Articles

Science Autonomy for

Rover Subsurface Exploration
of the Atacama Desert

David Wettergreen, Greydon Foil,

Michael Furlong, David R. Thompson

n As planetary rovers expand their capabili- obotic explorers communicate only intermittently
ties, traveling longer distances, deploying
complex tools, and collecting voluminous sci-
entific data, the requirements for intelligent
guidance and control also grow. This, coupled
R with scientists because of limited opportunities for vis-
ibility by Earth-based antennas and the growing num-
ber of spacecraft needing attention. The data rate of deep
space communication is also very limited. Autonomy can
with limited bandwidth and latencies, moti-
vates on-board autonomy that ensures the significantly improve science productivity in intervals
quality of the science data return. Increasing between communication opportunities. In particular, sci-
quality of the data requires better sample ence autonomy employs on-board analysis to make deci-
selection, data validation, and data reduc- sions affecting the scientific measurements that will be col-
tion. Robotic studies in Mars-like desert ter- lected or transmitted.
rain have advanced autonomy for long-dis- We define science autonomy as using information about sci-
tance exploration and seeded technologies for ence objectives and interpretation of science instrument
planetary rover missions. In these field exper-
data to determine rover actions. Science autonomy encom-
iments the remote science team uses a novel
control strategy that intersperses preplanned
passes detection and intelligent selection of measurements
activities with autonomous decision making. and samples, automatic acquisition of measurements. This
The robot performs automatic data collection, includes automated approach and instrument/tool place-
interpretation, and response at multiple spa- ment as well as calibration and verification, meaning col-
tial scales. Specific capabilities include instru- lecting the intended measurement or sample. Intelligent col-
ment calibration, visual targeting of selected lection of scientific measurements can increase both the
features, an on-board database of collected quantity and quality of information gathered. Science
data, and a long-range path planner that
autonomy describes utilizing scientific information to guide
guides the robot using analysis of current sur-
rover actions, for example, to execute an intelligent survey
face and prior satellite data. Field experi-
ments in the Atacama Desert of Chile over the
past decade demonstrate these capabilities
and illustrate current challenges and future
directions.

Copyright © 2014, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602 WINTER 2014 47
Articles

Figure 1. Zoë in the Atacama Desert.

or mapping strategy that adapts as data is collected. Deep Impact comet flyby (Mastrodemos, Kubitschek,
Decisions about which samples to acquire and where and Synnott 2005); target detection and response by
and when to travel next can be based upon metrics of the Mars Exploration Rovers (Castaño et al. 2008;
information gain. Similar metrics can also be used to Estlin et al. 2012); and spectral detection, discovery,
prioritize science data for download. Intelligent com- and mapping by the EO-1 spacecraft (Chien et al.
pression strategies use knowledge or models of con- 2005; Davies et al. 2006; Doggett et al. 2006; Ip et al.
tent to interpret and summarize in a compact form. 2006; Thompson et al. 2013). At the same time, new
The ultimate goal of science autonomy is to embody smart instruments are beginning to incorporate
sufficient understanding, quantified by models and autonomous science data analysis directly (Wagstaff
metrics, so that rovers can independently choose et al. 2013) and provide information that can be used
actions that best support the scientific investigation to guide the rovers’ targeting and operation.
in which they are engaged. Rovers will take their These techniques and others will enable surface
goals and guidance from scientists, but when isolat- rovers to achieve multiday autonomous operations.
ed they should make scientifically rational decisions Currently multiday rover plans do not travel over the
and when in communication they should provide horizon of yesterday’s imagery, which limits the dai-
the most relevant information possible. ly science yield. However, rover navigation already
Science autonomy is especially valuable for surface permits safe over-the-horizon traverses, and in prin-
rover operations because missions have finite lifetime ciple a rover could autonomously survey large areas
and rarely revisit sites after the first encounter — the of terrain with its full suite of instruments. In one
rover must make good decisions and get it right the natural arrangement, operators would direct the
first time. Recent demonstrations on spacecraft show rover using waypoints determined from satellite
increasingly sophisticated science autonomy capabil- images, relying on rover autonomy for low-level haz-
ities. Milestones include target tracking during the ard avoidance and science target selection en route.

48 AI MAGAZINE
 Articles

Figure 2. Locales Visited During the 2013 LITA Field Season.

A robot could even divert its path slightly to pursue remotely through the desert while exploring its geol-
science targets of opportunity (Woods et al. 2009). ogy and biology.
Multiday plans could therefore make very efficient This article describes the science autonomy system
use of communications and personnel resources, developed and tested with Zoë. It performed auto-
enhancing long-distance survey missions. matic acquisition of visible/near infrared (Vis-NIR)
The Life in the Atacama project is a NASA-spon- reflectance spectroscopy throughout the 2013 field
sored effort to evaluate these techniques in the con- season. This involved a range of different auton-
text of desert subsurface biogeology (Cabrol et al.
omous decisions exercised at various spatiotemporal
2007). It uses Zoë (Wettergreen et al. 2008), a rover
scales. We begin by describing the rover platform and
capable of traveling more than 10 kilometers per day
instrument payload. We then discuss instrument self-
and autonomously drilling up to 0.7 meter depth
(figure 1). As a mobility platform it combines navi- calibration, science feature detection, and targeting
gational autonomy with a changing payload of on- capabilities. We describe larger-scale path planning
board science instruments. Previous investigations used to select informative paths between waypoints.
have used a fluorescence imager capable of detecting We also detail the operational protocols used to com-
very specific organic compounds and neutron detec- mand the rover and the results of its autonomous
tors to measure hydrogen abundance. The current data collection. These experiments provide a case
configuration incorporates a Raman spectrometer, a study of science autonomy deployed continuously
visible near infrared point spectrometer, and naviga- over long distances. We report on system perform-
tion and science cameras. During a series of experi- ance, lessons learned, and plans for future develop-
ments in the summer of 2013, scientists guided Zoë ment.

WINTER 2014 49
Articles

mixed alluvium
mixed alluvium mixed alluvium
mixed-toned undivided

mixed-toned undivided mixed alluvium

mixed
alluvium

mixed-toned undivided mixed-toned undivided

mixed-toned undivided
dark-toned rock
younger mixed alluvium

mixed-toned undivided
mixed
dark stratified material alluvium

mixed-toned undivided

mixed-toned undivided
mixed
alluvium

mixed alluvium
mixed alluvium mixed alluvium
mixed-toned undivided

mixed-toned undivided mixed alluvium

mixed alluvium
mixed-toned undivided mixed-toned undivided

mixed-toned undivided
dark-toned rock
younger mixed alluvium

mixed-toned undivided
mixed
dark stratified material alluvium

mixed-toned undivided

mixed-toned undivided
mixed
alluvium

N 0 2 4 6 8
E Miles
W 0 2 4 6 8 10
S Km

Figure 3. Examples of the Different Data Products.

Top: Landsat image used in traverse planning. Bottom: Geologic classification map derived from ASTER data.

50 AI MAGAZINE
 Articles

Figure 4. Hazard Avoidance.

Rover Platform, Instruments, While the three visible and six SWIR bands are not
sufficient to conclusively identify mineralogical
and Operations composition, they do help discriminate the princi-
In typical field work, rover operations follow a daily pal units of surface material and suggested represen-
cycle in which a remote science team reviews the pri- tative sites to visit.
or data, decides the next day’s navigation waypoints Figure 3 shows examples of the different data
and measurements, and then sends these commands products: a Landsat image with three visible bands
to the rover over a satellite link. This is similar to the reveals terrain morphology and desirable outcrops,
and a multiband ASTER image provides a rough clas-
sporadic communications of a planetary mission.
sification of mineralogical units.
The rover then executes its commands over the
The rover itself is capable of driving more than 10
course of the subsequent day. In the Atacama cam-
kilometers per day on challenging desert terrain
paign, typical command cycles for Zoë cover 5–10
(Wettergreen et al. 2008). On-board obstacle avoid-
kilometers per day. Figure 2 shows the entire traverse
ance uses three-dimensional geometry from stereo
path: red dots show locations for imaging and spec- imagery to identify hazards above the ground plane
tral data collection, while white paddles indicate sites and plan local drive arcs that go around them (figure
of particular interest where more in-depth study is 4). Figure 5 shows the robot and the components
performed. used by its science autonomy system. A pair of for-
Scientists determine the waypoints for the next ward-facing navigation cameras provide hazard
day using geologic and compositional maps pro- avoidance capability through a local path planner.
duced from orbital remote sensing data. Here the The vertical drill structure delivers subsurface soil to
ASTER instrument proves particularly useful: its a microscopic imager and a Raman spectrometer
images have a spatial resolution of 15 meters (visible) inside the rover. Analyzing the drill samples takes an
and 30 meters (SWIR), making them capable of hour or more, so these are deployed judiciously at
resolving details such as isolated rock outcrops. specific locations. However, we found that autono-

WINTER 2014 51
Articles

Panoramic Spectrometer
camera foreoptic

Pan/Tilt
Unit

Navigation
cameras
White
Reference

Figure 5. The Main Components of Zoë’s Vis-NIR Spectrometer System.

A Raman spectrometer inside the rover measures pulverized samples from the subsurface drill.

my could play a role in improving the science data analysts can refine the remote view with detailed
collected by the Vis-NIR spectrometer. The spectrom- spectral information and specific mineral absorption
eter is a modified Analytical Spectral Devices Field- features. In this manner the Vis-NIR data serves as
spec Pro that acquires radiance spectra from 0.4–2.5 both a validation of the orbital data and a means to
micrometers at 0.001 micrometer resolution, housed better interpret the mineralogical constraints and
in the rover body and connected by a fiber optic context for biogeology studies. Each session of Vis-
cable to a foreoptic telescope mounted on a pan-tilt NIR acquisitions begins with the rover calibrating its
mechanism. The foreoptic provides a 1 degree field instrument for temperature, solar geometry, and
of view, and can be directed at specific targets in the atmospheric conditions using a white reference tar-
environment. Its field of regard spans a full 360 get mounted on the rover deck (figure 5 inset).
degrees azimuth and 90 degrees elevation. A colocat- Dividing the radiance from the target by the refer-
ed camera provides visual context to interpret the ence measurement produces reflectance data of the
spectra. form shown in figure 6. These spectra were acquired
Zoë’s Vis-NIR reflectance data overlaps in wave- at locations indicated in the adjacent panoramic
length with ASTER orbital images; it is a more spa- camera subframe, from a distance of approximately 2
tially and spectrally refined version of the satellite meters. The reflectance values represent the fraction
data. By visiting distinctive terrain units of figure 3, of light reflected at each wavelength; more specific

52 AI MAGAZINE
 Articles

0.5
II II
0.45
I
0.4
III
0.35

Reflectance
0.3

0.25 IV
0.2 V
0.15

0.1

0.05

0
0.5 1 1.5 2 2.5

Figure 6. Panoramic Camera Subframe.

Dive spectrometer fields of view (PP24), and associated reflectance spectra.

formulations are possible (Schaepman-Strub et al. Smart Targeting

2006), but we will use reflectance here in the ordi- Zoë began each autonomous target selection process
nary Lambertian sense. This assumption should gen- by acquiring a navigation camera image. On-board
erally hold for the geologic materials of interest. Note image processing then analyzed the scene to find
that the light-colored sediment area in spectra I-III is large contiguous regions of a desired terrain class.
associated with a higher average reflectance, as well Typically these classes were rough surface features
as unique spectral features such as the dip near 2 like rock outcrop or bright sediment patches with
micrometers. These spectra were smoothed using distinctive spectral signatures. Upon finding a feasi-
local linear regression, but some lingering noise ble target, the rover recalibrated its Vis-NIR spec-
spikes at longer wavelengths evidence the lower sig- trometer, pointed at the feature, and collected a
nal level in these spectral regions. small 3 x 3 raster of spectra centered on the target of
interest. For context, it also acquired a high-resolu-
tion color image of the scene.
Science Autonomy Methods The image analysis used a random forest pixel clas-
sification system described in previous work (Foil et
Zoë’s science autonomy system includes two basic
al. 2013; Wagstaff et al. 2013) and adapted to the
capabilities that operate on mesoscale and
Atacama environment. This supervised classification
macroscale features respectively. Smart targeting can method learns a mapping from local pixel intensities
identify science features in rover navigation imagery to the surface class of that pixel. The model is instan-
and use this information to point the Vis-NIR spec- tiated as an ensemble of decision trees trained in
trometer. Adaptive path planning navigates on scales advance. At run time, the rover tested each pixel in
of tens or hundreds of meters, using satellite images the new image and averaged the classification of
to select waypoints with distinctive or novel spectra. each tree in the ensemble. The end result was a clas-
We describe each of these techniques in turn. sification map of the entire image, along with asso-

WINTER 2014 53
Articles

ciated class posterior probabilities. By subsampling spectral image using a linear combination of library
each image by a factor of four prior to classification, spectra given by a d x m matrix Y. This gives the rela-
processing time was less than a second on Zoë’s on- tionship x = Y ϕ.
board laptop-scale CPU. In practice there is always residual error separating
After image classification, connected components the reconstruction from the measurement. This is
analysis was used to identify contiguous targets. The partly attributable to measurement noise, but unless
rover then promoted the single largest target of the the library is comprehensive there may also be
desired class for followup data collection. For each incompleteness errors (for example, spectral features
target, a center pixel was determined using the that are expressed in the observations but not present
largest inscribed circle heuristic (Estlin et al. 2012) in the library). A library that reconstructs all spectra
and transformed to a pan-tilt angle using the well can be said to have explained the scene, and pro-
assumption of a planar terrain surface. Use of navi- vides insight into the mineral compositions in the
gation camera stereo data would identify a true three- remote sensing data. This intuition provides a figure
dimensional position and enable more sophisticated of merit for an adaptive path-planning system to
kinematic solutions. Here we relied on an approxi- select future measurement locations. Zoë’s planner
mate planar solution coupled with rastering to selects locations, the measurements at which, when
ensure that an inaccurate pointing would still cap- used to augment the collected library Y, provide the
ture the target in at least one of the Vis-NIR spectra. largest expected reduction in unmixing error. The
Scientists developed several different ways to planner aims to visit locations that are spectrally dis-
incorporate this capability into multiday rover oper- tinctive, collecting samples that fully explain the
ations. The first approach was a target check used in orbital image.
the middle of long traverses. This only deployed the
In detail, we begin with a space of candidate meas-
spectrometer if a feature of interest was found at the
urement locations L. The robot collects a library of
check point. If there was no feature in the rover’s
spectra by sampling at sites B = {b : b ∈ L }. We define
field of view, it would carry on without spending the
a stochastic measurement function, y = f(b) + ε with
time to calibrate and deploy its spectrometer. In this
Gaussian-distributed measurement noise ε, that
fashion, Zoë could cover long distances without
yields spectral measurements Y = {yi : yi ∈ Rd, 1 ≤ i ≤
spending undue time on bare or uninteresting ter-
m}. Together the observations form a spectral library,
rain. This strategy was also useful near the boundary
a random m x d matrix written YB. Good measure-
of geologic contacts where the precise location was
ments reduce the total reconstruction error for select-
uncertain. A second strategy involved a paired
ed remote sensing observations given by X = {xi : xi ∈
panorama that acted as a supplement to a com-
Rd, 1 ≤ i ≤ n}. We impose a resource cost C(B) to rep-
manded Vis-NIR spectral raster. Here the rover com-
mitted all time resources in advance. It calibrated its resent limited time, power, and bandwidth; it must
spectrometer and acquired data in columns of five not exceed a total budget β. For simplicity we will ini-
spectra spaced at 10 degree increments directly in tially ignore the cost of point-to-point travel.
front of the robot and to either side. This provided a We define a risk function as the expected recon-
representative sampling of the terrain comprising the struction error incurred from unmixing the remote
rover’s current ASTER pixel. It then augmented this images with the library of spectra collected at the
dataset with a 3 x 3 raster centered on any target of surface:
interest. Together, these two products gave a better  
insight than either taken individually. They met the R(B) = E min ∑ YBφ − x 2 
 φ
x∈X  (1)
dual needs of having representative spectra as well as
for φ ≥ 0, C(B) ≤ β
capturing distinctive (outlier) features.
Here we are taking the expectation over the rover’s
Adaptive Path Planning observation matrix, which is a random variable.
The science autonomy system also operates on larg- Computing this expectation is analytically challeng-
er scales of tens or hundreds of meters, where it ana- ing, so instead we solve the related problem:
lyzes satellite data to adjust its traverse path. We
arg min B = min ∑ XBφ − x
model the explored environment using a standard φ
x∈X
2

geographic or area mixing model where each meas- (2)

for φ ≥ 0,C ( B) ≤ β
urement is a mixture of a small number of end-mem-
ber materials. End members’ spectra combine in pro- The linear geographic mixing assumption allows us
portion to their physical extent on the surface. Most to infer the expectation E[YB] since in situ spectra
scenes contain just a few end-member spectra, and combine in proportion to their extent on the surface.
any measurement x can be reconstructed with appro- Due to geographic mixing we can directly substitute
priate constituents and mixing fractions. For a scene the remotely observed spectra as the expectated
with m end members we define the mixing fractions observations. We rewrite the objective function using
to be vectors ϕ ∈ Rm. More generally we can model a remote measurements at sites B, written XB:

54 AI MAGAZINE
 Articles

Future spectra
YB

Orbital
data
XA

Collected
spectra YB

Figure 7. Formulation of Adaptive Path Planning.

R ( B A) = min ∑ XB φ − x at the future candidate locations. The objective is:

2
φ
x∈X
(3) R ( B A) = min ∑ [ Z A XB ] φ − x 2
φ
for φ ≥ 0,C ( B) ≤ β x∈X
(4)
for φ ≥ 0,C ( B) + C ( A) ≤ β
This allows direct computation of the objective for
any candidate set of measurement locations. To summarize, this final form reflects the key ele-
As the robot begins to collect spectra, some ele- ments of Vis-NIR surface exploration: the overall
ments of E[Y] become observed. The matrix ZA rep- goal of an accurate model using a handful of spectra,
resents the library of spectra collected at previous reflected in the squared error term; the physical
locations A = {a : a ∈ L.}. These measurements are a behavior of geographic mixing, which appears as a
realization of YA, and can be substituted into the positivity constraint; and the overall path-length
expectation as the library of in situ spectra grows. budget β representing finite rover energy and time
Consequently the library used for unmixing consists resources.
of (1) the actual spectra collected at previous locations Figure 7 portrays the planning process. Here the
concatenated with (2) the expected spectra collected robot has collected two spectra to form its library ZA.

WINTER 2014 55
Articles

It calculates the expected risk of a candidate path best of all. This performance boost happens because
using remote sensing data at locations XB as a proxy only the adaptive systems actively pursue the isolat-
for future measurements. In this manner, it can ed outcrops with unique mineralogy.
greedily (or nonmyopically) construct an optimal On-board the real rover, a low-level control system
path. For our tests, the path planning was purely is required to travel safely between these features of
greedy; we added waypoints one by one, inserting interest. Consequently, Zoë has on-board navigation
them into the optimal location in the waypoint software to turn high-level science waypoints, spaced
sequence and stopping when the total Euclidean on the order of tens or hundreds of meters, into low-
path cost was exceeded. level vehicle actions like drive arcs. It uses a software
During the Atacama field season we defined a suite known as the Reliable Autonomous Surface
budget defined in terms of path length, typically pro- Mobility (RASM) package (Wettergreen and Wagner
viding 1.5 times the straight-line distance to the goal. 2012) capable of local hazard avoidance and path
The time cost could be significant for longer travers- planning using a three-dimensional terrain represen-
es, particularly including the cost of the spectral tation. RASM extracts a cloud of three-dimensional
measurements at intermediate waypoint. For this rea- points from the stereo cameras, orients these points
son, most navigation actions were driving com- relative to previously collected data, and builds a tri-
mands with adaptive navigation actions at particular angulated mesh. An A* search algorithm projects
regions of special interest. When it encounters a sci- drive arcs across this mesh to compute the cost of
ence waypoint in the plan, the science path-planning local control actions. On longer scales, graph search
software finds the complete interpolating path that identifies the best path to the next waypoint. RASM
minimizes spectral reconstruction error of the nine- retains knowledge of the topology relating observa-
band ASTER image. It drives to the next intermediate tion locations to their neighbors, permitting efficient
waypoint along that path, collects a spectrum of the loop closure and pose estimation over long distances.
terrain, and replans the remainder using whatever
path budget remains. That remainder becomes the
next science plan, which is further refined in addi-
Field Season Results
tional planning rounds as the rover progresses for- We engaged smart targeting during three days of rover
ward. In this fashion, the science planner can be ful- operations. Table 1 shows the performance for typical
ly stateless and react to new data encountered during targets during these traverses. Columns indicate the
the traverse. day (Sol) of operations; the action sequence number,
Figure 8 shows the benefit of science-aware path with TC indicating a target check, PP a paired panora-
planning in a simple simulation. We simulate a vir- ma, and AT a more specific planned data collection
tual rover traversing the famous Cuprite, Nevada, activity; the analysts’ post hoc interpretation of the
mining district, which is known for containing many feature that was found; and two columns indicating
distinctive spectral features of interest in highly whether the result was a reasonable science target and
localized outcrops (Swayze et al. 1992). Here we whether the pointing was accurate. Pointing accuracy
planned rover paths using orbital ASTER data, simu- was evaluated based on the context image collected
lating 256 trials with random start and end points with each spectrum, allowing it to be placed within
within the scene. We also simulated high-resolution the navigation camera image.
in situ acquisitions using coregistered data from the Overall, target selection performed reliably. The
Airborne Visible Near Infrared Spectrometer (AVIRIS) majority of targets were either rocks or patches of dis-
(Green et al. 1998). tinctive sediment. The only arguable failure was
The comparison considers four different strategies when the system identified a distant car that fell into
to fill the path-length budget: a random path, which the rover field of view. The pointing solution was
bends the path toward randomly selected intermedi- slightly less reliable, since our groundplane assump-
ate waypoints; a direct path, which begins with a tion tended to break down at the periphery of the
straight line and then adds ”wiggles” until the total navigation camera image near the horizon.
length is reached; an unconstrained adaptive Occasionally very distant targets would result in
approach that minimizes equation 3 but without the the rover aiming its spectrometer too far into the dis-
positivity constraint; and an adaptive approach that tance. Only one scene was totally featureless — an
enforces positivity of mixing fractions. We recom- empty plain — and in this case the rover detected no
puted the reconstruction error for every trial by targets and successfully abstained from spending any
applying nonnegative least squares to the collected time resources.
high-resolution spectra. Figure 8 shows each Figure 9 shows several images from the real-time
method’s performance with a box indicating the classification. Here the system was trained to recog-
median and quartile of the data and the whiskers nize rocks and high albedo soil patches, and it suc-
indicating the extrema. Both adaptive methods sig- cessfully finds these features. In the center column, a
nificantly outperform the nonadaptive approaches, red overlay represents the surface labeled as the
with the constrained adaptive method performing belonging to the target class. The black rectangles

56 AI MAGAZINE
 Articles

Mean unmixing reconstruction error

0.04

0.035

0.03

0.025

0.02

0.015

0.01

Direct path Random path Adaptive Adaptive, constrained

Figure 8. Adaptive Path-Planning Performance, in Simulation.

Sol Action SR Target Found Target Valid? Pointing Accurate? Notes

12 TC30 122 Rocks OK OK
TC31a 123 Foreground rock OK OK
TC31b 125 Foreground rock OK OK
TC32 128 Rock pile, sediment OK OK
13 AT-13-09 172 Disturbed rocks and sediment OK OK
PP22 166 Distant rock patch OK BAD 1
PP23 160 Distant rock patch OK OK
PP24 154 Foreground rocks OK OK
PP25 148 Foreground rocks OK OK
TC34 158 Foreground sediment patch / rocks OK OK
TC34-recon 132 None OK n/a 2
TC41 170 Rocks OK BAD 3
TC42 164 Distant rock patch OK BAD 4
14 AT-13-10 190 Car BAD BAD 5
PP19 216 Foreground Rock OK OK
TC40 183 Rock patch OK OK

Table 1. Target Detection Results from Playa Exploration Phase.

1: Aims too high for distant targets. 2: No target in scene. 3: Targeted feature not the largest rock. 4: Very distant feature. 5: Cars in frame.

show the field of view of the high-resolution follow- may be attributed to low signal to noise. Alternative-
up image collected by the mast-mounted camera ly, these features may have spectral signatures that
(right column). Each follow-up image is accompanied were very similar to the background substrate.
by a 3 x 3 raster of spectra. Even when the target selec- We deployed the adaptive navigation system suc-
tion was successful, we did not notice a significant dif- cessfully in two instances during the 2013 field cam-
ference between the on- and off-target spectra. This paign. Near the end of the field season the rover vis-

WINTER 2014 57
Articles

Figure 9. Examples of Targets Detected in Midtraverse, and Associated Followup Images.

Sediment / Rock

end

start

Playa

Figure 10. A Composite of Rover Images Showing the Playa Where Adaptive Navigation Was Evaluated.

ited a playa — a dry lakebed approximately 2 kilo- large pixels represent 15-meter ASTER data. Here
meters in length that was spectrally distinctive from Zoë’s planned path, in blue, diverts to sample the
the surrounding terrain (figure 10). Figure 11 shows spectrally distinctive playa surface. The path
a typical round of adaptive path planning near the changed only slightly in subsequent replanning as
playa edge. Here the playa is visible as a bright area in the rover visited each waypoint and incorporated the
the lower right of the overhead satellite image. The new spectra into its solution.

58 AI MAGAZINE
 Articles

Original image Detected targets Followup image

Figure 11. Demonstration of Adaptive Path Planning.

While the on-board planning gave an intuitive directly address these registration errors with the use
and reasonable answer, the actual rover paths were of explicit ground control points (GCPs).
not as expected due to misregistration between
orbital data products and the rover’s on-board GPS Conclusions
estimate. Postanalysis of the data revealed the real This work demonstrates novel techniques integrating
position was offset by more than 100 meters from the adaptive autonomous science activities with pre-
intended location, so the actual rover path spent planned data collection. Zoë’s system will continue
most of its time on the playa. In the future we will to mature in the coming year.

WINTER 2014 59
Articles

Acknowledgments tano, R.; and Judd, M. 2012. Aegis Auto- In Situ Rock and Layer Analysis. Geophysical
mated Targeting for Mer Opportunity Research Letters 40(16): 4188–4193. dx.doi.
A portion of this research was per- Rover. ACM Transactions on Intelligent Sys- org/10.1002/grl.50817
formed at the Jet Propulsion Laborato- tems Technology (TIST) 3(3): Article 50. Wettergreen, D., and Wagner, M. 2012.
ry, California Institute of Technology. dx.doi.org/10.1145/2168752.2168764. Developing a Framework for Reliable
The Landsat map in Figure 3 is cour- Foil, G.; Thompson, D. R.; Abbey, W.; and Autonomous Surface Mobility. Paper pre-
tesy Ken Tanaka (USGS). The geologic Wettergreen, D. S. 2013. Probabilistic Sur- sented at the 12th International Sympo-
classification comes from Jeffery Moer- face Classification for Rover Instrument Tar- sium on Artificial Intelligence, Robotics and
sch (UTenn). Copyright 2013. This geting. In Proceedings of the 2013 IEEE Inter- Automation in Space (i-SAIRAS-12), Turin,
research is funded by the NASA Astro- national Conference on Intelligent Robots and Italy, 4–6 September.
biology and Technology for Exploring Systems (IROS), 775–782. Piscataway, NJ: Wettergreen, D.; Wagner, M. D.; Jonak, D.;
Institute of Electrical and Electronics Engi- Baskaran, V.; Deans, M.; Heys, S.; Pane, D.;
Planets (ASTEP) program under grant
neers. dx.doi.org/10.1109/IROS.2013.6696 Smith, T.; Teza, J.; Thompson, D. R.; Tomp-
NNX11AJ87G.
439 kins, P.; and Williams, C. 2008. Long-Dis-
Green, R. O.; Eastwood, M. L.; Sarture, C. tance Autonomous Survey and Mapping in
References M.; Chrien, T. G.; Aronsson, M.; Chippen- the Robotic Investigation of Life in the Ata-
dale, B. J.; Faust, J. A.; Pavri, B. E.; Chovit, C. cama Desert. Paper presented at the 9th
Cabrol, N. A.; Wettergreen, D.; Warren- J.; Solis, M.; Olah, M. R.; Williams, O. 1998. International Symposium on Artificial
Rhodes, K.; Grin, E. A.; Moersch, J.; Diaz, G. Imaging Spectroscopy and the Airborne Vis- Intelligence, Robotics and Automation in
C.; Cockell, C. S.; Coppin, P.; Demergasso, ible/Infrared Imaging Spectrometer (Aviris). Space (i-SAIRAS-08), Los Angeles, CA USA,
C.; Dohm, J. M.; Ernst, L.; Fisher, G.; Glas- Remote Sensing of Environment 65(3): 227– 26–29 February.
gow, J.; Hardgrove, C.; Hock, A. N.; Jonak, 248. dx.doi.org/10.1016/S0034-4257(98)
D.; Marinangeli, L.; Minkley, E.; Ori, G. G.; Woods, M.; Shaw, A.; Barnes, D.; Price, D.;
00064-9 Long, D.; and Pullan, D. 2009. Autonomous
Piatek, J.; Pudenz, E.; Smith, T.; Stubbs, K.;
Ip, F.; Dohm, J.; Baker, V.; Doggett, T.; Science for an Exomars Roverlike Mission.
Thomas, G.; Thompson, D.; Waggoner, A.;
Davies, A.; Castano, R.; Chien, S.; Cichy, B.; Journal of Field Robotics 26(4): 358–390.
Wagner, M.; Weinstein, S.; and Wyatt, M.
Greeley, R.; Sherwood, R.; Tran, D.; dx.doi.org/10.1002/rob.20289
2007. Life in the Atacama: Searching for Life
Rabideau, G. 2006. Flood Detection and
With Rovers (Science Overview). Journal of
Monitoring with the Autonomous Science- David Wettergreen is a research professor
Geophysical Research: Biogeosciences 112(G4):
craft Experiment Onboard EO-1. Remote at the Robotics Institute at Carnegie Mellon
28, 2007. dx.doi.org/10.1029/2006JG000
Sensing of Environment 101(4): 463–481. University. His research focuses on robotic
298
dx.doi.org/10.1016/j.rse.2005.12.018 exploration underwater, on the surface, and
Castaño, A.; Fukunaga, A.; Biesiadecki, J.; in air and space; and in the necessary ingre-
Mastrodemos, N.; Kubitschek, D. G.; and
Neakrase, L.; Whelley, P.; Greeley, R.; Lem- dients of perception, planning, learning
Synnott, S. P. 2005. Autonomous Naviga-
mon, M.; Castano, R.; and Chien, S. 2008. and control for robot autonomy. His work
tion for the Deep Impact Mission Encounter
Automatic Detection of Dust Devils and with Comet Tempel 1. Space Science Reviews spans conceptual and algorithm design
Clouds on Mars. Machine Vision and Appli- 117(1–2): 95–121. dx.doi.org/10.1007/s11 through field experimentation and results
cations 19(5-6): 467–482. dx.doi.org/10. 214-005-3394-4 in mobile robots that explore the difficult,
1007/s00138-007-0081-3 dangerous, and usually dirty places on
Schaepman-Strub, G.; Schaepman, M.;
Chien, S.; Sherwood, R.; Tran, D.; Cichy, B.; Painter, T.; Dangel, S.; and Martonchik, J. Earth, in the service of scientific and tech-
Rabideau, G.; Castano, R.; Davies, A.; Man- 2006. Reflectance Quantities in Optical nical investigations. Much of this work is
dl, D.; Frye, S.; Trout, B.; Shulman, S.; and Remote Sensing: Definitions and Case Stud- relevant to ongoing and future space activ-
Boyer, D. 2005. Using Autonomy Flight ies. Remote Sensing of Environment 103(1): ities.
Software to Improve Science Return on 27–42. dx.doi.org/10.1016/j.rse.2006.03. Greydon Foil is a Ph.D. student at the
Earth Observing One. Journal of Aerospace 002 Robotics Institute at Carnegie Mellon Uni-
Computing, Information, and Communication
Swayze, G.; Clark, R.; Sutley, S.; and Gal- versity. He is interested in large-scale map-
2(4): 196–216. dx.doi.org/10.2514/1.12923
lagher, A. 1992. Ground-Truthing Aviris Min- ping and planning for autonomous rovers,
Davies, A.; Chien, S.; Baker, V.; Doggett, T.; eral Mapping at Cuprite, Nevada. Summaries of with current work involving physical
Dohm, J.; Greeley, R.; Ip, F.; Cichy, B.; the Third Annual JPL Airborne Geosciences process modeling for better terrain under-
Rabideau, G.; Tran, D.; Sherwood, R. 2006. Workshop, Volume 1: AVIRIS Workshop. JPL standing.
Monitoring Active Volcanism With the Publication 92-14, 47–49. Pasadena, CA: Jet
Autonomous Sciencecraft Experiment on Propulsion Laboratory, California Institute Michael Furlong is a researcher with inter-
EO-1. Remote Sensing of Environment 101(4): of Technology. ests in theoretical neuroscience, decision
427–446. dx.doi.org/10.1016/j.rse.2005.08. making, and robotics.
Thompson, D.; Bornstein, B.; Chien, S.;
007
Schaffer, S.; Tran, D.; Bue, B.; Castano, R.; David R. Thompson is a researcher at the
Doggett, T.; Greeley, R.; Chien, S.; Castano, Gleeson, D.; and Noell, A. 2013. Jet Propulsion Laboratory, California Insti-
R.; Cichy, B.; Davies, A.; Rabideau, G.; Sher- Autonomous Spectral Discovery and Map- tute of Technology. His work involves the
wood, R.; Tran, D.; Baker, V.; et al. 2006. ping Onboard the EO-1 Spacecraft. IEEE application of computer vision, machine
Autonomous Detection of Cryospheric Transactions on Geoscience and Remote Sens- learning, and spectroscopy to robotic Earth
Change with Hyperion On-Board Earth ing 51(6): 3567–3579. dx.doi.org/10.1109/ and planetary science. His algorithms have
Observing-1. Remote Sensing of Environment TGRS.2012.2226040 run on autonomous robots and sensors
101(4): 447–462. dx.doi.org/10.1016/j.rse. Wagstaff, K.; Thompson, D. R.; Abbey, W.; fielded in North America, South America,
2005.11.014 Allwood, A.; Bekker, D. L.; Cabrol, N. A.; the Atlantic Ocean, airborne demonstra-
Estlin, T.; Bornstein, B.; Gaines, D.; Ander- Fuchs, T.; and Ortega, K. 2013. Smart, Tex- tions, low Earth orbit, and the surface of
son, R. C.; Thompson, D. R.; Burl, M.; Cas- ture-Sensitive Instrument Classification for Mars.

60 AI MAGAZINE