Available online at www.sciencedirect.

com

Planetary and Space Science 52 (2004) 23 – 30

www.elsevier.com/locate/pss

Toward a mobile autonomous robotic system for Mars exploration

P. Arenaa , P. Di Giamberardinob;∗ , L. Fortunaa , F. La Galab , S. Monacob ,
G. Muscatoa , A. Rizzoc , R. Ronchinib
a Universit
a di Catania, Dipartimento di Ingegneria Elettrica, Elettronica e dei Sistemi, Viale A. Doria 6, Catania 95125, Italy
b Universit
a di Roma “La Sapienza”, Dipartimento di Informatica e Sistemistica “Antonio Ruberti”, Via Eudossiana 18, Roma 00184, Italy
c Politecnico di Bari, Dipartimento di Elettrotecnica ed Elettronica, Via Re David 200, Bari 70125, Italy

Received 15 January 2003; received in revised form 11 June 2003; accepted 17 July 2003

Abstract

The paper deals with the results obtained up to now in the design and realization of mobile platforms, wheeled and legged ones, for
autonomous deployment in unknown and hostile environments: a work developed in the framework of a project supported by the Italian
Space Agency.
The paper is focused on the description of the hierarchical architecture adopted for the planning, the supervision and the control of
their mobility.
Experimental results validate the solutions proposed, evidencing the capabilities of the platforms to explore environments in presence
of irregular ground shape and obstacles of di6erent dimensions.
? 2003 Elsevier Ltd. All rights reserved.

Keywords: Robotic platform; Planetary exploration; Path planning; Wheeled robot; Legged robot

1. Introduction be responsible for its own health, to avoid possible damages

and to diagnose and correctly recover accidental software
The recent planetary explorative missions and their re- and hardware failures.
sults, like the NASA Path:nder, as well as the scheduled Nowadays it is commonly accepted the rover autonomy
ones for the next years, put in light the necessity, for plan- as a fundamental keyword in successful planetary explo-
etary unmanned missions, of using mobile platforms, with rations. In fact, in order to compensate the fact that the
a great navigational autonomy. In fact, a rover for plane- robotic platform can be too far to be safely reached in case
tary exploration has to move in rough terrains and unknown of malfunctions of unexpected situation, it is preferable that
environments, like the Martian one. Moreover, the overall an extreme self-reliance and functional autonomy capabili-
mission performance could be heavily a6ected by commu- ties are present on the rover.
nication delays between the rover and the Earth, thus re;ect- Essentially, the concept of autonomy could be sum-
ing in great di<culties for the implementation of a real-time marised in several issues like perception, locomotion, rough
remote control of the rover from the Earth. On the contrary, terrain navigation, self-awareness, functional and decisional
high-level tasks can be e6ectively imposed to the robot, autonomy, safeguard and system integration.
since the time required to perform them is fully compatible In addition to the intelligent behaviour, an important as-
with the communication delays. pect to be considered in the design of the rover is its mechan-
Then, a mobile platform should be able to explore, ical structure, especially for the way of locomotion, since
navigate and perform special tasks taking each time the the irregular ground shape in addition to the presence of ob-
right decisions on the basis of its perception of the external stacles of di6erent dimensions can make the motion very
environment: it results in a great advantage for a rover to di<cult. In our work we propose two di6erent solutions and,
on the basis of the experimental results we are collecting, a

Work partially supported by ASI Contract CT/IR 87/01.
∗
choice will be done for the :nal prototype.
Corresponding author. Tel.: +39-06-44585365;
fax: +39-06-44585367.
The :rst of such solutions makes use of a three wheels
E-mail addresses: digiamberardino@dis.uniroma1.it, robotic platform properly developed for autonomous plane-
paolo.digiamberardino@uniroma1.it (P. Di Giamberardino). tary exploration in presence of rough terrain and unexplored

0032-0633/$ - see front matter ? 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.pss.2003.07.002
24 P. Arena et al. / Planetary and Space Science 52 (2004) 23 – 30

unstructured environment. Each wheel has independent 2. The control architecture

orientation and driving capabilities, assuring high manoeu-
vrability. The use of aluminium ensures lightness and elas- Several intelligent control architectures for complex sys-
ticity, fundamental for payload dimensioning and insulation tems have been proposed in literature in recent years (Acar
from vibrations. Thanks to the particular structure chosen, F uner, 1990; Albus, 1991; Antsaklis and Passino,
and OzgF
described in Section 3, rocks overcoming capability become 1993; Lima and Saridis, 1999; Meystel and Isic, 1988). For
one of the most important features of such a robot. The our rovers we designed a hierarchical control architecture
whole payload has been arranged on the circular platform, based on a multiresolution task decomposition approach that
except the rechargeable storage batteries, fastened under distributes the control action in a hierarchy of closed-loop
the lower surface. controllers working at di6erent levels of resolution. At each
The second solution is based on the use of a legged level it communicates with the supervisor, sending a set of
structure: this kind of choice seems very attractive because commands that drive the system to the desired goal. Each
it allows the robot to reach places where only humans or level can work either in closed- or open-loop mode, receiv-
animals on foot can normally go. This requires great ing a di6erent multiresolutional feedback from the sensors.
adaptability to uneven or dangerous terrains. Many legged We refer to the supervisor as the “software module”:
robots have been built recently, paying particular attention to it manages the overall behaviour of the rover providing a
joint motion and leg coordination and a number of di6erent real operating system capable of implementing planning and
control approaches have been investigated, as discussed in control algorithms. Its functional architecture, as shown in
next sections. The problem of coordinating and controlling Fig. 1, is composed by four elements. The communication
legs is quite challenging and researchers are often helped modules supervise data exchange between the platform, its
by neurobiological studies on animal neuro-motor systems, on-board devices and the ground-based computer. The per-
since even the simplest animals are able to move on legs ception module acquires and organises sensorial informa-
and keep their balance. A Cellular Neural Network-based tion from the environment. In particular, making use of a
control architecture is the solution for motion coordination, stereo vision system, through the computation of the dispar-
as illustrated in the sequel, while the exact structure of such ity matrix from local images the terrain map is built and the
a robot will be discussed in Section 3. position, with respect to the known part of the built map, is
The paper follows this structure: in Section 2 the archi- computed.
tecture of the supervisor and the planner, i.e. the high-level The planning and control module plays a fundamental role
layers of the control architecture, independent from the par- in vehicle behaviour generation, and it will be discussed later
ticular choice of the robotic platform, are shortly illustrated. with more details. The “exception control module” handles
Section 3 is devoted to describe the mechanical solutions the wide ranges of emergency situations that can occur dur-
adopted, evidencing the peculiarities of each of them, and ing terrain exploration. Its tasks can range from the halt of
to describe the low-level controllers for both the mechanical the rover in presence of an imminent collision to a recovery
solutions. Some conclusive remarks in Section 4 complete procedure in absence of the radio-link or an emergency plan
the paper. execution when critical situations are detected. For example,

Fig. 1. The supervisor.

 P. Arena et al. / Planetary and Space Science 52 (2004) 23 – 30 25

Fig. 3. E6ect of HEAD, LINK and JUMP primitives.

Assuming a point-shaped robotic platform, so to make

the procedure independent from the actual robot mechan-
ics and its kinematics, at the top level the planner provides
a sequence of basic motion primitives GOTO(x; y) which
should permit the rover to reach the desired goal location
without colliding with large obstacles and avoiding unde-
sired regions. Only lower layers will take care of how to
avoid or overcome smallest obstacles like rocks or small
pits. In the example of Fig. 3 to reach the target point T the
following GOTO primitives have been generated: (from S)
GOTO N1 and (from N1 ) GOTO T .
Fig. 2. The hierarchical control architecture. The subsequent control layer locally modi:es the path
planned so far in order to improve its safeness, but trying
to preserve the original path. Next, the GOTO are parsed
when battery charge level decreases under a :xed threshold, into simpler motion primitives. Basically, each path segment
the supervisor must communicate to the power control card can be divided into a straight motion manoeuvre and a link
to switch to the auxiliary battery. with the next segment (HEAD and LINK, respectively). The
Fig. 2 shows the “multi resolution hierarchical control JUMP is similar to the HEAD, but operating in rough terrain
architecture” that implements the planning and control sys- with rocks and small obstacles. The JUMP must produce a
tems. Following a task decomposition approach, the higher sequence of basic commands that can steer the rover to a
level performs a :rst planning task, producing a sequence prescribed target avoiding rocks. The path will be a straight
of basic manoeuvres de:ning the commands that the rover line (the shortest) only if no obstacles lie on it.
must execute to reach the goal. Each manoeuvre is then it- The bottom layer of the architecture is dependent on the
eratively decomposed into simpler actions by lower layers. mechanical solution adopted and will be addressed in next
The top level performs a :rst planning making use of con- section after a short description of the two operating rovers.
ventional search algorithms, like A∗ (Hart et al., 1968). The
search space is constituted by a set of free terrain locations
either randomly or uniformly spaced in the grid. The x and 3. The two rovers: description of mechanics and low-level
y coordinates of each node denote the actual coordinates of controllers
the corresponding real location, while the z component rep-
resents an arti:cial dimension whose value is proportional In this section the two prototypal rovers developed are
to the degree of undesirability of the location. In this way A∗ brie;y illustrated in order to evidence their peculiarities
is able to :nd the shortest path avoiding undesired regions, and, for each of them, the description of the low-level local
because dangerous paths take an arti:cial length greater control is provided. The :rst, wheel based, is located at the
than the real one. This recalls the heuristic observation that Dipartimento di Informatica e Sistemistica “Antonio Ru-
a longer safe pathway is better than a shorter and unsafe berti”, University of Rome, Italy, while the second, leg
one. based, can be found at the Dipartimento Elettrico, Elettronico
26 P. Arena et al. / Planetary and Space Science 52 (2004) 23 – 30

e Sistemistico in the University of Catania, Italy. They are

available to the scienti:c community for performing tests
and experimentations. Moreover, a tool for remote control
via Internet of both the robots is under development, so dra-
matically improving the possibility of interaction.

3.1. The wheeled rover

As shown in Fig. 4, the wheeled robotic structure de-

veloped consists of a circular platform with three wheels
mounted at the end of three leg-shaped supports linked to
the platform in a symmetric way. The use of such supports
assures a su<cient highness to the body of the rover to pass
over small obstacles and, at the same time, a great stability
during the motion. Each wheel has independent orientation
and driving capabilities, provided by two electric motors,
with mechanical reductions. The wheels include the drive
motors and micro control units, both hidden in the wheels
hubs. The mechanical design of such wheels is shown in
Fig. 5. Fig. 5. The wheel.
A wired serial bus allows information exchange between
the peripherals and the main control unit.
The vehicle electronics, is basically composed of a The local distributed control is implemented by micro
single-board Pentium-based computer and a set of embed- controllers, which, in turn, generate voltage and current com-
ded microcontroller cards. Six of them are devoted to the mands to send through power drivers to the motors, getting
control of the motors, the other to power distribution and information from incremental encoders.
communication system management and monitoring. A path planner module, designed on the basis of the kine-
All the units exchange information sharing a fast RS485 matical characteristic of the rover, provides the reference
serial common ring. A RS232 bus is used to establish com- commands to each motor.
munication with the ground-station via a RF wireless mo- In past decades, the problem of autonomous mobile robot
dem. The drive and steer motor controller cards employ path planning has been studied extensively, and a great num-
a 20 MHz 16F876 Microchip processor, they provide the ber of techniques have been developed (Campion et al.,
PWM control signal needed to implement PID control laws 1996; Cherif, 1999; Foux et al., 1993; Kubota and Nakatani
for angular rate and position tracking. Tyoshmitsu, 1995; Latombe, 1991; Monaco and Normand-
The perception system is composed by a Videre Design Cyrot, 1992; Di Giamberardino et al., 1996, 2001). In this
stereo vision system mounted over the dome of the rover context, we propose two possible solutions. The :rst one is
trough a two degrees of freedom (azimuth and elevation) a heuristic approach. The basic idea is to steer the collid-
actuated head. ing wheel to avoid the obstacle when it occurs. If several
wheels are in proximity of collision at same time an urgent
exception request is send to the exception control module.
If it occurs, after rover stop, the exception control module
carries out an emergency manoeuvre getting out to stall the
rover.
The second solution is based on A∗ search algorithm. The
optimisation problem starts from a rover’s initial position
and heading. At each step the list of the successor nodes is
obtained performing small admissible elementary rotations
about a set of instantaneous centres of rotation, including the
actual one. In this way it is possible to generate a sequence
of elementary actions, which can drive the rover to the target
across a safe minimum length path.
HEAD, LINK and JUMP are then decomposed in a se-
quence of the simpler motion primitives PIVOT, SPIN,
FORWARD, AHEAD, SWING and TURN (see Fig. 6).
The trajectory generator parses the list of primitives so far
Fig. 4. The wheeled rover. obtained producing time-domain signals in the respect of
 P. Arena et al. / Planetary and Space Science 52 (2004) 23 – 30 27

Fig. 6. E6ect of primitives PIVOT, SPIN, FORWARD, AHEAD, SWING and TURN.

The methodology introduced took its inspiration from the

biological paradigm of the central pattern generator (CPG)
(Shepherd, 1994), which is able to functionally model the
neural structures devoted to generate and control animal
locomotion. Subsequently, a nonlinear slow-fast basic dy-
namics capable of producing signals that qualitatively match
neural dynamics was designed. A number of neurons were
locally connected to one another, with connections mim-
icking electrical synapses. This architecture, an array of lo-
cally connected nonlinear circuits, belongs to the paradigm
of the cellular nonlinear networks (CNNs) (Chua and
Yang, 1988; Roska and Chua, 1993). Subsequently, a CPG
Fig. 7. Example of motion across small obstacles. structure based on CNNs was designed to generate suitable
signals for locomotion generation and control purposes in
biological inspired multipods, like for example hexapods.
the kinematical constraints of the three-wheeled platform. Finally an example of real-time attitude control, by using
The trajectory resulting from a simulated experiment is an array of nonlinear PI connected in parallel with the
shown in Fig. 7. A graphic user interface (GUI) has been CNN, is brie;y presented. The main characteristics of the
also designed in order to provide easy access and full con- overall approach lie in the modularity, robustness, real time
trol of the rover to the remote operator. A di6erent GUI is generation and control, adaptability. These issues make the
available on the on-board computer for current maintenance methodology very promising to work in unstructured and
or direct control of the system during test session. complex environments, like those that met during space
missions.
3.2. The legged rover CNNs were introduced in 1988 as arrays of nonlinear,
locally connected, digitally programmable analog circuits,
For the legged structure, a new approach to the real-time called cells able to solve in real time classical image pro-
generation and control of locomotion patterns is overviewed. cessing tasks (Chua and Yang, 1988). Fig. 8 describes the
The methodology takes the biological aspects of walking main characteristics of CNNs: each single cell is an ana-
multipods into consideration, but never neglects the imple- log circuit, connected to its neighbours via programmable
mentation issues. The basic consideration is that living mov- connections. Each cell has an output piece-wise-linear satu-
ing structures have a great number of degrees of freedom, ration function. The methodology started from the matching
concurrently controlled. of the CNN paradigm with the CPG (Shepherd, 1994) is
28 P. Arena et al. / Planetary and Space Science 52 (2004) 23 – 30

Fig. 8. Characteristics of CNNs.

Fig. 9. Connections for robot gait.

the neurobiological paradigm that functionally describes realised (Arena et al., 1998), and the whole formulation of
the gait generation in animals. This is formulated through the CPG through CNNs was introduced and realised (Arena
a hierarchical structure, where a pool of neurons, the et al., 1999a, b).
command neurons (CNs) receives feedback from the en- In Fig. 9 a schematic representation of the realisation of
vironment and imposes a particular locomotion pattern on the locomotion control of an hexapod robot, able to change
a network of motor neurons, the local motion-generating among the slow, medium and fast gait, thank to a suitable
neurons (LMGNs). LMGNs generate a particular spatio- variation of the connections among the cells is reported. All
temporal dynamics over a network of actuators so as to the details of the implementation and realisation are fully
realise locomotion. Local feedback can of course modulate described in Arena et al. (1999b). One of the most inter-
the characteristics of the current gait. Both the dynamics esting characteristics of this type of implementation is the
related to the CNs and LPGNs have been designed and intrinsic robustness against fault. As described in Arena and
 P. Arena et al. / Planetary and Space Science 52 (2004) 23 – 30 29

Fig. 10. (a) REXABOT I. (b) REXABOT II.

Fig. 11. (a) The Roll and Pitch angles. (b) The attitude controlled variables.

Fortuna (2002), even if one neuron is cut from the ring gen-
erating the wave motions, locomotion succeeds in taking
place, even if a6ected by a lack of synchronisation among
the neurons. Therefore adaptability, robustness, local con-
nectivity and real-time control are among the major aspects
and advantages of this methodology.
Fig. 10a depicts the robot REXABOT I, a 12DOF au-
tonomous hexapod where the methodology previously
addressed was implemented (Arena and Fortuna, 2002).
Since an important aspect regards the possibility to have
a real-time attitude control on this kind of structures, for
example when multipod locomotion generation involves
undulatory dynamics on the robot body, that have to be
controlled, a 18 DOF hexapod, REXABOT II, was built
(Fig. 10b). The basic cell dynamical system that was able Fig. 12. The attitude control scheme.
to drive the 3DOF for each leg was an extended dynamics
with respect to the cell driving the 2DOF leg of Rexabot
I (Arena et al., 2002). Attitude control has to guarantee been used to measure the inclination of the robot. The two
that the robot keeps its body in horizontal position when it signals provided by the sensor reveal the orientation of the
walks on slopping planes. For this task a 2-axis acceleration robot with respect to the Earth plane and are used as error
sensor, located on the middle ventral side of the body, has signals, detecting the error on the roll ’ and pitch angles
30 P. Arena et al. / Planetary and Space Science 52 (2004) 23 – 30

Fig. 13. (a) Pitch control. (b) Roll control.

(Fig. 11a) with respect to the horizontal reference. These Arena, P., Branciforte, M., Fortuna, L., 1998. A CNN based experimental
signals, elaborated and merged to those coming from the frame for patterns and autowaves. Int. J. Circuits Theory Appl. 26,
CPG are used to control the joint variables. The 3 DOF leg 635–650.
Arena, P., Fortuna, L., Branciforte, M., 1999a. Reaction–di6usion CNN
allows us to perform the attitude control of the robot by algorithms to generate and control arti:cial locomotion. IEEE Trans.
applying very intuitive principles. A suitable control law de- Circuits Syst. I: Fund. Theory Appl. 46 (2), 259–266.
scribed in Arena et al. (2002) and shown in Fig. 12 achieves Arena, P., Fortuna, L., Branciforte, M., 1999b. Realization of a reaction–
these results. The two patterns of action can be schematised di6usion CNN algorithm for locomotion control in an hexapode robot.
through matrices containing the gains plus signs associated J. VLSI Sig. Process. Syst. 23 (2/3), 267–280.
Arena, P., Fortuna, L., Frasca, M., 2002. Attitude control in walking
to each of the two joint angles
and , that modulate the hexapod robots: an analogic spatio-temporal approach. Int. J. Circuit
attitude at the level of the single leg (Fig. 11b). The e6ects Theory Appl. 30, 441–458.
of the pitch and roll control are shown in Figs. 13a and b, Campion, G., Bastin, G., Novel, B.D., 1996. Structural properties and
respectively. It can be seen that the body of the robot is classi:cation of kinematic and dynamic models of wheeled mobile
kept horizontal against the slope of the ground plane. robots. IEEE Trans. Robot. Autom. 12 (1).
Cherif, M., 1999. Kinodynamic motion planning for all-terrain wheeled
vehicles. Proc. IEEE Int. Conf. Robot. Autom. 1, 317–322.
Chua, L.O., Yang, L., 1988. Cellular neural networks: theory and
4. Conclusions applications. IEEE Trans. Circuits Syst. I 35, 1257–1290.
Di Giamberardino, P., Monaco, S., Normand-Cyrot, D., 1996. Digital
control through :nite feedback discretizability. Proc. Int. Conf. Robot.
In the paper two di6erent robotic platforms developed and
Autom., 3141–3146.
located in our laboratories, a wheeled one and a legged one, Di Giamberardino, P., La Gala, F., Monaco, S., Normand-Cyrot, D.,
have been illustrated. The main aspects addressed through 2001. Experimental torque control of a wheeled robot under digital
the paper concern the design and the implementation of a motion planning. In: Proceedings of IASTED International Conference
hierarchical control architecture, in which only the lowest on Modelling, Identi:cation and Control, Innsbruck, Austria.
Foux, G., Heymann, M., Bruckstein, A., 1993. Two dimensional robot
layer is dependent on the particular robotic structure, so be-
navigation among unknown stationary polygonal obstacles. IEEE
ing used for any rover whose aim is to move in an unknown Trans. Robot. Autom. 9 (1), 96–102.
or partially known environment in presence of rough sur- Hart, P.E., Nilsson, N.J., Raphael, B., 1968. A formal basis for the
faces and obstacles. The results obtained up to now with heuristic determination of minimum cost paths. IEEE Trans. Syst. Sci.
the two kind of robots con:rm the feasibility of such im- Cybern. 4 (2), 100–107.
Kubota, T., Nakatani Tyoshmitsu, I., 1995. Path planning for planetary
plementation on mobile platforms for planetary exploration.
rover based on traversability probability. IEEE ICAR 739 –744.
Manganaro et al., 1999. Latombe, J.C., 1991. Robot Motion Planning. Kluver Academic, Boston,
MA.
Lima, P.U., Saridis, G.N., 1999. Intelligent controllers as hierarchical
stochastic automata. IEEE Trans. Syst. Man Cybern.—Part B: Cybern.
References 29 (2), 151–163.
Meystel, A., Isic, C., 1988. Pilot level of a hierarchical controller for a
F uner, U., 1990. Design of knowledge-rich hierarchical
Acar, L., OzgF unmanned mobile robot. IEEE J. Robot. Autom. 4 (3), 241–255.
controllers for large functional systems. IEEE Trans. Syst. Man Cybern. Monaco, S., Normand-Cyrot, D., 1992. An introduction to motion planning
20(4) 791–803. under multirate digital control. Proc. 31st Cont. and Dec. Conf.
Albus, J.S., 1991. Outline for a theory of intelligence. IEEE Trans. Syst. pp. 1780 –1785.
Man Cybern. 21 (3), 473–509. Roska, T., Chua, L.O., 1993. The CNN universal machine: an analogic
Antsaklis, P.J., Passino, K.M. (Eds.), 1993. An introduction to intelligent array computer. IEEE Trans. Circuits Syst. I 40, 163–173.
and autonomous control. Kluwer Academic Publishers, Norwell. Shepherd, G.M., 1994. Neurobiology. Oxford University Press,
Arena, P., Fortuna, L., 2002. Analog cellular locomotion control of Oxford.
hexapod robots. IEEE Contr. Syst. Mag. 22 (6).