Professional Documents
Culture Documents
I. INTRODUCTION
In recent years, modeling and control of flexible joint
robots (FJR) are gathering more and more interests [1][3].
Flexibility in robot joints originates from flexible mechanisms such as tendon-driven, belt-pulley transmission, harmonic drive as well as the installation of sensors, e.g., joint
torque sensor (JTS). With the diverse growth of the robotics
field, various types of the mechanisms and the sensors are
adopted/developed in robot design, and naturally, control of
FJR is emerged as one of the important topics.
In controlling the FJR, the following model which is
derived under the Spongs assumption [4] is widely adopted.
M (q)
q + C(q, q)
q + g(q) = s
B + s = m + f ,
(1)
(2)
wkchung}@postech.ac.kr
978-1-4799-6934-0/14/$31.00 2014 IEEE
Controller
Motor-side
Dynamics
Link-side
Dynamics
(nonlinear)
Fig. 1. Motor-side dynamics and link-side dynamics are coupled via joint
torque s . The purpose of this paper is to reshape motor-side dynamics into
the user-defined one using DOB structure.
2381
Link-side
Dynamics
+
++
+
-
+ -
1
1
V = qT M (q)q + T B
2
2
1
(a)
Link-side
Dynamics
+
+
+
++
+
(b)
Fig. 2. (a) Motor-side dynamics reshaping using DOB. (b) An equivalent
structure to the DOB.
Link-side
Dynamics
Controller
+
+
+
+
++
= Pn (s)(m s ).
Fig. 3.
A controller is applied to the FJR with motor-side dynamics
reshaping. Note that is fed back into the controller instead of , and
k is additionally introduced.
Q(s)
.
Pn (s){1 Q(s)}
(5)
By virtue of the alternative structure, the analysis and implementation become simpler as will be investigated in next
section.
Moreover, it should be pointed out that, when the controller is applied, nominal (estimated) signals should be fed
back into the controller instead of real (measured) signals as
shown in Fig. 3. Also, to make the theory clear, k in Fig.
3 should be designed to satisfy
e Tnr s + T k 0.
(6)
otherwise,
This section summarizes main results of the paper excluding derivations which will be introduced in next section.
(4)
e nr s ,
k = sgn()
k = 0,
(7)
(8)
2382
3( s) + 1
.
( s)2 + 3( s)2 + 3( s) + 1
1
)(s + KP ),
2
(11)
g
,
s+g
(12)
B + s = m + f + k + a
Bn + Dn + s = m + k
M (q)
q + C(q, q)
q + g(q) = s .
(16)
(17)
(18)
(19)
(20)
(10)
(13)
= Vkin (q,
+ Vpot (q, ) + Va (q, q,
V (q, q,
, )
)
, ),
(21)
wT wdt,
C(s) = (K +
1
1
)(s + KP + KI ),
2
s
(14)
where
Vpotential
(xTnr Sxnr
aT Ra )dt
= T s qT s + qT g(q).
w wdt,
(15)
1
1 T
q M (q)q + T B
(22)
2
2
1
Vpot (q, ) = ( q)T K( q) + Ug (q).
2
(23)
=
Vkinetic Vkin (q,
)
(24)
(25)
Thus,
2383
+ V a (q, q,
V = T m (q, q,
, )
, ).
(26)
Motor brake
Electric motor
+ V a (q, q,
).
)
,
,
V < T m (q, q,
(31)
Here, recall that the nominal signal is fed back into the
).
Noting that (31) has
controller, i.e., m = m (q, q,
,
identical form to (26), and using the invariance principle, the
global asymptotic stability of (17)-(18) can be concluded.
=
For example, consider a PD control of the form m (, )
where d is a desired value and
KC ( d ) KD ,
the gravity g(q) is neglected. Defining Va () = 12 (
= Vkin +
d )T K(d ), the Lyapunov function V (q, q,
, )
Vpot +Va () guarantees the global asymptotic stability of (1)(2). The dynamics (17)-(18) is also globally asymptotically
)
= KC ( d ) KD using
stabilized by m (,
,
s ) = Vkin + Vpot + Va ().
V (q, q,
,
Conversely, since global asymptotic stability of (17)-(18)
is guaranteed, there exists a Lyapunov function Vq,n =
Vq,n (xq , xn ) that satisfies the following by the Lyapunov
converse theorem (theorem 23, and definitions (viii), (x) in
[22]).
Vq,n > 0
Vq,n : radially unbounded
V q,n ||xq ||2 ||xn ||2
(32)
(33)
(34)
(36)
(30)
Belt-pulley mechanism
Harmonic drive
and therefore,
Thus,
(
)
V = e Tnr s + T k T D + T m + V a
Link
Motor-side encoder
,
s ) = Vkin (q,
+ Vpot (q, s ) + Va (q, q,
).
V (q, q,
,
)
,
(27)
= qT (s g(q)) + T (D s + m + k )
V kin (q,
)
(28)
T
T
T
V pot (q, s ) = s q s + q g(q).
(29)
Link-side encoder
(37)
(38)
(39)
V nr x (xnr ) + w (||w||),
(40)
(41)
2384
Fig. 5. PD control only, PD control with inertia reshaping, and PD control with 3-2 order DOB are compared. The performance of the controller is
improved by the exact reshaping of motor-side dynamics using DOB.
m (, )
where KP = 50 N/m, KD = 5 Ns/m represent stiffness
and damping, respectively, and d = 0 is desired motorside position. The controller was implemented on a real-time
OS (RTX) and ran at 500Hz control frequency. It should be
noted that the system significantly suffers from large frictions
caused by belt-pulley mechanism (which is amplified by
160:1 harmonic drive). The nominal dynamics (Pn (s) in
DOB structure) was selected as 1/(Bn s2 + Dn s), where
Bn = 2.3 kg m2 which was obtained by CAD data, and
Dn = 0.01 kg m2 /s which was set arbitrary.
The following three sets of experiments were considered.
1) In the first set of experiments, PD control only, PD
control with inertia reshaping, and PD control with
DOB were compared.
2) In the second set of experiments, DOB was implemented with measured signal () feedback.
3) In the third set of experiments, DOB was implemented
with estimated signal feedback, as proposed in the
paper.
In the second and third sets of experiments, it is verified
feedback is indeed necessary to
that the estimated signal ()
ensure the stability. Experiment scenarios were as follows:
During 5 to 10 sec, external torque was applied (about 15
Nm), and removed. Again, during 20 to 25 sec, external
torque was applied in opposite direction, and then removed.
and
Measured motor position , estimated motor position ,
measured JTS signal are shown in figures. Since the 1DOF
testbed has very high joint stiffness, link-side position is
almost the same as motor-side position, and thus, only the
motor side position is presented in figures.
For the first set of experiments (Fig. 5), when PD control
was applied alone, due to the high friction occurred in the
(42)
where c is a new control input, and = B/Bn = 4, was applied, the friction and motor inertia were effectively reduced,
and the response showed much better result than the previous
one (Fig. 5b). However, due to the remaining uncompensated
frictions, the motor position could not follow estimated
and moreover, the motor position even could not
position ,
return to its original position. The realized stiffness was about
60 N/m which is larger than the desired value. When 32 order DOB is applied, the motor position followed
well, and the desired stiffness was realized (50 N/m). In
overall, due to the high friction caused in motor-side, the
control performance was significantly degraded. Fortunately,
applying DOB effectively eliminated the friction, and the
performance of the controller was successfully improved.
For the second set of experiments (Fig. 6), DOB was
implemented with measured signal () feedback. When 1st
order Q(s) filter was applied, although could not follow
well, the resulting motion was somewhat plausible (Top
row in Fig. 6a). Nice oscillatory was motion occurred (see,
e.g., 10-15 s), and even returned to its original position.
However, this observation does not mean that the stability
is guaranteed; which is one of the states of the entire
system dynamics does not converge to equilibrium (Bottom
row in Fig. 6a). The resulting plausible motion was due
to, paradoxically, weak compensating ability of 1st order
Q(s) filter: although some amount of enr exists, the 1st
order DOB has no ability to compensate the error. This
observation becomes more clear when higher order Q(s)
2385
Fig. 6. DOB is implemented with measured state () feedback. Left column: results of 1st order Q(s) filter. Right column: results of 3-1 order Q(s)
Middle row: JTS signal. Bottom row: magnified view of dotted box of the top figure. The
filter. Top row: measured position (), estimated position ().
asymptotic stability is not guaranteed when the measured signal is fed back into the controller.
feedback. Left column: results of 1st order Q(s) filter. Right column: results of 3-1 order Q(s)
DOB is implemented with estimated state ()
2
Middle row: JTS signal. Bottom row: shows that e nr is dominated by D . P1 = e nr ,
filter. Top row: measured position (), estimated position ().
2
Fig. 7.
2386
filter was applied (Fig. 6). The 3-1 order DOB made
track better, but even though the friction in motor-side was
compensated well, the asymptotic stability was not satisfied
either (Bottom row in Fig. 6b). Therefore, in overall, it
is shown that feeding back the measured signal does not
guarantee the stability.
For the third set of experiments (Fig. 7), DOB was
feedback; i.e., the
implemented with estimated signal ()
controller has the form of
)
= KC ( d ) KD
m (,
was applied. When 1st order Q(s) filter was adopted,
although steady state error was not zero, the asymptotic
stability was guaranteed (Fig. 7a). The steady state error
can be understood by the fact that the corresponding C(s)
to the 1st order Q(s) filter is PD control input (recall
Remark2). On the other hand, when higher order Q(s) filter
was applied, the performance was significantly improved
(Fig. 7b). Moreover, the bottom rows in Fig. 7a, b show that
(43)
2387
R EFERENCES
[1] G. Pratt and M. Williamson, Series elastic actuators, in 1995
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS), vol. 1, Aug 1995, pp. 399406 vol.1.
[2] M. Zinn, O. Khatib, B. Roth, and J. K. Salisbury, Playing it safe
[human-friendly robots], Robotics & Automation Magazine, IEEE,
vol. 11, no. 2, pp. 1221, 2004.
[3] A. Albu-Schaffer, O. Eiberger, M. Grebenstein, S. Haddadin, C. Ott,
T. Wimbock, S. Wolf, and G. Hirzinger, Soft robotics, Robotics &
Automation Magazine, IEEE, vol. 15, no. 3, pp. 2030, 2008.
2388