-bandwidth communication between Future Work for the Planning System processors to share relatively concise data models. One improvement for the planning component is to en- There has also been a significant amount of work on able dynamic planning capabilities. To accomplish this cooperating robots. One related system is GRAMMPS we will utilize a dynamic version of ASPEN,which mon- (Bummit & Stentz 1988), which coordinates multiple itors plan execution and allows re-planning when nec- mobile robots visiting locations in cluttered, partially essary (Chien et al. 1999). To perform autonomous known environments. GRAMMPS also has a low-level rover-operations, an on-board planning system must be planner on each robot and uses a similar approach to able to respond in a timely fashion to a dynamic, un- distribute targets, however GRAMMPS does not look predictable environment. Rover plans may often need at multiple resources or exogenous events. Most other to be modified due to events such as traverses com- cooperative robotic systems utilize reactive planning pleting early and setbacks such as failure to reach an techniques (Mataric 1995; Parker 1999). These systems observation site. focus on behavioral approaches and do not explicitly We also intend to extend the planning model to be reason about assigning goals and planning courses of more robust to failure situations. For instance, if failure actions. Furthermore, none of these systems utilize a occurs, the planning system should recognize it (e.g. learning componentto drive the system goals. the rover has not responded for a certain amount of time), not send any new goals to that rover, and re- Future Work assign any current goals assigned to that rover. A number of extensions are planned for each component of MISUS.One major extension already under way is Future Work for the Simulator to interface with a multiple rover execution architec- Weare interested in improving the environment simula- ture (Estlin et al. 1999) being developed at JPL that tor by adding different data sources, and by improving includes a number of additional components including: the sophistication of the hypotheses investigated. a real-time multi-rover hardware simulator which mod- The visual texture of a rock’s surface can give clues els rover kinematics and sensor feedback and control to the composition and geological history of the rock, and is a source of information that should be used when semi-autonomousrovers: exploring simulations of martian attempting to sample from the distribution of rock com- hydrothermal deposits. In American Geophysical Union positions rather than the spatial distribution of rock lo- Spring Meeting. cations. Weintend to incorporate visual texture as a Estlin, T.; Hayati, S.; et al. 1999. Anintegrated architec- source of information for the rovers to help them choose ture for cooperating rovers. In Proceedingsof the 1999In- a sampling strategy. ternational Symposiumon Artficial Intelligence, Robotics Modeling the distribution of rock compositions is a and Automationfor Space. task that can yield useful information for geologists. Fukanaga, A.; Rabideau, G.; Chien, S.; and Yan, D. 1997. Consider the scenario where an impact excavates an Towardsan application frameworkfor automated planning ancient hydrothermal system, in which there was a sta- and scheduling. In Proceedingsof the 1997 International Symposiumon Art]ieial Intelligence, Robotics and Automa- ble supply of hot water beneath the surface of Mars tion forSpace. at some time in the past. It may be possible to deduce Gil, Y. 1993. Efficient domain-independentexperimenta- the existence of such a system from study of the impact tion. In Proceedingsof the Tenth International Conference ejecta scattered on the surface, and examination of the on Machine Learning, 128-134. crater interior walls and rim deposits; hydrothermal ac- Hathaway,R. J. 1986. Another interpretation of the EM tivity would have altered the mineral characteristics of algorithm for mixture distributions. Statistics and Proba- the excavated deposits. bility Letters 4:53-56. Johnson, D., and McGeoch,L. 1997. The traveling sales- Conclusions manproblem: A case study in local optimization. In Aarts, This paper outlines a framework for coordinating mul- E. H. L., and Lenstra, J. K., eds., Local Search in Com- tiple rover behavior in generating and achieving geo- binatorial Optimization. London: John Wiley and Sons. logical science goals. This system integrates techniques 215-310. from machine learning and planning and scheduling to JPL 1999. http://www.jpl.nasa.gov/missions/. autonomously analyze and request new science data and Mataric, M. 1995. Designing and understanding adaptive generate the action sequences to retrieve that data. We group behavior. Adaptive Behavior 4(1):51-80. discuss a number of integration issues including devel- Mjolsness, E. 1997. Symbolic neural networks derived oping shared goal and plan representations, coordinat- from stochastic grammardomain models. In Sun, R., and ing systems asynchronously, and adjusting interface pa- Alexandre, F., eds., Connectionist Symbolic Integration. rameters to best serve the overall system goal. Wehope Lawrence Erlbaum Associates. the techniques and issues presented in this paper will Nordhausen, B., and Langley, P. 1993. An integrated prove useful to other designers of integrated systems. framework for empirical discovery. Machine Learning 12:17-47. Acknowledgments Parker, L. E. 1999. Cooperative robotics for multi-target This work was performed by the Jet Propulsion Laboratory, observation. Intelligent Automation and Soft Computing California Institute of Technology,under contract with the 5(1):5-19. National Aeronautics and Space Administration. This work Provost, F. J., and Hennessy,D. N. 1999. Scaling up: Dis- was supported by the NASAAdvanced Concept Program, tributed machinelearning with cooperation. In Proceedings Code S, managed by Neville Marzwell, and the NASAAu- of the Thirteenth National Conferenceon Ariticial Intelli- tonomy Program, Code SM, managed by David Atkinson. gence. The authors acknowledgethe invaluable contributions of Rabidean, G.; Estlin, T.; and Chien, S. 1999. Working AshleyDaviesfor his help in defining relevant geologysce- together: Automaticgeneration of command sequences for narios for this workand for providing general knowledgeof multiple cooperating rovers. In Proceedings of the 1999 planetary geology. Wealso thank Padhraic Smythfor pro- IEEE Aerospace Conference. viding his environmentsimulation code, and Rich Maclin for Rajamoney,S. 1990. A computational approach to theory leading the developmentof the spectra generation process. revision. In Shrager~J., and Langley, P., eds., Computa- tional Models of Scientific Discovery and Theory Forma- References tion. San Mateo, CA: MorganKanfman. 225-254. Bishop, C. M. 1995. Neural Networks for Pattern Recog- Redner, R. A., and Walker, H. F. 1984. Mixture densities, nition. NewYork, NY:Oxford Univ. Press. maximum likelihood, and the EMalgorithm. SIAM Review Bummit, B., and Stentz, A. 1988. GRAMMPS: A general- 26:195-239. ized mission planner for multiple mobilerobots in unstruc- Tsioutsias, D., and Mjolsness, E. 1994. Optimization tured environments. In Proceedings of the IEEE Confer- dynamics for partitioned neural networks. International ence on Robots and Automation. Journal of Neural Systems 5(4). Chien, S.; Knight, R.; Stechert, A.; Sherwood,R.; and Tsiontsias, D., and Mjolsness,E. 1996. A multiscale atten- Rabideau, G. 1999. Integrated planning and execution for tional frameworkfor relaxation neural networks. Advances autonomousspacecraft. In Proceedings of the 1999 IEEE in Neural Information Processing Systems 8. AerospaceConference. Zweben, M.; Daun, B.; Davis, E.; and Deale, M. 1994. Davies, A.; Mjolsness, E.; Gray, A.; Mann,T.; Castano, Scheduling and rescheduling with iterative repair. In R.; Estlin, T.; and Saunders, R. 1999. Hypothesis- Zweben,M., and Fox, M., eds., Intelligent Scheduling. San driven active data analysis of geological phenomenausing Francisco, CA: MorganKaufmann. 241-256.