TELKOM
NIKA
, Vol. 11, No. 6, June 20
13, pp. 3044
~ 305
2
e-ISSN: 2087
-278X
3044
Re
cei
v
ed
De
cem
ber 2
4
, 2012; Re
vi
sed
March 19, 20
13; Accepted
April 13, 201
3
Optimized Passive Coupling Control for Biped Robot
Lipeng YUAN*
1,2
, Am
ur Al Yahm
edi
3
, Liming Yuan
4
1
School of Mec
hanical and
Electrical Engineering, Ha
rbin Institute of T
e
c
hno
logy
, P.O.
Box
420, Har
b
in,
Chin
a 15
00
01
2
Departme
n
t of Mechan
ical a
n
d
Aerosp
ace E
ngi
neer
in
g,
Col
l
eg
e of Engi
ne
erin
g, Corne
ll
Univers
i
t
y
, 3
0
6
Kimbal
l Ha
ll, C
o
rne
ll Un
iversit
y
, Ithaca, NY, USA 1485
3
3
Departme
n
t of Mechan
ical &
Industria
l Engi
neer
ing,
Su
ltan
Qaboos Un
ive
r
sit, P.O. Box 3
3
, OMAN,
Al
kh
ou
d
12
3
4
Capital Aerospace Machiner
y
Company
, No. 1, Nan Dahongmen Road
, Sub-box
91, P.O. Box
34, Beijing,
Chin
a 10
00
76
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: hit
y
l
p
@
126.c
o
m
A
b
st
r
a
ct
A pop
ular hy
p
o
thesis re
gar
di
ng le
gg
ed l
o
co
moti
on is
th
at hu
ma
ns an
d ot
her lar
ge a
n
i
m
als w
a
lk
and ru
n in
a
ma
nn
er that mi
ni
mi
z
e
s the
meta
bol
ic en
ergy exp
e
n
d
itu
r
e for loco
moti
on. Here, w
e
just
consi
der the w
a
lkin
g g
a
it p
a
tterns.
And w
e
prese
n
ted
a hy
brid
mo
de
l for
a pass
i
ve 2
D
w
a
lker w
i
th kn
ees
and
po
int fe
et. T
he dy
na
mi
cs of this
mo
del
w
e
re ful
l
y
der
ived
a
naly
t
ically. W
e
ha
ve a
l
so
pro
p
o
s
e
d
opti
m
i
z
e
d
v
i
rtu
a
l
passiv
e
an
d virtu
a
l c
o
u
p
l
i
ng
contro
l
la
w
s
. This is a
l
so a
si
mp
le
a
nd
effective
g
a
it-
gen
eratio
n
method
bas
ed
on
t
h
is k
nee
d w
a
lk
er
mo
del,
w
h
ic
h i
m
itat
es the
e
nergy
be
hav
ior
in
every
w
a
lki
n
g
cycle. The con
t
rol strategy is forme
d
by taki
ng into
acc
o
u
n
t
the featur
es of mech
an
ical
ener
gy dissi
pat
ion
and restor
atio
n. F
o
llow
i
ng th
e prop
os
e
d
method, w
e
use
comp
uter opti
m
i
z
at
io
n to find w
h
ich gaits
a
r
e
ind
eed
e
nerg
e
t
ically
opti
m
al
for this
mo
del.
And w
e
also
prove s
o
me w
a
lkin
g r
u
les
maybe tr
ue
by t
h
e
results of simulations.
Ke
y
w
ords
: bip
ed rob
o
t, virtua
l passiv
e
, opti
m
i
z
at
io
n, ener
gy
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Why d
o
p
eop
le not
wal
k
o
r
even
ru
n
wi
th a
smooth
l
e
vel gait [1],
like
a waiter
holdin
g
two
cup
s
b
r
i
m
-full of b
o
ili
ng
coffee?
Why do p
eopl
e
sel
e
ct
wal
k
in
g and
runnin
g
from th
e ot
her
possibilitie
s?
We ad
dre
s
s su
ch qu
estio
n
s by
modeli
ng a perso
n as a ma
chin
e
describ
able
with
the eq
uation
s
of ne
wtoni
an
me
cha
n
ics.
We
wi
sh
to
fi
nd h
o
w a
pe
rson
can
get from o
ne
pla
c
e
to
anothe
r with the lea
s
t muscle work.
Passive
dyna
mic wal
k
e
r
s exhibit
a sta
b
le
gait
[2] when placed o
n
a downward
sl
ope
with n
o
a
c
tu
ation. The
s
e
system
s d
e
m
onstrate
h
o
w the inh
e
rent
dynami
c
s of
wal
k
e
r
s can
be
exploited to a
c
hieve n
a
tura
l and ene
rgy-efficient
gaits.
Our main g
o
a
ls are the followin
g
aspect
s
.
1) T
o
p
r
e
s
e
n
t
a math
ema
t
ical m
odel
for
a
simple
two-dime
nsio
nal pl
ana
r
kn
eed
wal
k
e
r
with
point feet an
d kn
ee
s, ma
king it b
o
th a
logi
cal exten
s
ion
of the compa
ss
gait
model a
nd a
physi
cally rea
lizabl
e model
2) Reali
z
atio
n of safe vi
rtual pa
ssive
dynamic
co
ntrol ag
ain
s
t human
bei
ng an
d out
side
environ
ment.
3) A
c
cordi
n
g
to the
co
ntrol strategy,
energy
optim
ization
will
b
e
carried
out
. We
se
ek a
n
explanation
o
f
gait choice with no e
s
se
n
t
ial depen
den
ce on el
asti
c energy storag
e.
2. Model of a
Kneed
Bipe
d
This sectio
n addresse
s th
e walki
ng ro
bot model. We de
al with
a planer bip
ed model
whi
c
h ha
s
kn
ee joints. Fi
g
u
re 1
sh
ows
the model
of a kn
eed bi
pe
d wal
k
ing
rob
o
t and Ta
ble
1
lists its notati
ons a
nd num
erical setting
s for simulatio
n
s.
At the start of
each
step, th
e stan
ce
l
eg i
s
mo
deled
a
s
a si
ngle lin
k
of length
L
, while
the
swi
ng le
g is
modele
d
a
s
t
w
o lin
ks con
necte
d by
a f
r
ictionl
ess joi
n
t. The sy
ste
m
is g
o
vern
e
d
by
its unlo
c
ked
kne
e
dynami
cs
until the swing le
g st
rai
ghten
s out.
Whe
n
the leg
is fully extended,
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Optim
i
zed Pa
ssi
ve
Cou
p
lin
g Control for
Biped Ro
bot (Lipen
g YUAN)
3045
the kn
ee
strike occu
rs. At this point, th
e velo
citie
s
chang
e insta
n
t
ly due to the colli
sion, a
n
d
immediately
afterwa
r
d
s
, we swit
ch to a two-li
nk
sy
st
e
m
in it
s loc
k
e
d
kne
e
dy
na
mics p
h
a
s
e.
Table 1. Nota
tions an
d Nu
meri
cal Settings
Notations Name
Number
&unit
m
t
m
s
Thigh mass
Shank mass
0.5 kg
0.05 kg
m
H
Hip mass
0.5 kg
a
1
b
1
a
2
b
2
q
1
q
2
q
3
u
1
u
2
u
3
Shank length (be
l
ow
point mass)
Shank length (ab
o
ve point
mass)
Thigh length (b
elow
point mass)
Thigh length (a
b
o
ve point mass)
Virtual slope
Stance leg angle w
.
r.t. vertical
Thigh leg angle
w
.
r.t. vertical
Shake leg angle w
.
r.t. vertical
Ankle torque
Hip torque
Knee torque
0.375 m
0.125 m
0.175 m
0.325 m
Rad
Rad
Rad
Rad
N
m
N
m
N
m
Figure 1. Fou
r-lin
k Knee
d Biped Mod
e
l
2.1. D
y
namic
Equations
1) Unlo
cked Knee
Dyn
a
mi
cs. Du
ring
th
e
unlo
c
ked
swing
pha
se, t
he sy
stem i
s
a thre
e-
link pe
ndul
u
m
[3, 4]. The dynamics [5]
are
sho
w
n in
the form of pl
anar m
anip
u
l
a
tor dynami
c
s in
Equation (1).
,
T
c
Hq
q
B
q
q
q
G
q
J
(1)
Whe
r
e,
T
J
is th
e con
s
traint f
o
rce at
kn
ee
-joints.
is the
control in
put
and
c
is the
vector du
e to
the environm
ental forces o
f
the r
obot. The sp
ecifi
c
in
ertia, velocity-dep
end
ent a
nd
gravitational
matrices a
r
e
given in Equa
tion (2).
1
1
12
13
12
22
2
3
1
3
23
33
HH
H
HH
H
H
HH
H
122
2
133
3
21
1
1
233
3
311
1
322
2
0
0
0
hq
h
q
B
hq
h
q
hq
h
q
12
1
22
13
sin
sin
si
n
st
s
h
s
t
ts
t
s
ma
m
l
a
m
m
m
L
g
q
Gm
b
m
l
g
q
mb
g
q
(2)
2) Lo
cked K
nee
Dynami
c
s. After the
kne
e
stri
ke, t
he knee
rem
a
ins l
o
cke
d
and the
system
swit
ch to doubl
e-li
nk pe
ndul
um
dynamic
s. T
he rem
a
ind
e
r of the swin
g
phase o
c
curs
with
straig
ht l
egs.
The
dyn
a
mics fo
r the
ne
wly-
lo
cked
syste
m
a
r
e
exactly tho
s
e
of the
co
mp
ass
gait dynami
c
s but
with a
different ma
ss configu
r
atio
n. The mat
r
ices of dyn
a
mi
cs
are sho
w
n
in
Equation (3) for co
mplete
n
e
ss.
11
12
12
22
HH
H
HH
2
1
0
0
hq
B
hq
12
1
21
2
si
n
si
n
st
s
h
s
t
ts
t
ma
m
l
a
m
m
m
L
g
q
G
mb
m
l
b
g
q
(3)
After the swi
ng foot touch
e
s
the groun
d, the syste
m
swit
ch the
stance and
swi
ng leg.
This
compl
e
t
e
s a f
u
ll st
ep.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3044 – 3
052
3046
2.2. Collision E
v
ent
1) Kneestri
ke Dynami
cs.
We model the kneest
rike as
a di
screte collisi
on event in a
three
-
lin
k ch
a
i
n and switch
to the comp
ass gait mo
d
e
l afterwa
r
d
s
.
Since the on
ly external force
on this
syste
m
is at the st
ance foot, an
gular m
o
me
ntum is p
r
e
s
e
r
ved for the e
n
t
i
re sy
stem ab
out
the sta
n
ce fo
ot and
for the
swing
leg
ab
out the
hip.
Whe
n
lo
oki
n
g
at the
lo
wer
link
of the
swing
leg, howeve
r
,
the knee
strike acts a
s
an external
impu
lse. The
r
efore, angula
r
momentum is
not
con
s
e
r
ved
ab
out the
kne
e
. But the kn
ee j
o
int angl
e correspon
ding to
the knee i
s
l
o
cked
after th
e
colli
sion.
Therefore, its post-collisi
o
n velocity
will
be
that of the second link. We express the
cha
nge in vel
o
citie
s
as:
1
1
2
2
3
q
q
QQ
q
q
q
32
qq
(4)
2) Heelst
rike
Dynamics. T
he
heel
stri
ke
is model
ed
as an in
ela
s
tic colli
sio
n
ab
out the
collidin
g foot. This heel
stri
ke event is, again, i
denti
c
al to the heelstrike for the
compa
s
s ga
it.
Angula
r
mom
entum is then
con
s
erve
d for the
entire system about the collidi
ng foot and for the
swi
ng leg afte
r impact ab
ou
t the hip. Right after
the event, the model switche
s
b
o
th legs an
d th
e
impact fo
ot b
e
com
e
s the n
e
w
stan
ce fo
ot. The mo
de
l also
switch
e
s
ba
ck to the
unlo
c
ked thre
e-
link
dynami
c
s to
sta
r
t a
new ste
p
cycle. Th
e th
i
r
d joint
angle
sta
r
ts
with t
he
same
an
gular
positio
n and
velocity as th
e se
con
d
one
. This co
lli
sio
n
event is expre
s
sed in Eq
uation (5
).
01
10
10
qq
Qq
Q
q
32
qq
(5)
3. Optimized
Virtual Pass
iv
e D
y
namic
Walking
Passive
dyna
mic wal
k
e
r
s exhibit
a sta
b
le,
natu
r
al a
nd
e
nergy-eff
icient gait.
Howeve
r,
the pa
ssive
walke
r
cann
ot
wal
k
on th
e l
e
vel ground
without a
n
y e
x
ternal e
nerg
y
sou
r
ces,
so
we
will introduce “optimized virt
ual grav
ity” for bi
ped robots
to create the wal
k
ing pattern
automatically with the le
ast muscle work an
d with
out
loss of prope
rties of p
a
ssi
v
e dynamic
walk
on the floor.
3.1. Virtual Passiv
e
D
y
na
mic Walking
1) Virtual
Gra
v
ity and Active Wal
k
ing. A
virtual
gravity toward the h
o
rizontal di
re
ction is
as a driving f
o
rce to walk f
o
rward [6]. We can tran
sfo
r
m the virtual gravity e
ffect to the actuato
r
’s
torque a
nd it can b
e
expre
s
sed a
s
follo
ws:
is the virtual slo
pe an
g
l
e.
12
1
22
13
co
s
co
s
t
an
co
s
hs
t
s
t
s
ts
t
s
mL
m
a
m
l
a
m
L
m
L
q
mb
m
l
q
g
mb
q
(6)
The tran
sfo
r
mation of con
t
rol inputs a
r
e
:
1
2
3
11
0
01
1
00
1
u
Su
u
u
(7)
From Equ
a
tio
n
(6) a
nd (7
), we get:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Optim
i
zed Pa
ssi
ve
Cou
p
lin
g Control for
Biped Ro
bot (Lipen
g YUAN)
3047
11
2
1
2
2
1
3
co
s
c
o
s
co
s
t
an
hs
t
s
s
t
t
s
t
s
um
L
m
a
m
l
a
m
L
m
L
q
m
b
m
l
q
m
b
q
g
22
2
1
3
co
s
c
o
s
t
a
n
ts
t
s
um
b
m
l
q
m
b
q
g
31
3
co
s
t
a
n
s
um
b
q
g
2) Multi Virtual Gravity. The dynamics o
f
kneed bip
e
d
robots i
s
mo
re com
p
lex than that
of comp
ass-g
a
it ones. So t
he stea
dy gai
t of a k
nee
d wal
k
er
with si
ngle virtual g
r
avity cannot
be
obtaine
d ea
si
ly without sui
t
able parame
t
er ch
oice.
At the same ti
me, we al
so
want to optim
ize
the total ene
rgy that the bi
ped
robot
ha
s con
s
um
ed
on the a
c
tuat
ors du
ring th
e wal
k
in
g. Ba
sed
on the ob
se
rvation, we propo
se
“M
ulti Virtual Gravity” for the kn
eed bip
ed ro
bot in ord
e
r
to
gene
rate
ste
ady wal
k
ing
pattern
s an
d
optimize
th
e
wal
k
ing e
nergy withou
t
loss of
virtual
passivity (Fig
ure 2.). Th
e transfo
rme
d
torque of
Multi
Virtual Gravit
y effect is given by:
ta
n
cn
k
s
Rq
g
(8)
Whe
r
e,
1
2
3
4
ta
n
ta
n
ta
n
t
a
n
ta
n
ta
n
H
1
2
3
co
s
cos
co
s
c
q
Rq
q
q
12
2
1
hs
t
s
t
s
ts
t
s
mL
m
a
m
l
a
m
L
m
L
mb
m
l
mb
And
nk
s
is the se
nsitivity function of cont
rol inputs (vi
r
tual
gravity scal
e
).
3) Virtu
a
l En
e
r
gy. The
pa
ssivity of virtual pa
ssive
walker i
s
al
so
abl
e to b
e
sho
w
n u
s
ing
“virtual en
erg
y
”, that is:
VV
1
,
2
T
Eq
M
q
q
P
q
(9)
Figure 2. Optimizing M
u
lti Virtual Gravit
y Fields
The virtual po
tential energy is given by:
4
V
1
,
co
s
c
o
s
i
H
i
H
i
J
g
Jg
Pq
(10
)
3.2. Optimized Multi Virtual Grav
it
y
Why do hu
m
ans a
nd oth
e
r anim
a
ls m
o
ve t
he way
they do? An anci
ent hypothe
sis,
dating ba
ck a
t
least to a co
ntempo
rary o
f
Galil
eo and
Ne
wton (Bo
r
elli [7]), is that animals mo
ve
in a m
ann
er that minimi
ze
s effort, pe
rha
p
s
as qu
antified by m
e
tabo
lic
co
st
pe
r di
stan
ce t
r
avell
e
d
[8-10].
In orde
r to op
timizing the t
o
tal actuato
r
’
s
torq
ue en
ergy and keep
steady walki
n
g pattern
without lo
ss
of virtual pa
ssivity, we use
the ‘S
NOPT’
softwa
r
e to
calcul
ating the
suitabl
e cont
rol
para
m
eter of
the biped ro
b
o
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3044 – 3
052
3048
1) Optimi
zing
Target. The
optimizin
g target is
the total actuato
r
’s t
o
rqu
e
ene
rgy
.
During
the Unlo
cked
Knee Stage the optimizi
n
g
target is setting as follo
ws:
22
2
0
12
3
t
dt
(11
)
2) O
p
timizin
g
Co
ndition. A
gait is
ch
aracterize
d by th
e
po
sition a
nd
velocity of th
e body
at the start of a stan
ce ph
a
s
e rel
a
tive to
the stan
ce fo
ot, by the step perio
d, and
by
.
The o
p
timal
solutio
n
s
hav
e cost a
r
bitrarily
cl
ose to
ze
ro
unle
s
s the optimi
z
ation is
further
con
s
trained. Firstly, the co
st
P
can be redu
ce
d nea
rly to zero by taki
ng
small ste
p
s [
11-
13]. So, we
o
p
timize fo
r va
riou
s fixed va
lues of
step l
ength
d
. Se
condly, the
co
st
P
ha
s
a no
n-
anthro
pom
orphic
l
o
wer b
ound (corre
spondi
ng
to st
andi
n
g
on
on
e leg fo
r a
n
i
n
finite time
mid
-
step), a
pproa
che
d
as the a
v
erage
spe
e
d
v goes to ze
ro, so we
con
s
train
v
.
Becau
s
e
we just wa
nt to find the wal
k
in
g gait with the least mu
scl
e work
W
, we
choo
se
low sp
eed
v
=0.185,
a
nd sh
ort step
l
engt
h
d
=0.1.
The
s
e can
gua
ran
t
ee that the
robot o
perate
s
a
cla
ssi
c invert
ed-p
end
ulum
wal
k
ing g
a
it but not an impulsive ru
nnin
g
gait.
We can find a
steady fixed point for the Tabl
e 1 pa
ra
meters whe
n
the virtual slo
pe angle
is 0.050
4 rad.
We use the fixed point and
as the optimi
z
ing initial
co
ndition.
3) Optimi
zed
Nume
rical
Simulation Result. By using the ‘SNOPT’ softwa
r
e an
d
according
to
the optimi
z
in
g target a
nd
con
s
trai
nt
co
ndition, we can
g
e
t
the
re
sult whi
c
h m
a
ke
s
the total a
c
tu
ator’s torque
ene
rgy mini
mal. At
the
same
time, the result ha
s stea
dy wal
k
i
ng
pattern
s
and
can
walk with
out lo
ss of virtual pa
ss
ivity. A limit
cycle
for
the
up
per
link
of on
e le
g
starting from this fixed point
is sho
w
n in F
i
gure 3.
The in
stanta
neou
s velo
ci
ty change
s
from t
he he
e
l
strike an
d knee
strike eve
n
ts can
be observed in this limit cycle
as
straig
ht lines w
here the cycle j
u
mps
with the instanta
neo
u
s
velocity ch
an
ges
whil
e th
e po
sition
s remain the
same. In cont
rast
with the
com
pass
g
a
it,
however, in
a
ddition to th
e
two h
eel-stri
k
es, th
e
r
e
are
two mo
re i
n
st
antane
ou
s ve
locity chang
e
s
prod
uced by the kn
ee
strike
s.
The An
gle t
r
a
j
ectory
of the
virtual
pa
ssi
ve
dynami
c
wal
k
ing
after
energy optimi
z
ation
is
sho
w
n
on
th
e left ha
nd
of Figu
re
4. An
ene
rgy pl
ot, sho
w
in
g p
o
te
ntial en
ergy,
as
well
a
s
tot
a
l
mech
ani
cal e
nergy is
sho
w
n on the right
hand of Figu
re 4.
In this plot, there a
r
e ste
p
increa
se
s in t
he potential e
nergy on
ea
ch foot transfe
r, sin
c
e
we
write th
e
energy of the
system
rel
a
tive to
the sta
n
ce
point. Th
ese i
n
cre
a
se
s a
r
e
sho
w
n
to
exactly bala
n
c
e
out the
ki
netic
ene
rgy
lost th
ro
u
g
h
out on
e ste
p
.
We
ca
n se
e that the fi
nal
mech
ani
cal e
nergy is
con
s
tant through
o
u
t.
And by using
the optimize
d
re
sult to co
ntro
l the bipe
d robot, we can get a batch stabl
e
Eigenvalue
s for the linea
ri
zed map.
Figure 3. Limit Cycle Traj
ectory
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Optim
i
zed Pa
ssi
ve
Cou
p
lin
g Control for
Biped Ro
bot (Lipen
g YUAN)
3049
So, this app
roach can a
c
hieve walkin
g
gait wi
th the
least mu
scle
work at fixe
d sp
eed
and ste
p
leng
th, and kee
p
steady walki
n
g pattern
s wit
hout loss of virtual pa
ssivity.
This maybe
mean
s th
at th
e optimi
z
e
d
multi virtual
g
r
avity pa
ssive
wal
k
in
g p
a
ttern
on th
e
level groun
d
is al
so
anoth
e
r p
eopl
e’s n
a
tural
wal
k
in
g motion. A
n
d pe
ople
ca
n
also u
s
e
th
e
simila
r
walki
n
g pattern o
n
t
he level
grou
nd a
s
on the
slop
e. That i
s
why
as we
wal
k
, e
s
pe
cia
l
ly
downhill, we
do no stop ourselves at
every step. On the contrary,
we let our body fall forward,
only stoppi
ng
ourselves at
each footstep
.
Figure 4.
Virt
ual Passive Dynamic
Walki
ng with Ene
r
g
y
Optimizatio
n
4. Optimized
Virtual Cou
p
ling Contr
o
l
In this se
ctio
n, we intro
d
u
c
e “Optimizi
n
g Vi
rtual Co
u
p
ling Control
”
in orde
r to g
enerate
variable
wal
k
i
ng pattern wi
th respe
c
t to the ene
rgy
le
vels witho
u
t loss of pro
perties of pa
ssiv
e
dynamic walk on the
floor.
The b
a
si
c
co
nce
p
t is
to
re
gard
hybri
d
d
y
namical
sy
stems
as impa
ct-
less (smo
oth) dynamical systems virtu
a
lly and the
e
nergy info
rm
ation is
only utilized fo
r
the
control. In
o
r
de
r to
re
alize
the
rob
o
t
system
pa
ssive and sm
ooth, we co
nsi
der a
virtual
flywh
eel
in co
mpute
r
.
Then th
e bip
ed robot
com
b
ined
with
flywhe
el exhibit
s
pa
ssive an
d sm
ooth a
s
an
augme
n
ted
mech
ani
cal system. After that, we try to
find the walking gait with the lea
s
t muscle
work by usi
n
g
this cont
rol
method.
4.1. Augmen
ted Me
chani
cal Sy
stem
The dynami
c
equatio
n of the
biped robot
is given by:
,
e
MC
g
(12
)
And that of flywhe
el:
f
ff
M
(13
)
The aug
ment
ed mechani
cal system i
s
g
i
ven by the following
equati
on:
,
0
0
0
e
f
f
ff
M
Cg
M
(14
)
We de
note (1
2) as:
,
aa
a
a
a
e
Mq
q
C
q
q
q
g
q
(15
)
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3044 – 3
052
3050
4.2. Couplin
g Control La
w
The co
ntrol in
put
a
for the sy
stem (1
5) i
s
g
i
ven by:
ta
n
0
c
a
f
f
Rg
(16
)
The second t
e
rm of the
rig
h
t hand i
s
the
cou
p
ling
con
t
rol input for t
he de
cou
p
led
system
whic
h is
defined as
following:
T
f
p
P
qq
p
(17
)
At first, we will introdu
ce a
ugmente
d
virtual
ene
rgy which is the
su
m of virtual energy of
the biped rob
o
t and kin
e
tic energy of the
flywheel:
VV
,,
,
,
a
f
Eq
q
E
K
(18
)
The
control formul
ation (1
7)
can
be ex
plaine
d from
the followi
ng
viewpoi
nt. The target
con
d
ition of the augm
ente
d
mech
ani
cal
energy is giv
en by:
V
0
f
d
EK
dt
(19
)
And then we
get:
2T
1
2
ff
f
f
f
f
f
f
dd
KM
M
dt
d
t
(20
)
Hen
c
e, from
(20) the follo
wing re
sult is o
b
tained.
1T
ff
(21
)
This is equiv
a
lent to
(17
)
whe
r
e
1
f
p
. Con
s
i
derin
g the
st
ructure of
(1
0
)
, let u
s
set
the vector
p
in (17) as
the
following form:
1
V
ta
n
a
cf
pR
E
(22
)
Whe
r
e,
*
VV
V
aa
EE
E
is the
cha
nge
of the au
gmente
d
virtual e
nergy and
>0 i
s
the
feedba
ck gai
n.
*
V
E
is the nomi
nal value of the
V
a
E
.
4.3. Optimized Virtual Coupling Con
t
rol
In this
se
ctio
n, we
will
use virtual
co
up
li
ng controller
to
drive
the biped ro
bot, optimize
the total
actu
ator’s torque
energy, an
d
get the
walk
i
ng g
a
it with
t
he le
ast
mu
scle
wo
rk. Th
e
n
we
try to find so
me rul
e
s
of human
wal
k
i
ng through
this
kind
of walkin
g control
pattern th
at i
s
without any reactio
n
forc
e again
s
t external forces.
In orde
r to o
p
timizing th
e
total actuato
r
’
s
torq
ue e
nergy, we u
s
e th
e ‘SNOPT’
software
to calculating
the suitabl
e contro
l pa
ram
e
ter of the biped ro
bot.
1) Nume
rical
Simulation
Re
sult. By using the ‘S
NOPT’ software and a
c
cord
ing to th
e
optimizin
g target and
co
nst
r
aint conditio
n
, we
can
get
the re
sult wh
ich ma
ke
s th
e total actu
ator’s
torque
ene
rg
y minimal. At
the sam
e
time, the metho
d
can
reali
z
e
the rob
o
t syst
em pa
ssive a
n
d
smooth, the energy information
is only utilized for the
control.
An ene
rgy
pl
ot, sho
w
in
g
real a
nd virtu
a
l total m
e
ch
anical e
nergy and
pote
n
tia
l
ene
rgy,
as
well
as
au
gmented
me
chani
cal e
nerg
y
is sho
w
n in
Figure 5. Th
e
gra
ph o
n
the
left hand i
s
t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Optim
i
zed Pa
ssi
ve
Cou
p
lin
g Control for
Biped Ro
bot (Lipen
g YUAN)
3051
result with
e
nergy
optimization, and
th
e graph
on
t
he ri
ght ha
n
d
is th
e resu
lt without e
n
ergy
optimizatio
n.
The a
c
tuato
r
’
s
torque
s
of the virtual
cou
p
ling
co
n
t
rol dynami
c
wal
k
ing
is
sho
w
n i
n
Figure 6.
Th
e left ha
nd i
s
the
result
with e
ner
gy optimizatio
n, and
th
e right
han
d
i
s
with
out
energy optimi
z
ation.
2) Result Analysis. F
r
om
Figure 5 an
d
Figure
6, we
can
see th
at the rob
o
t an
d virtual
flywheel sto
r
a
ge
e
n
e
r
gy
e
a
c
h othe
r
a
nd as a re
sult
of
it
the
total en
ergy
val
ue of the
au
gment
ed
system i
s
ke
p
t
consta
nt. Fu
rtherm
o
re, fro
m
Figure
6 we can
se
e tha
t
the con
s
tant
-like to
rqu
e
is
also
su
cceed
ed. It means that these
wa
l
k
ing p
a
ttern
s are natu
r
al m
o
tion.
Acco
rdi
ng to the simulatio
n
result, we can al
so kno
w
that after energy optimi
z
a
t
ion, the
total mecha
n
i
c
al en
ergy an
d augme
n
ted
mecha
n
ic
al energy are b
o
th smalle
r than those with
out
optimizatio
n.
From th
e
co
mpari
s
o
n
bet
wee
n
the fig
u
re
s of
actu
a
t
or’s to
rq
ue
s, we
ca
n
see
that the
amplitude
s of
the torque
s
of hip joint an
d kne
e
joint keep si
milar b
e
twee
n after
optimizatio
n and
before
o
p
timization,
and
t
he valu
e of
ankl
e
to
rque
is mu
ch
big
ger befo
r
e
o
p
timization.
This
maybe mea
n
s
anoth
e
r
wal
k
ing
rule that
the co
st
co
nsumed by the t
o
rqu
e
s
on kn
ee joint and
h
i
p
joint is simil
a
r in
different
kinds
of walking ga
its,
a
n
d
the bigg
est
different mu
scl
e work ma
inly
come
fro
m
th
e cost
of a
n
kl
e torque. S
o
,
if
human
wan
t
to raise the total input wa
lking en
ergy to
increa
se
the
wal
k
ing sp
ee
d
and enla
r
g
e
step
length
,
the most p
r
actical an
d ef
fective way i
s
to
increa
se the
push-off imp
u
lse
s
force, e
v
en though
p
eople u
s
e th
e muscle of t
h
igh o
r
sh
an
k to
enforce.
Figure 5. Energy Plot
Figure 6. Actuator’
s
Torqu
e
s
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3044 – 3
052
3052
5. Conclusio
n
Here,
u
s
ing n
u
meri
cal opti
m
ization
and
sup
porti
n
g
an
alytical argu
ments,
we
re
alize th
e
safe p
a
ssive
dynamic co
ntrol ag
ain
s
t
human
bein
g
and o
u
tsid
e
environ
ment
and o
b
tain t
h
e
energy minim
i
zing
gaits
by usin
g optimi
z
ed vi
rtual
p
a
ssive and
virtual
coupli
ng control strate
gy.
At low spee
d
s
the optimization discove
r
s the in
ve
rted-p
end
ulum
walk
with the least mu
scle
work. We co
mpare the re
sults of si
mul
a
tions, an
d we discover
so
me wal
k
ing
rules mayb
e true.
(1)
The the o
p
timized
multi virtual gravity pas
sive wal
k
ing
pattern i
s
al
so an
oth
e
r pe
ople’
s
natural
wal
k
in
g motion on the level gro
u
nd.
(2)
The big
g
e
s
t different mu
scle
work m
a
i
n
ly come f
r
o
m
the co
st of
ankl
e
torq
ue
in different
kind
s
of
walki
ng g
a
its. So,
if huma
n
wan
t
to chan
ge t
he
wal
k
ing
p
a
ttern, the
m
o
st p
r
a
c
tical
and effective
way is to ch
a
nge the pu
sh
-off impulse
s force.
Referen
ces
[1]
R McN Alex
ander.
Mecha
n
ics
of biped
al l
o
co
moti
on
. Perg
a
m
on Press, Ne
w
York. 1
976;
1: 493-5
04.
[2]
Vaness
a
F
,
Hsu Ch
en.
Passi
ve Dyn
a
m
ic Walkin
g w
i
th Kn
ees: A Poi
n
t Foot Mod
e
l
. Ma
ssachus
ett
s
Institute of
T
e
chno
log
y
Pr
ess, Massachus
etts. 2007.
[3]
Simon Moc
hon
,
T
homas A McMaho
n, Ballisti
c
w
a
lkin
g.
Jour
nal of Bio
m
ech
anics
. 19
80; 13
: 49-57.
[4]
Simon Moc
h
o
n
,
T
homas A McMahon.
Ballistic
w
a
lki
ng: An impro
v
ed mod
e
l.
Ma
th
em
a
t
i
c
al
Biosci
ences
.1
9
80; 52(3-
4): 24
1-26
0.
[5]
Jean-J
a
cqu
e
s
E, Slotine, W
e
i
p
in
g Li.
App
lie
d Non
lin
ear C
o
ntrol
. Prentice
Hall. 1
990.
[6]
F
u
mihiko
Asa
n
o
, Masaki
Yam
a
kita. Virtu
a
l G
r
avit
y an
d C
o
u
p
lin
g C
ontro
l for Ro
botic Ga
it S
y
nthesis
.
Systems a
nd H
u
mans
. 20
01;
31(6): 73
7-7
4
5
.
[7]
JA Bore
lli. On
the m
o
veme
nt of a
n
ima
l
s
(De
Mot
u
An
i
m
alium, P
a
rs
prima). P. Ma
quet (tra
ns.)
.
Springer-Ver
lag
. Berlin. 19
89:
152.
[8]
R McN Ale
x
an
der. Optimizati
on a
nd
gaits i
n
the l
o
com
o
ti
on of verte
b
rat
e
s.
Physiol. Rev
. 198
9; 69
:
119
9-12
27.
[9]
R McN Alex
ander.
Optima for anim
a
ls
. Princ
e
ton Un
iversit
y
Press, Princeton, NJ. 199
6.
[10]
Srinivas
an. Op
timal s
pee
ds f
o
r
w
a
lki
ng
an
d ru
nni
ng,
an
d
w
a
lk
in
g o
n
a
movi
ng
w
a
lk
w
a
y
.
C
haos
.
200
9; 19: 026-
112.
[11]
R McN Alexan
der. A model o
f
biped
al loc
o
m
o
tion o
n
compl
i
ant legs.
Phi
l
. T
r
ans. R. Soc.
Lon
d
. 199
2;
B338: 18
9-1
9
8
.
[12]
A Ruin
a, JE Bertram, M Sr
inivas
an. A co
llisi
o
n
a
l mod
e
l
of t
he energ
e
tic cost of supp
ort
w
o
rk
qua
litative
l
y
e
x
plai
ns le
g-seq
u
enci
ng in
w
a
lki
ng an
d gal
lop
i
ng, pseu
do-e
l
a
s
tic leg beh
avi
o
r in runn
in
g
and the
w
a
lk-to
-run transiti
on.
J. T
heor. Biol.
200
5; 237: 17
0
-
192.
[13]
AD Ku
o. A si
mple m
o
d
e
l
pr
edicts th
e st
e
p
le
ngth-sp
ee
d relati
onsh
i
p in
huma
n
w
a
lki
n
g
.
Journ
a
l of
Bio
m
ec
han
ical Engi
neer
in
g
. 2001; 12
3: 264-
269.
Evaluation Warning : The document was created with Spire.PDF for Python.