Professional Documents
Culture Documents
I International Conference on
Intelligent Robots and Systems
September 28. October 2,2004, Sendai, Japan
Abslmcf-In walking of quadruped robots, a better performance is obtained by changing the order of swing leg
at a gait. In this paper, we study a walking with obstacles
avoidance of a quadruped robot using neural network (NN)
for determining the moving distance of the robot and for
determining the order of a swing leg. The training data is
considered with initial positions of each leg. The effecti~enes
oflhe present method is demonstrated with some experiments
using TITAN-VIII.
I. INTROIIUCTION
Although legged mobile robots are inferior to wheeled
or crawler types in mobility efficiency on the flat ground,
they demonstrate high motion performance and adaptation
capability to the ground by utilizing their high degrees of
freedom (DOF). Since such robots can choose stable legplacement, stable movements can be performed on irregular
terrains [I]. Moreover. they demonstrate some unique functionalities: e.g., they turn without any slippage; they realize
a stable scaffold when they stop; and their supporting legs
can he selected at any point on the ground[2].
However, it is very difficult to decide robot gait due
to its high DOE When the legs of the robot are simply
controlled by a fixed command, adaptation capability to
the terrain is remarkably restricted and sometimes it is
impossible to maintain a stable walk. Moreover, when a
leg is unable to he placed properly, optimum leg placement
must be efficiently found from among other candidates.
Therefore, it needs for a legged mobile robot to sequentially decide the progression of legs. For that purpose, the
robot predictively perceives and recognizes geographical
features of the terrain, and it consequently gets over any
obstacle by using adaptation ability acquired in advance.
From this fact, legged robots are not necessary to avoid
all the obstacles by altering their path, unlike wheeled
or crawler types. Because, they can avoid an obstacle by
nawling-over or striding, according to the obstacles nature
and the current state of the robot. Thus, it can be found that
the mobility efficiency to reach a destination is improved
by such action. Moreover, when robots have many legs like
4-legged or 6-legged types, the movement range is affected
3400
Authorized licensed use limited to: Khajeh Nasir Toosi University of Technology. Downloaded on December 21, 2009 at 05:43 from IEEE Xplore. Restrictions apply.
PC:RT- Linm
I ~ o r sensor
c ~ I I I Potentiometer I
where
zt = Z f i
Fig. 2. TITAN-VI11
- isi + 13
(6)
-90.0, -65.0 5
oii 5 65.0
3401
Authorized licensed use limited to: Khajeh Nasir Toosi University of Technology. Downloaded on December 21, 2009 at 05:43 from IEEE Xplore. Restrictions apply.
I I
Fig. 4.
0: 1-2-3-4
1 : 1-2-4-3
2: 1 - 3 - 2 4
3: 1-3-4-2
3:4-3-2-1
Fig. 3.
as a basic walk and the order of swing leg is determined. 24 kinds of the order exist in a static walk of the
quadruped robot. Such kinds of the order can be tried to
implement whenever the quadruped robot walks, hut the
hrder of swing leg in this research is determined by a
three-layered NN shown in Fig. 5 . Inputs to the NN are
assumed to be the robot's leg position yfl(k), . . . , yfr(k).
the amount of x- and y-directional movements of the robot
{AXr(k),Al'~(k)} and the tuming angle of the robot
AEr(k). Moreover, we prepare 24 units at the output,
corresponding to 24 kinds of the order. The order of swing
leg fed to the quadruped robot uses the order of the unit
whose output value is closest to one among output units.
= c2ui34i(z)
(8)
i=l
3402
Authorized licensed use limited to: Khajeh Nasir Toosi University of Technology. Downloaded on December 21, 2009 at 05:43 from IEEE Xplore. Restrictions apply.
Generation
Pis. 7. Coordinatcs of the robot's legs
IV. ACQUISITION
OF T H E ORDER
OF SWING LEG
an individual is defined by
fitnesss = error,,
(11)
(12)
2403
Authorized licensed use limited to: Khajeh Nasir Toosi University of Technology. Downloaded on December 21, 2009 at 05:43 from IEEE Xplore. Restrictions apply.
VJl
1/13
z/f3
I"[
"I[
[mm1
[mm1
250.0
250.0 -250.0
2 ~ 0 . 0 -200.0 .
-.
220.0 -270.0
220.0 -240.0
280.0 -270.0
250.0
~
.
..
230.0
260.0
250.0
AXr(k)
Imml
yf4
~~
-250.0
-5.0
-2m.n
-260.0
-280.0
-260.0
30.0
0.0
10.0
-2.0
~~
~~
Order of
swing leg
AYr(k) ABr(k)
"I[
150.0
Idegl
3.0
~00.0
-5.0
200.0
160.0
180.0
0.0
-3.0
5.0
1+4-2*3
4-2-3-1
4-2+1&+3
2-4-3-1
1-2+3&+4
T I
hidden Ihyrr
,-onnection weights of
and parameters of m~~
are trained by GA as Same as the
for the NN
determine the order of swing leg. The associated fitness
function ,,fan individual is defined by
c
ob,
fitness =
+ fitness, + fitness,)
(fitness,
(13)
i=l
fitness, =
0; if there is no collision
10(4,(k)
yze(k)), otherwise
(14)
3404
Authorized licensed use limited to: Khajeh Nasir Toosi University of Technology. Downloaded on December 21, 2009 at 05:43 from IEEE Xplore. Restrictions apply.
TABLE 11
THEMOVEMENT DATA B Y T H E PROPOSED METHOD
-500
0
x"1
fitnesu, =
(15)
fitness
T
~
c -
50
(&(k)
+ yie(k))
(16)
(0.1,0.85),
(%T3? ~ o r 3 =
) (0.5,0.55),
(%II;YOII)
= (-0.5,0.86)
(loi2; ~ o t z =
) (-0.5,0.66)
(2013:Y~13)= (-0.15,0.65)
(%i4;%i4)
= (-0.15,0.85)
(%ri,~ori)
VIII. CONCLUSION
500
(0.1,0.65)
( + o r z : ~ o r z ) = (0.5: 0.75)
(%2,~or2) =
[ml.
REFERENCES
Ill K.K. $ a r k and G.G. A d m s , "Dynamic modclin8 and hydrodynamic prrfamancc of hiomimaic undenualer rahot locomotion:'
Auroriomoas Robots, vol. 13. no. 3, pp. 221-240, 2002.
121 S. Hirose and K. Yoneda "Towuard dcvelopmcnt of pmaical
quadruped wulluq vehicle:' J. of Robotic9 Sociep of Jopon. wl.
I I , no 3. pp. 160,165. 1993.
131 W.S . MacDonald and R. A. Gmpcn, "Building Walking Gaits for
lmgular lcrrain from Baris Conuollcn:' Proc. of the 1997 IEEE
In,. Conference on Robolic~(11111 Auromofioa, vol. 1, pp. 481486,
1997.
141 M. Huher and R. A. Gmpen, '"Prior Smcture for On-line I-caming."
3405
Authorized licensed use limited to: Khajeh Nasir Toosi University of Technology. Downloaded on December 21, 2009 at 05:43 from IEEE Xplore. Restrictions apply.