Professional Documents
Culture Documents
### Dynamic-Identification-Of-Robots-With-Power-Model
### Dynamic-Identification-Of-Robots-With-Power-Model
Maxime GAUTIER
(4)
which have no effect on the dynamic model and by JqTrdt = H(q,q)(tb)- H(q,q)(ta)+ jqTrfdt =Ah X (9)
regrouping some others in linear relations. Symbolic and ta ta
numerical solutions have been proposed for any open or Ah is the (IxNp) regressor row matrix defmed from the
closed chain manipulator [7, 8, 91 to get a minimal energy function row matrix h(q, q) :
dynamic model :
Ah(% 4) == h(q7 q)(tb ) - h(q, q>(ta)
The energy model as the power model is a scalar equation
r = D ( q , q, q ) X = CD:iXi (5) whose symbolic equations are easier to calculate and
i=l,Np
manipulate than the vector equations of the dynamic
The ith column D:,i of D is given as following : model [ 101.
4. THE IDENTIFICATION METHOD
Generally, ordinary least squares (OLS) technique is used
From Newton-Euler algorithm it can be seen that the to estimate the minimum dynamic parameters solving an
torque of motor j is function of the inertial parameters of overdetermined linear system obtained from a sampling of
links j, j+l,.., n., thus the matrix D has an upper triangular the dynamic model (Eq.7) or the energy model (Eq.9)
form, such that : along a trajectory (9, q or q) [ l , 3,4,5,6] :
- -
- rl - -D1,1 DIJ ... D'," XI
r2 0 D2,2 ... D2.n x2
. -- . (7) The L.S. solution ? minimizes the 2 norm llp1I2 of the
Dn-l,n-l Dn-l,n xn-l vector of errors p. The unicity of ?depends on the rank
rn-1 0 ...
... of the observation matrix W.
-rn - 0 0 Dn,n
__ xn - W numerical rank deficiency can come from two origins :
the - structural rank deficiency which stands for any samples
the of (9, q or q) in W.
the This is the structural parameters identifiability problem
i
parameters of link which are contained i n the base which is solved using base parameters.
parameters of the robot. - data rank deficiency due to a bad choice of noisy
D',J is a row vector, containing the coefficients of the (q, q or q) samples in W. This is the problem of optimal
parameters XJ in the equation of the torque of motor i. measurement strategies which is solved using closed loop
identification to track exciting trajectories.
3. THE POWER MODEL In order to decrease the sensitivity of the L.S. solution of
In order to eliminate any derivation of velocity in the system (Eq.10) to errors in Y and W, the condition
identification process, a model based on the energy number of the observation matrix W, Cond(W), must be
theorem has been proposed [6,11]. This model can be close to one before computing 2 . Exciting trajectories
obtained in a differential form calculating the power of the can be obtained by non linear optimization of a criterion
system with the Lagrange equation (Eq. 1) : function of the condition number of W, under constraints
of the equation of an interpolator and the joints positions
and velocities limits [16]. In the following, we suppose
that this stage has been reached, that is W is a (rxNp) fill
1923
rank and well conditioned matrix obtained by tracking an random matrices [l]. Then it is essential to filter data in Y
exciting trajectory [ 121, which needs a closed loop control and W, before computing the L.S. solution.
of the robot. 5.2. Filtering the dynamic model
Standard deviations 02, are estimated using classical and
Samples of VT and q at rate w, allow to calculate samples
simple results from statistics, considering the matrix W to
of (Eq.7) at times ti, i=l ,...,ne, to get the (Ixr)
be a deterministic one, and p to be a zero mean additive measurement vector Y(r) and the (rxNp) observation
independent noise, with standard deviation up such that :
matrix W(q, q, q) , with r=nxn,>Np, where matrices have
T 2
c,, = E(P P) = 0, 1, been arranged to regroup the data of a same joint all
where E is the expectation operator. together :
An unbiased estimation of CY,, is used
-2
!/Y - w kif
( 3 = (1 1)
r-Np r .-I r -.
The covariance matrix of the estimation error and standard
deviations can be calculated by :
L J L A
In fact this is not true because the vector of measurements Ywfpd = wwfp (q fq 3 q fq ,q fq x -k Pwfpd (19)
Yfpdis obtained from the concatenation of n
with :
measurements vectors of the n motor torques Y J , with
Ywfpd = G Yfpd Wwfpd = G wfpd &f$d = G Pfpd
9 9
r .=
J [DJ.J
"
...
1
DJ," @J
, dim(OJ) = (NpJJ)
X"
One point is how to choose the sample times ta(i). In order
. 2 = /bjpd
0; - @&cl 6Jll2
, with :
to avoid offset in y(i) due to constant perturbation in
(qfqT r), tb(i)-ta(i) must be bounded. In [12, 151,
rJ - NpJ
sampling times are optimized to get a well conditioned
observation matrix W (Eq.20). This formulation needs a
lower resampling of the energy function h at times ta(i),
tb(i). Then the function h calculated at the acquisition rate
w, must be lowpass filtered with @(s) in order to avoid
aliasing. So it is easier and natural to define the sampling
times from the decimate procedure with a ratio nd, which
results in choosing tb(i) and t,(i) with a constant value
tb(i)-ta(i)=nd*2*n/w,. As a result, the parallel decimate
procedure fpd must be applied to (Eq. 20,21) as following:
1925
This is equivalent to parallel filter and decimate the
differential expression of the energy model (Eq.8), which
is the power model :
Pfpd =('$fc
r)fpd = S hfpd x = dhfpj x
The derivative dhfpd is calculated without phase shift
(23)
D:,6 = DLMYz =
-(2ql + q2) sinq2 - '$2 (241 + '$2) cosq2
ql2 cosq2 - q l sinq2
1
using a central difference algorithm.
Filtering the power model clearly defines the choice of the
sampling times t,(i) of the energy model which depends on
nd = 10, while keeping the main advantage of this model -power model (Eq. 8) :
which is the simplicity of the energy functions h.
6. VALIDATION
6.1. Description of the robot
The comparison is carried out on a 2 joints planar direct
drive prototype robot (Fig. l), without gravity effect. The
description of the geometry of the robot uses the modified
Denavit and Hartenberg notation [6, 101.
Y2
1926
Table 1 : Comparison of the systematic errors of the providing to use the same filters and the same information.
filtering process. Our conclusion is that the filtered power model is very
attractive for it's simplicity and accuracy and should be
I Dynamic 11 Dynamic 11 Power I used instead of the filtered dynamic model.
REFERENCES
[ 11 M. Gautier, P.P. Restrepo, W. Khalil, "Identification
of an industrial robot using filtered dynamic model ' I ,
Proc. 3rd ECC, Rome, Sept. 1995, pp. 2380-2385.
[2] C. Canudas de Wit, A. Aubin, "Parameters
identification of robots manipulators via sequential
hybrid estimation algorithms", IFAC Congress, Tallin,
1990, pp. 178-183.
[3] M. Prufer, C. Schmidt, F. Wahl, "Identification of
Robot Dynamics with Differential and Integral
Models : a Comparison", IEEE Conf. on Robotics and
Automation, San Diego, 1994, pp. 340-345.
[4] C.H. An, C.G. Atkeson, J.M. Hollerbach, " Estimation
6.3. Experimental comparison of inertial parameters of rigid body links of
The 2 models and the same filtering process as defined for manipulators". Proc. 24th Conf. on Decision and
simulation have been used to calculate the L.S estimation control, 1985, pp. 990-995.
[5] P.K. Khosla, T. Kanade, "Parameter identification of
X of the dynamic parameters of the prototype robot. Four robot dynamics ", 24th CDC, 1985, pp. 1754-1760.
independent realizations of the trajectory have been used [6] Gautier, W. Khalil, "On the identification of the
to get rJ = 3200 for the dynamic model and r=3200 for inertial parameters of robots", 27th CDC, 1988, pp.
the power model. 2264-2269.
[7] M. Gautier, W. Khalil, " Direct calculation of
Table 2 : Comparison of experimental estimation minimum inertial parameters of serial robots", IEEE
Trans. on Robotics and Automation, Vol. 6, No 3,
I Dynamic 11 Dynamic 11 Power I 1990, pp. 368-373.
[SI H. Mayeda, K. Yoshida, K. Osuka, "Base parameters
of manipulator dynamics", Proc. IEEE Conf. on
Robotics and Automation, 1988, pp. 1367-1373.
191 M. Gautier, "Numerical calculation of the base inertial
parameters", Journal of Robotics Systems, Vol. 8, No
4, 1991, pp. 485-506.
[ 101 W. Khalil, "SYMORO+ : A system for generating the
symbolic models of robots", SYROCO, Capri, Sept.
1994.
111 S.Y. Sheu, M. Walker, "Estimating the essential
parameter space of the robot manipulators dynamics",
Proc. 28 th CDC, 1989, pp. 2135-2140.
121 M. Gautier, "Optimal motion planning for robot's
inertial parameters identification", 3 1st CDC, Tucson,
Results given in Tab.2 show that the new filtered power 1992, pp.70-73.
model with OLS gives the same accuracy than that 131 P. P. Restrepo, M. Gautier, "Calibration of drive
obtained with the filtered dynamic model and WLS, using chain of robot joints", 4th IEEE Conf. on Control
exactly the same amount of information and the same Applications, Albany, 1995, pp. 526-53 1.
filters (wfp,wf,), that is to say the 2 methods have the same [ 141 P.R. Belanger, X. Meng, P. Dobrovolny, "Estimation
convergent speed. So, it is possible to take advantage of a of angular velocity and acceleration from equally-
very simple identification model which is less sensitive to spaced shaft encoder measurements", IEEE Conf. on
the use of exciting trajectories compared to the energy Robotics and Automation, Nice, 1992.
model which is it's integral formulation [15]. [15] M. Gautier, W. Khalil, C. Presse, P.P. Restrepo,
7. CONCLUSION "Experimental Identification of dynamic Parameters of
robot", 4th SYROCO, Capri, Italy, Sept. 1994, pp.
A new formulation of the energy model, based on a 625 -63 0.
filtered power model, is proposed. Theoretical, simulation [16] C. Presse, M. Gautier, "New Criteria of exciting
and experimental studies show that OLS estimation using trajectories for Robot Identification", Proc. IEEE
the filtered power model gives close estimations, accuracy Conf. on Robotics and Automation, Atlanta, 1993, pp.
and convergent speed as those obtained with a new WLS 907-9 12.
estimation using the classical filtered dynamic model,
1927