OVERCOMING THE CHALLENGES OF SOLAR ROVER AUTONOMY: ENABLING LONG-DURATION PLANETARY NAVIGATION
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
OVERCOMING THE CHALLENGES OF SOLAR ROVER
AUTONOMY: ENABLING LONG-DURATION PLANETARY
NAVIGATION
Olivier Lamarre1 and Jonathan Kelly2
1
STARS Laboratory, University of Toronto, Canada, E-mail: olivier.lamarre@robotics.utias.utoronto.ca
2
STARS Laboratory, University of Toronto, Canada, E-mail: jkelly@utias.utoronto.ca
ABSTRACT Mars 2020 rover to cache samples and drop them
The successes of previous and current Mars rovers at a prespecified location [1]. The second phase
have encouraged space agencies worldwide to will then land both a solar-powered rover and an
pursue additional planetary exploration missions ascent vehicle on the surface, to fetch the cache
arXiv:1805.05451v2 [cs.RO] 13 Jul 2018
with more ambitious navigation goals. For exam- container and launch it into Martian orbit, respec-
ple, NASA’s planned Mars Sample Return mis- tively. As part of the third and final phase, an or-
sion will be a multi-year undertaking that will biter will capture the sample container and return
require a solar-powered rover to drive over 150 it to Earth.
metres per sol for approximately three months. In 2011, it was reported that the “fetch” rover
This paper reviews the mobility planning frame- for the MSR mission would be expected to drive
work used by current rovers and surveys the major up to fourteen kilometres in approximately three
challenges involved in continuous long-distance months [2], corresponding to more than 150 me-
navigation on the Red Planet. It also discusses tres per sol. Technically, Opportunity and Curios-
recent work related to environment-aware and ity are already capable of driving such a distance
energy-aware navigation, and provides a perspec- during a single sol, but, in reality, this pace can-
tive on how such work may eventually allow a not be maintained. In fact, since the beginning of
solar-powered rover to achieve autonomous long- its mission, Curiosity has driven over 130 metres
distance navigation on Mars. during a single sol only four times [3]. This can
be attributed to numerous factors, including op-
erational restrictions, flight hardware limitations,
1 INTRODUCTION
and constraints related to the mission’s science-
Surface mobility on Mars has tremendously accel- driven goals. Nevertheless, such limitations raise
erated the study of our solar system neighbour. In questions about the ability of future solar-powered
1997, the National Aeronautics and Space Admin- mobile robots, like the MSR fetch rover, to navi-
istration (NASA) successfully delivered the So- gate long distances at a relatively fast pace across
journer rover to the surface of the Red Planet, as the diverse Martian terrain.
part of the Mars Pathfinder mission. This little
rover was the first robot to navigate on another Numerous survey and review papers have already
planet. Several years later, in 2004, the successful discussed the future of mobile exploration of Mars
landings of the Mars Exploration Rovers (MERs), (e.g., the authors of [4] review generic adaptive
Spirit and Opportunity, marked the beginning of capabilities and behaviours of rovers operating
an active and ongoing mobile presence on Mars, within a fully autonomous framework). How-
which continues to the present day. In 2012, the ever, to the best of our knowledge, no paper has
Curiosity rover, part of the Mars Science Labora- described the challenges of rover navigation in
tory (MSL) mission, became the largest rover to energy-limited settings, considering the present
date to explore the Martian surface. operational context and near-future mission goals.
This paper surveys how the current operational
Despite the very large amount of data collected by and technological challenges of planetary mobil-
rovers across different regions of Mars, the capa- ity planning and execution can be addressed by
bility of these rovers to conduct experiments on recent research in the field. More specifically, we
soil samples, for example, is very limited. This identify work that could be applied in the context
is why NASA is preparing a three-phase endeav- of long-distance navigation autonomy with solar-
our, called the Mars Sample Return (MSR) mis- powered rovers, to meet the navigation require-
sion, to return Martian soil samples to Earth. The ments of the MSR mission and beyond.
first phase of the mission will involve using theThe remainder of the paper is structured as fol- bilities [5]. On MSL, a single science campaign
lows: Section 2 describes the state of the art in can last more than a week, due to both the large
navigation planning for Mars rovers, while Sec- number of international researchers involved in
tion 3 highlights the current challenges related to the process and the complexity of the numerous
mobility on Mars. Section 4 introduces promis- sensors and analyzers on board the rover. This
ing work towards energy-aware navigation; Sec- timeframe is too slow (fast) to be supported by
tion 5 discusses how the per-sol driving range the tactical (strategic) planners. The supratac-
of solar-powered rovers may be substantially im- tical cycle therefore serves as an intermediary
proved over the next few years. for multi-sol scientific campaigns, and acts as a
bridge between the international scientific teams
and the engineering personnel directly communi-
2 CURRENT ROVER cating with the rover.
NAVIGATION FRAMEWORK
Navigating on Mars is very challenging. In order
to understand the implications of the research pre- 2.2 Curiosity’s Navigation Modes
sented in the remainder of this paper, it is neces- Curiosity was initially equipped with three pri-
sary to be familiar with the current rover naviga- mary navigation modes, each with a different level
tion planning and execution process. This section of autonomy [7]. Blind-drive (the first mode) in-
provides a very brief description of the operational volves letting the tactical team assess the terrain
framework used to drive Curiosity and outlines the around the rover and then prepare a sequence of
rover’s navigation capabilities. instructions for the exact manoeuvres to accom-
plish (e.g., drive 2 metres forward, turn 90 de-
2.1 Navigation Planning grees right, etc.). As the rover executes these
commands, it keeps track of its own motion using
The operational framework employed to control
wheel odometry. Since this is a dead-reckoning
Curiosity is composed of three distinct cycles:
technique (which causes the uncertainty of the
strategic, supratactical, and tactical. The strategic
state of the rover to increase with distance), blind-
cycle can vary from a few weeks to a few months,
drive can only be used over short distances in most
and involves high-level and long-term activity
situations.
planning; it heavily relies on input from the sci-
ence teams to select sites to investigate (based pri- The second navigation mode, auto-navigation,
marily on orbital data), incorporates preliminary employs a level of autonomy in order to reach a
traversability and activity assessment, and contin- distant waypoint identified by the planning team.
uously ensures that the mission maintains focus The rover frequently stops to survey the sur-
on its objectives while accounting for changes in rounding terrain using stereo vision, assess the
the rover’s condition and capabilities [5]. The tac- traversability of the local region, and select a safe
tical cycle, on the other hand, lies at the opposite path to move closer to its goal. This process is
end of the planning spectrum, typically lasting a accomplished using the Grid-based Estimation of
day (or a few days at most). The purpose of this Surface Traversability Applied to Local Terrain
cycle is to ensure that daily mission goals, formu- (GESTALT) algorithm, which is only able to de-
lated by the strategic and supratactical teams, are tect and avoid geometric obstacles [8]. In auto-
fulfilled. Tactical mission planners mainly utilize navigation mode, Curiosity can also use visual
surface data collected by Curiosity’s sensors (such odometry (VO) at a very low frequency (roughly
as stereo imagery and telemetry). Tactical plan- every 10 metres) to verify the rover’s egomotion
ning is highly reactive: on every cycle, the tactical (by performing “slip checks”).
planners analyze the rover’s current state and ad-
The final navigation mode is the slowest, but ar-
just the next sequence of instructions accordingly
guably the safest. In addition to navigating using
before uplinking commands to the rover [6]. Both
obstacle detection and avoidance, Curiosity uses
the strategic and tactical cycle structures were in-
VO over short distances (1.5 metres or so) as an
herited from the MER planning framework.
additional source of relative motion information.
As mentioned above, the MSL mission introduced This is useful when fine positioning of the rover is
a ‘supratactical’ cycle into the standard planning required, or when driving on terrain with an ele-
framework. This is a direct consequence of Cu- vated risk of slip. Since 2012, other drive modes
riosity’s advanced and extensive scientific capa- have been uploaded to the rover (such as visualtarget tracking) or derived from already-existing on a regular basis must be detailed instructions
capabilities (e.g., combinations of blind-driving, rather than more generic goals. Such detailed
VO, and geometric terrain assessment) [9]. command sequences necessitate an assumption of
the rover’s state and resources throughout the en-
Curiosity inherited its VO-based capabilities for
tire daily plan. Since such knowledge is uncertain
online slip detection and improved state estima-
(especially for long command sequences), highly
tion from the MER program [10]. Although rela-
conservative estimates are often made, resulting in
tively simple to measure once it is happening, slip
a recurrent under-use of the rover’s time and en-
is very hard to predict from vision data only. Mis-
ergy. On “traverse” sols (when driving is the main
sion planners have therefore largely relied upon
activity), these issues can be mitigated to some
empirical ‘slip versus slope’ curves, for bedrock,
extent by leveraging Curiosity’s navigation auton-
cohesive soil, and loose sand, derived from ter-
omy capabilities (which enable goal-driven driv-
restrial experiments in simulated Martian environ-
ing). On long traverses, it remains the case that
ments [11].
safety margins (for example attitude boundaries,
maximal slip, or wheel current draw) tend to be
3 LONG-DISTANCE NAVIGATION conservative and can slow the mission down.
CHALLENGES Another operational difficulty arises from the dif-
As with any robotic exploration mission in an ference in the rotation periods of the Earth and
uncontrolled and partially unknown environment, Mars (a day on Mars is 37 minutes longer than a
a number of unexpected factors may affect mis- day on Earth). Because tactical planners normally
sion execution or lead to outcomes different from operate during daytime hours on Earth, the contin-
those initially anticipated. In the case of MSL, uously varying time difference between Pasadena
two key parameters for evaluating mission perfor- (where the NASA Jet Propulsion Laboratory is lo-
mance are the number of soil samples success- cated) and Gale Crater generates recurrent ineffi-
fully analyzed by the onboard laboratory and the cient periods. For example, if the uplink window
total distance driven [7]. Initially, mission plan- at the end of the day on Earth occurs near the end
ners anticipated that Curiosity should be capable of the day on Mars, Curiosity will have to wait un-
of driving 18 kilometres and analyzing 11 sam- til the following sol to execute the requested tasks.
ples within its warranty period (the first 687 sols This leads to a “restricted sol,” during which few
of the mission). However, as of sol 1237, almost or no activities can be scheduled (since the tacti-
twice the mission’s warranty period, the rover’s cal planners have to wait for Curiosity to finish the
onboard laboratory had just analyzed its 12th sam- assigned tasks and communicate with Earth).
ple, and it had traversed less than 12.9 kilometres
[3, 12]. Although there is no doubt that the MSL
mission has been highly successful, such a realiza- 3.2 Environmental Factors
tion raises questions about what factors, exactly, The highly heterogeneous Martian terrain may be
may have contributed to slower overall navigation the most significant element affecting the navi-
progress (besides those purely related to various gation abilities of current rovers. Although Cu-
science campaigns), and how their impact can be riosity was designed to handle sandy, hard, and
reduced in future missions. The following sub- rocky terrains better than the MERs, it was not in-
sections provide an overview of contributing op- tended to be driven over sharp embedded rocks
erational and Martian environmental factors. (formed through wind erosion and called “ven-
tifacts”). Driving over such rocks early in the mis-
3.1 Operational Factors sion resulted in numerous punctures to Curiosity’s
wheels, which in turn has dramatically changed
In 2016, a study of the various mission aspects
how the rover is driven on Mars [14].
influencing the productivity of the MSL opera-
tions group was conducted, with the goal of reduc- One resulting change involves driving more in
ing the quantity of labor-intensive planning tasks sandy environments, to avoid concentrated loads
and improving current and future mission perfor- on the wheels as much as possible—the gen-
mance levels [13]. A key element that stood out eral result is lower driving distances per sol.
from the study is that, because Curiosity has very This is particularly the case during the traverse
little computational power and limited general au- of megaripples, which are aeolian sand accu-
tonomy [9], the commands uplinked to the rover mulations covered by a coarser sand layer [15].The successful crossing of megaripples including the environment. This type of sensing is used ex-
Dingo Gap and Moon Light Valley, and the fail- tensively by Curiosity whenever it arrives at a new
ures experienced in Hidden Valley, revealed that site: imagery provides dense details about the spa-
the traversability of such formations is influenced tial content around the rover. As such, efforts to
by both their geometry (shape, wavelength and extract useful information from these data involve
amplitude) and their material properties. The ex- vision-based terrain classification. At present, due
act material properties cannot be determined by to the low processing power on board Curiosity,
the rover itself and are very difficult for mission dense image processing must happen on Earth.
planners to infer remotely.
One of the solutions to the accelerated wheel dam-
Similar traversability issues were encountered age issue on MSL was the development of a risk-
several times by the MERs on sandy terrain. For aware navigation planning tool to assist tactical
example, Opportunity remained stuck in the Pur- operations and reduce risks associated with hu-
gatory dunes for 38 sols and took six sols to man errors [20]. The goal of this work was to
leave the Jammerbugt ripple [16]. Orbital im- easily distinguish safe and hazardous terrain (es-
agery was later used to identify additional fields of pecially embedded sharp rocks), and plan safe
large ripples, requiring the rover to make several paths based on the physical configuration of the
detours. Unforeseen terrain properties, in fact, surrounding ground and the detected local ter-
brought Spirit’s mission to an end: in April 2009, rain types. A random forest-based algorithm was
the rover broke through a poorly cemented thin used to classify NAVCAM (navigation camera)
crust and embedded itself into unconsolidated soil images, categorizing each pixel as belonging to
[17], where it remained trapped. one of five terrain types. These terrain types were
determined using a set of meaningful intensity and
Operating a solar powered-rover on the Red
gradient-based features extracted from the spatial
Planet adds another source of vulnerability. En-
context around each pixel in each image. The ran-
ergy availability became a major constraint when
dom forest architecture was suitable for this task,
planning activities for the MERs [18] after their
primarily because of its speed of execution, in-
missions extended beyond their warranty periods.
herent robustness to irrelevant or noisy data, and
Energy generation rates are heavily influenced by
ability to capture nonlinear relationships between
the opacity of the atmosphere, the amount of dust
features (which are common in planetary environ-
covering the solar panels, and the frequency of
ments). The classified image data were finally
natural cleaning events caused by the wind [19].
combined with more traditional attitude-related
In addition, as the seasons change, the paths uti-
constraints in a random geometric graph frame-
lized by the MERs must be adjusted due to the
work, to find optimal, safe paths that considered
change in the maximal elevation of the sun in the
the placement of the rover’s wheels.
sky. The typically lower energy levels limit the
activities that can be accomplished in a single sol, A more modern terrain classification approach
while also decreasing the amount of data that can was employed as part of the Mars 2020 landing
be downlinked to Earth [6]. site selection campaign. The Mars Reconnais-
sance Orbiter is equipped with the High Resolu-
4 RECENT RELEVANT WORK tion Imaging Science Experiment (HiRISE) cam-
era, which is able to capture images of the Mar-
Based on the issues raised in the previous section, tian surface with a resolution of 25 centimetres
it is clear that increased situational awareness and per pixel. Although downsampled or cropped
effective energy modelling would allow for bet- HiRISE images are used on a regular basis by mis-
ter predictions of the behaviour of a rover as it sion planning teams, it is impossible for humans
is navigating; in turn, a more productive use of to thoroughly inspect large areas (i.e., Mars 2020
battery resources would be possible. This section landing site candidates) at the full resolution. The
presents a selection of recent research that could Soil Property and Object Classifier (SPOC) sys-
potentially contribute to an increase of navigation tem, built on a convolutional neural network, was
autonomy under energy constraints. trained to assign one of 17 terrain classes to each
HiRISE pixel. SPOC was able to achieve high ac-
4.1 Increased Situational Awareness curacy with only very sparse training labels sup-
Passive exteroceptive sensors, such as cameras, plied by humans experts [21].
perceive natural electromagnetic radiation fromA separate SPOC deep classifier with a similar ar- velocity profiles using a virtual rover model on
chitecture was trained with MSL NAVCAM data, simulated terrains [23]. An estimate of energy
and then used to annotate images taken on sols consumption can easily be computed from these
0 to 938. The results were correlated with thou- values. Such simulations are however computa-
sands of previously-recorded MSL slip events, for tionally expensive and are not suitable for use in
each terrain type, and compared to the empiri- an optimization framework. More ‘convenient’
cal Earth-based slip versus slope curves used by energy models are required for long-distance nav-
tactical planners. As expected, the slip versus igation planning.
slope data retrieved from Martian traverses were
Sakayori et al. have suggested a deep learning-
slightly different from the Earth-calibrated mod-
based approximation [24] to the terramechanics
els, except for the case of sand, where the differ-
simulations for rovers driving on sandy inclines.
ences were dramatic. Once again, this is mainly
The authors used the Wong-Reece wheel-soil in-
due to the failure of vision-based methods to fully
teraction equations and the dynamic model of a 4-
characterize sandy soil.
wheeled rover to generate a set of predetermined
In order to resolve the inability to predict slip terrain and robot configurations and the corre-
on sandy terrain, a different exteroceptive sensing sponding theoretical energy consumption values.
technique has recently been suggested: thermal These data were used to train a feedforward neu-
measurements of the ground, from which thermal ral network to output power consumption based
inertia can be derived. The thermal inertia of a on three input parameters: the rover’s target ve-
sandy terrain describes the rate at which the ter- locity, its heading angle, and the slope angle of the
rain gains or loses heat relative to the surrounding local terrain. The high accuracy achieved by the
environment. This property is strongly character- network was heavily influenced by the assump-
ized by the physical characteristics of the sand, tion that the simulated sandy environment was
such as density, particle size distribution, cemen- uniform; however, this network concept could be
tation and others. Cunningham et al. [22] re- readily extended to incorporate soil mechanical
cently demonstrated that considering thermal iner- properties as inputs.
tia can increase slip prediction accuracy on sandy
A similar idea for reducing computations related
terrain. Although thermal inertia estimates from
to energy consumption is presented in [25]. In-
orbital data have already been used to assess gen-
stead of direct calculation, a series of lookup ta-
eral traversability, work in [22] represents the first
bles store energy consumption values for differ-
use of in-situ measurements for this purpose. The
ent ‘bins’ of slope angle and terrain type, ob-
intuition behind the correlation between thermal
tained from dynamic and terramechanics simula-
inertia and traversability is that, at Martian at-
tions. The bin reference values are used in con-
mospheric pressure, the physical factors affecting
cert with solar energy generation predictions (de-
thermal inertia are also key factors influencing the
pendent on the alignment of the rover’s top plate
amount of slip experienced by a rover on gran-
normal with respect to the sun) to provide an en-
ular terrain. The work in [22] presented a two-
ergy profile for a given, discretized path. In this
experts model approach (considering both thermal
case, the optimality of a path depends on the net
inertia and terrain slope) for in-situ thermal mea-
energy balance, path length, and a risk factor (re-
surement (using Curiosity’s ground temperature
lated to the terrain configuration). Similar to [24],
sensor) and orbital measurement (using the Mars
this framework can also be applied to real-world
Odyssey spacecraft’s thermal imager). For each
platforms.
case, a threshold value separating low from high
thermal inertia sand and a slip versus slope curve A great deal of work to date has focused on pro-
for each regime were learned. This model exhib- ducing energy models and maps from empirical
ited a lower error than the traditional single-expert data collected by proprioceptive sensors on board
model, which only considers terrain slope. planetary rovers. Although such models are typ-
ically biased towards simple terrain types in the
context of real missions because of safety con-
4.2 Energy Models cerns, they generally offer accurate terrain rep-
Existing wheel-soil interaction simulators such as resentations and mobility predictions. Research
the Adams-based Rover Terramechanics and Mo- by Martin et al. [26] has examined methods to
bility Interaction Simulator (Artemis) can already generate energy-optimal paths by driving through
provide an estimate of required wheel torques andan unknown environment several times and post- tactical planners in the context of the MSR mis-
processing the gathered data. The energy cost of sion. Better integration of the strategic and tac-
a path is represented as a function of parameters tical teams during long-distance optimal naviga-
related to mobility (rover velocity, terrain config- tion planning would reduce reactivity and improve
uration) and includes a constant energy sink defin- predictability.
ing the power needs of the internal rover com-
An efficient use of the resources on board the
ponents. Since power consumption data points
MSR fetch rover will be vital to the success of the
are collected along discrete paths in a continu-
mission. To achieve this, more accurate empirical
ous space, gaussian process regression is used for
energy models for predicting energy consumption
interpolation purposes and to provide an under-
and generation will need to be employed. More
standing of where detail is lacking in the path
efficient use of the rover’s time will also be highly
map. This information is used incrementally dur-
valuable—this may be accomplished by carrying
ing the exploration phase until optimal planning
out navigation planning for multiple sols in a row,
across the whole space is possible. The approach
lessening the impact of restricted sols, solar con-
has been demonstrated (statistically) on simulated
junctions, reduced planning efforts during week-
flat terrains and through simplified tests in natural
ends or holidays, and other inefficiencies. Multi-
environments [27].
sol navigation with solar-powered rovers will re-
Otsu et al. [28] have also analyzed ways to ex- quire a dynamic activity planner to prioritize dif-
ploit proprioceptive measurements, developing a ferent tasks (driving along energy-optimal paths,
self-supervised learning framework that uses vi- stopping to replenish batteries, selecting a proper
bration data to train a system to associate visual location to “sleep” for the night, etc.) and to main-
information with energy consumption. A SVM- tain robustness against varying energy generation
based terrain classifier is first trained on features rates. Goal-driven planning and online task prior-
extracted from processed time-series acceleration itization has recently been investigated in [29].
signals (in a supervised manner) to identify dif-
It is also important to note that greater levels
ferent terrain types. Separately, linear regression
of autonomy, through improved online situational
(with two parameters only) is applied to fit em-
awareness (via exteroceptive sensing) and/or in-
pirical energy versus slope data for each terrain
cremental energy model learning, may be possi-
type. The linear relationship is based on several
ble over the next decade with the development of
assumptions: the rover drives at constant veloc-
space-qualified Field Programmable Gate Arrays
ity and the traversable slope is limited and de-
(FPGAs). These co-processors are expected to ac-
forms linearly. A vision-based classifier is then
celerate image processing through parallel com-
trained with the output of the vibration terrain
puting on board the Mars 2020 rover [30].
classifier, to identify visible soil types. Combined
with the corresponding terrain slope information, Lastly, our laboratory is currently developing
energy consumption can then be predicted. This methods to improve solar rover autonomy in
approach has been successfully tested using a real Martian environments by incorporating empirical
rover platform on three different soil types across rover energy consumption information with or-
multiple terrains with different slope angles. bital data, including thermal inertia maps, digi-
tal elevation models, and imagery pre-categorized
with identified terrain classes. These techniques
5 DISCUSSION will be tested on real rover platforms and will
The recent research presented in the last sec- be validated in several Martian analogue environ-
tion may lead to new opportunities for sustain- ments on Earth.
able long-distance navigation planning with solar-
powered rovers such as the MSR fetch rover. In
the past, detailed navigation planning at the strate- 6 CONCLUSION
gic level has been very restricted due to the limited In summary, long-distance planetary navigation
amount of information provided by orbital data. involving solar-powered rovers will be critically
Now, with the ability of algorithms such as SPOC important in the next decade, especially for
to provide richer information about the Martian NASA’s upcoming Mars Sample Return mission.
terrain, more reliable planning is possible from This paper highlighted the operational and ex-
orbital imagery. Identifying preliminary and then traterrestrial environmental factors limiting the
global energy-optimal paths could better inform current mobility planning framework, and re-viewed recent work tackling issues related to situ- the MER Tactical Uplink Process. In: Proc.
ational awareness and improved energy models. AIAA SpaceOps Conf., Rome, Italy.
A more detailed analysis of energy-efficient and [7] Grotzinger JP, Crisp J, Vasavada AR, Ander-
long-distance navigation in a completely au- son RC, Baker CJ, Barry R, Blake DF, Con-
tonomous manner could extend the current survey. rad P, Edgett KS, Ferdowski B et al. (2012)
The present paper did not report on this specific Mars Science Laboratory Mission and Sci-
research area, in order to keep the discussion fo- ence Investigation. In: Space Science Re-
cussed on existing operational processes and flight views, 170(1–4):pp.5–56.
hardware limitations.
[8] Goldberg SB, Maimone MW and Matthies
Acknowledgements L (2002) Stereo Vision and Rover Navi-
This work was supported in part by the Natu- gation Software for Planetary Exploration.
ral Sciences and Engineering Research Council In: Proc. IEEE Aerospace Conf. (AERO’02),
(NSERC) of Canada. The authors wish to thank volume 5, Big Sky, Montana, USA,
Dr. Masahiro Ono and Dr. Kyohei Otsu from the pp.2025–2036.
NASA Jet Propulsion Laboratory for their invalu- [9] Maimone M (2016) A Martian Vision: Im-
able advice regarding ongoing research, which pact of JPL Robotics Vision and Mobil-
greatly assisted in the writing of this paper. ity Research on the Mars Rovers. In:
References JPL Robotics Section Senior Lecture Series,
NASA Jet Propulsion Laboratory, Pasadena,
[1] Mars 2020 Mission Overview. Available California, USA.
at: https://mars.nasa.gov/mars2020/
mission/overview/. Accessed: 2018-04- [10] Maimone M, Cheng Y and Matthies L
10. (2007) Two Years of Visual Odometry on
the Mars Exploration Rovers. In: J. Field
[2] Mattingly R and May L (2011) Mars Sam- Robot., 24(3):pp.169–186.
ple Return as a Campaign. In: Proc. IEEE [11] Heverly M, Matthews J, Lin J, Fuller D,
Aerospace Conf. (AERO’11), Big Sky, Mon- Maimone M, Biesiadecki J and Leichty J
tana, USA, pp.1–13. (2013) Traverse Performance Characteriza-
tion for the Mars Science Laboratory Rover.
[3] The Navigation and Ancillary Infor- In: J. Field Robot., 30(6):pp.835–846.
mation Facility (NAIF). Mars Science
Laboratory Project Kernel. Available [12] Sutter B, Mcadam AC, Mahaffy PR, Ming
at: https://naif.jpl.nasa.gov/ DW, Edgett KS, Rampe EB, Eigenbrode JL,
pub/naif/MSL/kernels/. Accessed: Franz HB, Freissinet C, Grotzinger JP et al.
2018-04-01. (2017) Evolved gas analyses of sedimentary
rocks and eolian sediment in Gale Crater,
[4] Wong C, Yang E, Yan XT and Gu D (2017) Mars: results of the Curiosity Rover’s Sam-
Adaptive and Intelligent Navigation of Au- ple Analysis at Mars (SAM) instrument from
tonomous Planetary Rovers – A Survey. In: Yellowknife Bay to the Namib Dune. In: J.
Proc. NASA/ESA Conf. Adaptive Hardware Geophys. Res, 122(12):pp.2574–2609.
and Systems (AHS’17), Pasadena, Califor-
[13] Gaines D, Anderson R, Doran G, Huff-
nia, USA, pp.237–244.
man W, Justice H, Mackey R, Rabideau G,
Vasavada A, Verma V, Estlin T et al. (2016)
[5] Chattopadhyay D, Mishkin A, Allbaugh
Productivity Challenges for Mars Rover Op-
A, Cox ZN, Lee SW, Tan-Wang G and
erations. In: Proc. AAAI Int. Conf. Auto-
Pyrzak G (2014) The Mars Science Lab-
mated Planning and Scheduling (ICAPS’16)
oratory Supratactical Process. In: Proc.
Workshop on Planning and Robotics (Plan-
AIAA SpaceOps Conf., Pasadena, California,
Rob), London, United Kingdom, pp.115–
USA.
125.
[6] Mishkin A and Laubach S (2006) From [14] Lakdawalla E (2014) . Curiosity wheel
Prime to Extended Mission: Evolution of damage: The problem and solutions.Available at: http://www.planetary. [22] Cunningham C, Whittaker WL and Nes-
org/blogs/emily-lakdawalla/2014/ nas IA (2017) Improving Slip Prediction on
08190630-curiosity-wheel-damage. Mars Using Thermal Inertia Measurements.
html. Accessed: 2018-04-12. In: Proc. Robotics: Science and Systems
(RSS’17), Cambridge, Massachusetts, USA.
[15] Arvidson RE, Iagnemma KD, Maimone M,
Fraeman AA, Zhou F, Heverly MC, Bel- [23] Zhou F, Arvidson RE, Bennett K, Trease
lutta P, Rubin D, Stein NT, Grotzinger JP B, Lindemann R, Bellutta P, Iagnemma K
et al. (2017) Mars Science Laboratory Cu- and Senatore C (2014) Simulations of Mars
riosity Rover Megaripple Crossings up to Rover Traverses. In: J. Field Robot.,
Sol 710 in Gale Crater. In: J. Field Robot., 31(1):pp.141–160.
34(3):pp.495–518. [24] Sakayori G and Ishigami G (2017) Energy
Efficient Slope Traversability Planning for
[16] Arvidson RE, Ashley JW, Bell J, Chojnacki Mobile Robot in Loose Soil. In: Proc. IEEE
M, Cohen J, Economou T, Farrand WH, Fer- Int. Conf. Mechatronics (ICM’17)), Gipps-
gason R, Fleischer I, Geissler P et al. (2011) land, Victoria, Australia, pp.99–104.
Opportunity Mars Rover mission: Overview
and selected results from Purgatory ripple to [25] Fallah S, Yue B, Vahid-Araghi O and Khaje-
traverses to Endeavour crater. In: J. Geo- pour A (2013) Energy Management of Plan-
physical Res., 116(E7). etary Rovers Using a Fast Feature-Based
Path Planning and Hardware-in-the-Loop
[17] Arvidson RE, Bell J, Bellutta P, Cabrol Experiments. In: IEEE Transactions on Ve-
NA, Catalano J, Cohen J, Crumpler LS, hicular Technology, 62(6):pp.2389–2401.
Des Marais D, Estlin T, Farrand W
[26] Martin S and Corke P (2014) Long-Term
et al. (2010) Spirit Mars Rover Mission:
Exploration & Tours for Energy Con-
Overview and selected results from the
strained Robots with Online Propriocep-
northern Home Plate Winter Haven to the
tive Traversability Estimation. In: Proc.
side of Scamander crater. In: J. Geophys.
IEEE Int. Conf. Robotics and Automation
Res, 115(E7).
(ICRA’14), Hong Kong, China, pp.5778–
[18] Bresina JL, Jónsson AK, Morris PH and 5785.
Rajan K (2005) Activity Planning for the [27] Martin S and Corke P (2015) Long Term Op-
Mars Exploration Rovers. In: Proc. AAAI timisation of a Mobile Robot with Proprio-
Int. Conf. Automated Planning and Schedul- ceptive Perception. In: Proc. Australasian
ing (ICAPS’05), Monterey, California, USA, Conf. Robotics and Automation (ACRA’15),
pp.40–49. Canberra, Australia.
[19] Landis GA (2005) Exploring Mars with [28] Otsu K and Kubota T (2016) Energy-aware
Solar-powered Rovers. In: Proc. IEEE terrain analysis for mobile robot explo-
Photovoltaic Specialists Conf., Lake Buena ration. In: Field and Service Robotics, vol-
Vista, Florida, USA, pp.858–861. ume 113 of Springer Tracts in Advanced
Robotics, pp.373–388.
[20] Ono M, Fuchs TJ, Steffy A, Maimone M and
[29] Gaines D, Rabideau G, Doran G, Schaffer
Yen J (2015) Risk-aware Planetary Rover
S, Wong V, Vasavada A and Anderson R
Operation: Autonomous Terrain Classifica-
(2017) Expressing Campaign Intent to In-
tion and Path Planning. In: Proc. IEEE
crease Productivity of Planetary Exploration
Aerospace Conf. (AERO’15), Big Sky, Mon-
Rovers. In: Proc. Int. Workshop on Plan-
tana, USA, pp.1–10.
ning and Scheduling for Space (IWPSS’17),
[21] Rothrock B, Kennedy R, Cunningham C, Pittsburgh, Pennsylvania, USA.
Papon J, Heverly M and Ono M (2016) [30] Seablom MS (2016) Computing Advances
SPOC: Deep Learning-based Terrain Classi- to Enable Speedy New Rover on the Red
fication for Mars Rover Missions. In: Proc. Planet. In: 2015 Science Mission Direc-
AIAA SPACE Forum, Long Beach, Califor- torate Technology Highlights, pp.25–26.
nia, USA.You can also read
