Shared autonomy system for tracked vehicles on rough terrain based on continuous three-dimensional terrain scanning.

Full text

Page 1

Shared Autonomy System for Tracked Vehicles on Rough
Terrain Based on Continuous Three-Dimensional
Terrain Scanning
••••••••••••••••••••••••••••••••••••
Yoshito Okada, Keiji Nagatani, and Kazuya Yoshida∗
Department of Aerospace Engineering, School of Engineering, Tohoku University, 6-6-01, Aramaki-Aoba,
Aoba-ku, Sendai, Miyagi 980-8579, Japan
e-mail: yoshito@astro.mech.tohoku.ac.jp, keiji@astro.mech.tohoku.ac.jp, kazuya@astro.mech.tohoku.ac.jp
Satoshi Tadokoro
Department of Applied Information Sciences, School of Information Sciences, Tohoku University, 6-6-01,
Aramaki-Aoba, Aoba-ku, Sendai, Miyagi 980-8579, Japan
e-mail: tadokoro@rm.is.tohoku.ac.jp
Tomoaki Yoshida and Eiji Koyanagi
Future Robotics Technology Center, Chiba Institute of Technology, 2-17-1, Tsudanuma, Narashino, Chiba 275-0016, Japan
e-mail: yoshida@furo.org, koyanagi@furo.org
Received 25 January 2011; accepted 23 August 2011
Tracked vehicles are frequently used as search-and-rescue robots for exploring disaster areas. To enhance their
ability to traverse rough terrain, some of these robots are equipped with swingable subtracks. However, man-
ual control of such subtracks also increases the operator’s workload, particularly in teleoperation with limited
camera views. To eliminate this trade-off, we have developed a shared autonomy system using an autonomous
controller for subtracks that is based on continuous three-dimensional terrain scanning. Using this system, the
operator has only to specify the direction of travel to the robot, following which the robot traverses rough ter-
rain using autonomously generated subtrack motions. In our system, real-time terrain slices near the robot are
obtained using two or three LIDAR (laser imaging detection and ranging) sensors, and these terrain slices are
integrated to generate three-dimensional terrain information. In this paper, we introduce an autonomous con-
troller for subtracks and validate the reliability of a shared autonomy system on actual rough terrains through
experimental results. C ?2011 Wiley Periodicals, Inc.
1.
1.1.
INTRODUCTION
Shared Autonomy
Mobility and ease in teleoperation are both important for
mobile robots on rough terrain. A general approach to en-
hance the mobility of the robot is to increase the degrees
of freedom (DOF) of its leg, wheel, steering, or track. How-
ever, this approach naturally takes the ease in teleoperation
away from the robot because the human driver has to han-
dle the increased DOF.
A typical solution for overcoming the trade-off be-
tween mobility and the ease of teleoperation is to generate
the locomotion of the robot by collaboration between the
human driver and an autonomous control. This approach
is known as shared autonomy; for instance, the NASA
Mars Exploration Rovers Spirit and Opportunity (Mai-
mone, Biesiadecki, Tunstel, Cheng, & Leger, 2006) could au-
∗Website: http://www.astro.mech.tohoku.ac.jp.
tonomously travel to the location specified by the driver.
Also, BigDog developed by Boston Dynamics (Raibert,
Blankespoor, Nelson, & Playter, 2008) autonomously gen-
erates its leg motions to maintain stability and realize the
desired direction of travel specified by the driver.
1.2.Search-and-Rescue Robots
A number of research institutes are currently developing
search-and-rescue robots for exploring disaster areas and
obtaining information on victims during the initial stages
of investigation (Tadokoro, Matsuno, & Jacoff, 2005). These
robots are expected to support rescue operations and min-
imize the risk of injury to rescuers and victims from sec-
ondary disasters.
It is extremely important for these robots to have high
mobility over the rough terrains of disaster areas that are
so often strewn with rubble; therefore, tracked vehicles are
frequently used for such applications (Arai, Tanaka, Hirose,
Kuwahara, & Tsukui, 2008; Borenstein and Granosik, 2007;
Guarnieri, Debenest, Inoh, Fukushima, & Hirose, 2005;
Journal of Field Robotics 28(6), 875–893 (2011)
View this article online at wileyonlinelibrary.com • DOI: 10.1002/rob.20416
C ?2011 Wiley Periodicals, Inc.

Page 2

876
•
Journal of Field Robotics—2011
Figure 1.
sensors.
Tracked vehicle test bed Kenaf with three LIDAR
Figure 2.
sors.
Quince, successor of Kenaf, with two LIDAR sen-
Miyanaka et al., 2007; Borenstein & Granosik, 2007; Micire,
2008).
To enhance the traversal ability and stability of
tracked vehicles for search and rescue, some are equipped
with swingable subtracks, which negotiate steps and
bumps in hazardous environments. Our tracked vehicle
test bed Kenaf (Yoshida et al., 2007) and its successor
Quince (Rohmer, Yoshida, Nagatani, & Tadokoro, 2010)
are shown in Figures 1 and 2, respectively. Both robots
possess four subtracks, one at each of the four corners
of the body, covered with the two main tracks. This has
proven to be one of the best configurations, which en-
ables good locomotion over rough terrain, and Kenaf won
the best mobility award in the RoboCupRescue Robot
League (Jacoff, Messina, Weiss, Tadokoro, & Nakagawa,
2003) in 2007 and 2009. In recent years, a number of
robots that have been developed possess this configuration
(Figure 3).
However, we observed that the use of subtracks also
increases the workload of the operator controlling the
robot.Inparticular,manualoperationoftherobotwithsub-
tracks becomes more difficult when the operator teleoper-
ates it with limited camera views sent from the robot itself.
Figure 3.
in group photograph of RoboCupRescue Robot League 2010.
Similarly shaped tracked vehicles having subtracks
1.3. Research Objective and Approach
Our research objective is to overcome the dilemma between
thetraversalabilityandeaseofteleoperationthatisderived
from the use of subtracks. In particular, we aim to achieve
smooth traversal and turning of a tracked vehicle, even by
an unskilled operator, using autonomous control for the
subtracks. Our shared autonomy system comprises a man-
ual controller for the main tracks and an autonomous con-
troller for the subtracks. The performance of an unskilled
operator should be comparable to that of a skilled operator
using fully manual control.
Our system was successfully incorporated into Kenaf
and Quince, and it was confirmed that the autonomous
controller reduces the operator’s workload and maintains
a stable pose of the robot while it traverses and turns on
an actual rough terrain. Some parts of this study were re-
portedin2009(Okada,Nagatani,&Yoshida,2009)and2010
(Okada, Nagatani, Yoshida, Yoshida, & Koyanagi, 2010a,
2010b).
The controller is based on real-time terrain scanning,
which is achieved by two or three LIDAR (laser imag-
ing detection and ranging) sensors (Kawata, Ohya, Yuta,
Santosh, & Mori, 2005). Two LIDAR sensors are attached
on both sides of the robot, and the omissible one is located
at the front of the robot. The use of the front LIDAR sen-
sor enriches the terrain shape information and enables the
controller to produce better subtrack motion for traversing
rough terrain.
Figure 4 shows the configuration of LIDAR sensors.
The autonomous controller for the subtracks can be used
with up to three LIDAR sensors. The optional front LIDAR
sensor obtains a slice of the shape of the terrain being
traversed by the robot at any instant. Moreover, the con-
troller integrates the current terrain slices obtained from
the LIDAR sensors and various recent terrain slices ob-
tained from the front sensor if available, on the basis of
the estimated positions and postures tagged to each slice
to estimate the three-dimensional shape of the terrain near
the robot. Thus, the controller with three LIDAR sensors
Journal of Field Robotics DOI 10.1002/rob

Page 3

Okada et al.: Shared Autonomy System for Tracked Vehicles on Rough Terrain
•
877
Table I.
Basic specifications of Kenaf and Quince.
SpecificationKenaf Quince
Dimensions (mm)
Weight (kg)
Length of subtracks (mm)
DOF
W 400 × L 500
20
235
6 (2 main tracks and 4 subtracks)
W 370 × L 710
26
250
Figure 4.
systems.
Comparison between two-LIDAR and three-LIDAR
can produce better subtrack motions, negotiating narrow
steps/bumps between the subtracks, which are out of the
scanning range of the side sensors.
The rest of this paper is organized as follows. In
Section 2, we introduce related work on tracked vehicles
automatically traversing rough terrain. In Section 3, we
describe our tracked vehicle test beds Kenaf and Quince,
in brief. In Section 4, we present our strategy for the au-
tonomous control of subtracks; this strategy is based on the
motions of subtracks teleoperated by expert operators. In
Section 7, we explain an algorithm for realizing the strat-
egy described in Section 4. We then report our experiments;
we applied the proposed shared autonomy system, includ-
ing the autonomous controller for the subtracks, to Ke-
naf and Quince and performed actual experiments in sim-
ulated disaster environments to validate it. In Section 8,
we report our experimental results and discuss our find-
ings. Finally, we present the conclusions of our study in
Section 9.
2. RELATED WORK
There have been several studies on mobile robots auto-
matically traversing unknown rough terrain using a con-
troller/mechanical behavior.
RHex (Saranli, Buehler, & Koditschek, 2001) is a robot
with six compliant legs that rotate full circle. The design
of RHex was biologically inspired by hexapods. Although
the legs are simply actuated by open-loop control with-
out any intelligence, it was experimentally confirmed that
RHex enables good locomotion like a hexapod on various
general terrains such as grass, bumps, and step field pallets
(Jacoff, Downs, Virts, & Messina, 2008). This simple mech-
Figure 5.
ROBHAZ-DT3.
anism, which does not employ external sensors, is tough
and practical. However, RHex seems to be more likely to
shake its body while traversing rough terrain than tracked
vehicles with active subtracks. Therefore, it could be said
that tracked vehicles with subtracks are more suitable for
equipping a precision manipulator or continuous acquisi-
tion of environment information by external sensors.
ROBHAZ-DT3 (Lee, Kang, Kim, & Park, 2004) con-
tains a passive joint between the anterior and the posterior
tracks; as shown in Figure 5, the joint and tracks are spe-
cially designed for the case in which the robot has to ascend
or descend stairs. The passive motion of the anterior track
triggers the rotation of this joint to enable good mobility
over stairs.
TheHELIOScarrier(Guarnierietal.,2009) isequipped
with an active tail-like mechanism; this tail is intended to
maintain a stable attitude of the robot body on stairs or
steps. It can be operated either manually or by means of
an autonomous controller, and it stabilizes the attitude of
the robot body by pressing its tail to the ground as shown
in Figure 6. The autonomous controller produces motion of
the tail on the basis of the attitude of the robot and distance
from the ground. It assists the robot to move over stairs or
steps.
Ohno, Morimura, Tadokoro, Koyanagi, and Yoshida
(2007) also proposed an autonomous controller for
subtracks(Figure7);thiscontrolleremployscurrentsensors
Journal of Field Robotics DOI 10.1002/rob

Page 4

878
•
Journal of Field Robotics—2011
Figure 6.
HELIOS carrier.
Figure 7.
Autonomous subtrack controller by Ohno.
that measure the torque of each subtrack and position sen-
sitive detector range sensors that are located at the front
and back of the robot body to judge whether the robot is
in contact with the ground. The velocity of each subtrack is
determined on the basis of the above-mentioned judgment
and the posture of the robot. This controller is quite simple
and useful for traversing stairs and steps; however, it can-
not generate advanced motions for the subtracks against a
more complex, rough terrain because the controller is not
based on detailed terrain shape information.
Kadous, Sammut, and Sheh (2006) reported on au-
tonomous traversal of a tracked vehicle having no active
subtracks using behavioral cloning. Their approach was to
use a learning technique that is known as situation-action
behavioral cloning. They simply represented the situation
by the slope of the ground in front of the robot and the pos-
tureoftherobot.Theypreviouslygavepairsofthesituation
and corresponding appropriate action of the main tracks to
their autonomous robot. In an autonomous traversal, the
robot continuously obtained the current situation using a
range imager and accelerometer and controlled the main
tracks by cloning the associated action with the similar sit-
uation. Their application could be called one of shared au-
tonomy systems because it allowed a human observer to
intervene to specify the action of the main tracks. With this
system, they realized semiautonomous navigation on step
field pallets (Jacoff et al., 2008) having short gaps.
Chonnaparamutt and Birk (2008) reported a simula-
tion study on fuzzy control for tracked vehicles with active
rear flippers; they constructed two different autonomous
controllers for the main tracks and the nontracked flippers
to realize fully autonomous exploration of the tracked vehi-
cle. Either controller is based on a fuzzy rule derived from
manual navigation induced by expert operators; one con-
trols the torques of the main tracks according to the differ-
ential of the center of gravity of the vehicle, and the other
controls the position of the flippers according to the pitch
angle of the robot and its differential. Those fuzzy con-
trollers effectively work for stairs and bumps in the sim-
ulation experiments.
Thus, related studies focused on traversing stairs or
steps have not employed detailed information on the ter-
rain shape around the robot. These works are reasonable
approaches for studying traversal of uncomplicated stairs
or steps; however, our interest lies in traversal of unknown
and more complex rough terrain.
3.CONTROL TARGET
The shared autonomy system introduced in this paper em-
ploys a generic algorithm for tracked vehicles with several
subtracks that can widely change the attitude of the robot
body by swinging themselves. In particular, we have con-
sidered a robot with four subtracks as a good application of
the system, because the use of two front and two rear sub-
tracks enables better stabilization performance, and in this
configuration it is easy to watch the ground shape along
the subtracks using LIDAR sensors. It should be noted that
our approach is not designed for robots with only front
or rear subtracks such as a PackBot developed by iRobot
(Yamauchi, 2004) because their subtracks would not take
strong control of the attitude of the robot body in these de-
signs.
In this study, the shared autonomy system was incor-
porated into the tracked vehicle test bed Kenaf (Figure 1)
and its successor Quince (Figure 2). Kenaf and Quince are
six-DOF tracked vehicle test beds for rescue operations.
They have two main tracks covering the body and four sub-
tracks, one at each corner of the body.
Kenaf contains three LIDAR sensors at the front and
both sides of the body to obtain real-time terrain slices. All
motors in Kenaf are encoder equipped, and the circumfer-
ential velocities of the main tracks and angular positions of
the subtracks are obtainable. Kenaf also contains a three-
DOF gyroscope and a gravity sensor.
Moreover, we have incorporated a three-dimensional
odometrymethod(Nagatani,Tokunaga,Okada,&Yoshida,
2008) in Kenaf, which uses the outputs of the main
tracks’ encoders, gyroscope, and gravity sensor to estimate
the position and posture of its body. This method is a
variation of gyrodometry (Borenstein & Feng, 1996;
Maeyama, Ishikawa, & Yuta, 1996) that includes three-
dimensional posture estimation using a gyroscope and a
kinematic model of tracked vehicles with compensation for
track slippage (Nagatani, Endo, & Yoshida, 2007).
Quince’s configuration is almost the same as that of
Kenaf. However, Quince does not have the front LIDAR
Journal of Field Robotics DOI 10.1002/rob

Page 5

Okada et al.: Shared Autonomy System for Tracked Vehicles on Rough Terrain
•
879
sensor because it interferes with the manipulator on its top
face.
On rough terrain, it is quite difficult to estimate a
position with high accuracy over long distances when a
dead-reckoning technique such as odometry is used. How-
ever, the proposed controller requires reliable positions
only along short trajectories over the entire length of the
test bed, which is approximately 90 cm. This is why we
have used position estimation based on three-dimensional
odometry.
4.CONTROL STRATEGY
As mentioned in Section 1.3, we aim to achieve smooth
traversal and turning by a tracked vehicle even by an un-
skilled operator using our shared autonomy system, which
comprises a manual controller for the main tracks and an
autonomous controller for the subtracks. The performance
of the unskilled operator should be comparable to that of a
skilled operator using a fully manual controller. Thus, we
apply control strategies based on the subtrack motions of
the robot when it is operated by skilled operators to the
autonomous controller for the subtracks. Through experi-
ments and robot competitions we participated in, we ob-
served the following four features of fully manual opera-
tions performed by skilled operators:
• To enable the robot to smoothly traverse the terrain, its
posturemustbemaintainedaccordingtotheslopeofthe
ground.
• To enable good locomotion, the main tracks and sub-
tracks should be in contact with the ground as much as
possible.
• An operator spreads the subtracks while directing the
robot along a straight path and folds them while turning
the robot.
• When the pose of the robot becomes unstable, rolling
over should be prevented using the motion of the sub-
tracks.
Considering the above-mentioned four features, we ap-
plied the following strategy to the subtracks and the robot
body:
1. Thepostureoftherobotbodymustbemaintainedparal-
lel to the least-squares plane of the ground surface, and
the robot body must make contact with the ground.
2. The desired posture can be realized by changing the an-
gular positions of the subtracks.
3. The subtrack controller must employ the spreading and
folding modes. The operator can switch between these
modes manually.
4. The desired pose (desired posture and subtrack posi-
tions) must be evaluated and redefined if it is unstable.
To realize the above strategy, we (1) constructed a system to
scan in detail the terrain shape, (2) considered the geome-
try of a generic-shaped subtrack, and (3) designed an algo-
rithm to realize the proposed control strategy. These three
issues are discussed in the following sections.
5.TERRAIN SCANNING
Inthisstudy,LIDARsensorsaretypicallyused.Theleftand
right LIDAR sensors obtain slices of the terrain shape along
the subtracks, and the omissible front LIDAR sensor ob-
tains a slice of the shape of the terrain in front of the robot,
which is immediately traversed by it. The terrain slices ob-
tained from the LIDAR sensors are stored and tagged with
the estimated position and posture of the robot body at the
instant of each terrain scan. These stored slices were inte-
grated according to the procedure described later in this
section to generate three-dimensional information of the
terrain near the robot, which is used in the algorithm de-
scribed in Section 7.
First, we describe the coordinate system used in this
study. Let the robot’s coordinate system be right handed,
its origin be the center of the robot, its x axis be orthogonal
to the front face, and its z axis be orthogonal to the top face.
The position and posture of the robot can be represented
by the relationship between the global and the robot coor-
dinate systems.
In addition, we adopted the quaternion representation
(Horn, 1987) to describe the positions and postures of the
robot. For example, let quaternion p denote the position
vector (xpos,ypos,zpos)Tin the global system and quater-
nion q denote the θrotrotation about the axis of the unit vec-
tor (xrot,yrot,zrot)T. The coordinate conversion from the lo-
cal system {p,q} to the global system can then be described
by the following equations:
pglobal= q × plocal× q−1+ p,
p = [0,xpos,ypos,zpos]T,
q = [cos(θrot/2),xrotsin(θrot/2),
yrotsin(θrot/2),zrotsin(θrot/2)]T.
(1)
(2)
(3)
The scanned points U in the robot coordinate sys-
tem at the moment of scanning are first obtained by
the LIDAR sensor on the robot. We tag U with the es-
timated position p and posture q in the global system
at the moment of the scan and define them as the two-
dimensional terrain information S = {U,p,q}. Let sub-
scripts l, r, and f denote terrain slices from the left, right,
and front sensors, respectively, and let subscript n denote
a terrain slice obtained during the nth control loop. We
Journal of Field Robotics DOI 10.1002/rob