Model Identification for Robot

Manipulators using Regressor-Free

Adaptive Control

Rameez Hayat1,2

Martin Buss1,2

1 Institute for Advanced Study

Technical University of Munich

2 Chair of Automatic Control Engineering

Technical University of Munich

UKACC International Conference on Control, 09/01/2016

Presentation Overview

Introduction
Related Work
Contribution
Results
Conclusion

Introduction

Related

Contribution

Results

Conclusion

Feedback Linearization
Manipulator Dynamics:
M (q)
q +C(q, q)
q+G(q)+F

q = .

Introduction

Related

Contribution

Results

Conclusion

Feedback Linearization
Manipulator Dynamics:
M (q)
q +C(q, q)
q+G(q)+F

q = .
Let:
= M (q)
a + N (q, q)
+ F q,

where, a
= qd K v e K p e,
N (q, q)
= C(q, q)
q+G(q)

Introduction

Related

Contribution

Results

Conclusion

Feedback Linearization
Manipulator Dynamics:
M (q)
q +C(q, q)
q+G(q)+F

q = .
Let:
= M (q)
a + N (q, q)
+ F q,

where, a
= qd K v e K p e,
N (q, q)
= C(q, q)
q+G(q)

Introduction

Related

Contribution

Results

Conclusion

Feedback Linearization
Manipulator Dynamics:
M (q)
q +C(q, q)
q+G(q)+F

q = .
Let:
= M (q)
a + N (q, q)
+ F q,

where, a
= qd K v e K p e,
N (q, q)
= C(q, q)
q+G(q)

PD
e
+ K v e + K p e = 0.

Introduction

Related

Contribution

Results

Conclusion

Motivation
System Dynamics estimation:
(
+ F q,
=M
q d K v e K p e) + N

Introduction

Related

Contribution

Results

Conclusion

Motivation
System Dynamics estimation:
(
+ F q,
=M
q d K v e K p e) + N

M (q)
q + C(q, q)
q + G(q) + F q =

Introduction

Related

Contribution

Results

Conclusion

Motivation
System Dynamics estimation:
(
+ F q,
=M
q d K v e K p e) + N

M (q)
q + C(q, q)
q + G(q) + F q =
1 (M
q + N
+ F q),
e
+ K v e + K p e = M

=M
M, N
=N
N and F = F F .
where, M

Introduction

Related

Contribution

Results

Conclusion

Motivation
System Dynamics estimation:
(
+ F q,
=M
q d K v e K p e) + N

1 (M
q + N
+ F q),
e
+ K v e + K p e = M

=M
M, N
=N
N and F = F F .
where, M
Goal
Design an adaptive controller that removes any mismatch to
achieve better feedback linearization.

Introduction

Related

Contribution

Results

Conclusion

Related Work
Siciliano et al.
Regressor-based adaptive controller:
Let: Y (q, q,
q)p = M (q)
q + N (q, q)
+ F q.

e
+ K v e + K p e = M

Introduction

Related

Contribution

Y (q, q,
q)p,

Results

Conclusion

Related Work
Siciliano et al.
Regressor-based adaptive controller:
Let: Y (q, q,
q)p = M (q)
q + N (q, q)
+ F q.

e
+ K v e + K p e = M

Y (q, q,
q)p,

State space representation:

x = Ax BM 1 Y (q, q,
q)p,


 

0
0
In
R2n2n , B =
R2nn
A=
K p K v
In
and
x = [eT e T ]T

Introduction

Related

Contribution

Results

Conclusion

Related Work
1
1
V (x, p)
= xT P 1 x + pT P 2 p,

2
2

Introduction

Related

Contribution

Results

Conclusion

Related Work
1
1
V (x, p)
= xT P 1 x + pT P 2 p,

2
2
If we put:
1 Y )T B T P 1 x.
p = P 1
2 (M

Introduction

Related

Contribution

Results

Conclusion

Related Work
1
1
V (x, p)
= xT P 1 x + pT P 2 p,

2
2
If we put:
1 Y )T B T P 1 x.
p = P 1
2 (M
Then

1
V = xT Qx 6 0
2

Introduction

Related

Contribution

Results

Conclusion

Related Work
1
1
V (x, p)
= xT P 1 x + pT P 2 p,

2
2
If we put:
1 Y )T B T P 1 x.
p = P 1
2 (M
Then

1
V = xT Qx 6 0
2

Drawbacks
 Require joint acceleration
 Calculating inverse of Inertia Matrix

Introduction

Related

Contribution

Results

Conclusion

Related Work
Slotine et al. [1987]
Slotine and Lis adaptive controller:
Slotine et al. reconfigured the system dynamics by considering
s = e + e, where > 0 Rnn . Using this new formulation:
M s + N + F s + M v + F v = ,
v = q d e.

Introduction

Related

Contribution

Results

Conclusion

Related Work
Slotine et al. [1987]
Slotine and Lis adaptive controller:
Slotine et al. reconfigured the system dynamics by considering
s = e + e, where > 0 Rnn . Using this new formulation:
M s + N + F s + M v + F v = ,
v = q d e. Considering:
v + F v + N
K D s = Y (q, q,
=M
v, v)
p,

where K D is a positive definite matrix,

Introduction

Related

Contribution

Results

Conclusion

Related Work
Slotine et al. [1987]
Slotine and Lis adaptive controller:
Slotine et al. reconfigured the system dynamics by considering
s = e + e, where > 0 Rnn . Using this new formulation:
M s + N + F s + M v + F v = ,
v = q d e. Considering:
v + F v + N
K D s = Y (q, q,
=M
v, v)
p,

where K D is a positive definite matrix, we get:

M s + F s + K D s = Y (q, q,
v, v)
p.

Introduction

Related

Contribution

Results

Conclusion

Related Work
Slotine et al. [1987]
Slotine and Lis adaptive controller:
Slotine et al. reconfigured the system dynamics by considering
s = e + e, where > 0 Rnn . Using this new formulation:
M s + N + F s + M v + F v = ,
v = q d e. Considering:
v + F v + N
K D s = Y (q, q,
=M
v, v)
p,

where K D is a positive definite matrix, we get:

M s + F s + K D s = Y (q, q,
v, v)
p.

Following the same steps as before:

v, v)s.

p = P 1 Y T (q, q,
2

Introduction

Related

Contribution

Results

Conclusion

Related Work
Kai et al. [2013]
Regressor-free adaptive controller:
M 1 Y (q, q,
q)p = W T Z,
where W is a weighting matrix and Z is a matrix of basis
functions.
1
1
TM QW
)
V = xT P x + tr(W
2
2
The control law would be:
= Q1 ZxT P B.
W

Introduction

Related

Contribution

Results

Conclusion

Proposed Adaptive Technique

Let = M (q)
q + N (q, q)
+ F q q + a , where a represents
disturbances and unmodeled parameters. Then:
+ qd K v e K p e.
=

Introduction

Related

Contribution

Results

Conclusion

Proposed Adaptive Technique

Let = M (q)
q + N (q, q)
+ F q q + a , where a represents
disturbances and unmodeled parameters. Then:
+ qd K v e K p e.
=
Using the above equation:
,
e
+ K v e + K p e =

Introduction

Related

Contribution

Results

Conclusion

Proposed Adaptive Technique

Let = M (q)
q + N (q, q)
+ F q q + a , where a represents
disturbances and unmodeled parameters. Then:
+ qd K v e K p e.
=
Using the above equation:
,
e
+ K v e + K p e =
converge to and thus e 0, we introduce the
To make
following differential equation:
m
X
i=0

ai

di
= f (e, e),

dti

is
and am = 1
where the system mismatch
Introduction

Related

Contribution

Results

Conclusion

Proposed Adaptive Technique

at every time instance, is
To evaluate the value of
approximated by a polynomial of order m 1 using Taylor series.
The update law becomes:
m1

X di X

dm
ai i + f (e, e),
=

dtm
dt
i=0

where X = e + K v e + K p e. Assuming Kx = f (e, e),

if
T
1
K = P 2 B P 1 , then:
Convergence
T Q2
6 0.
V = xT Q1 x

Introduction

Related

Contribution

Results

Conclusion

10

Parameter Identification

Using least squares:

Y p = B,
p = (Y T Y)1 Y T B,
where,
Y = [Y T1 Y T2 ... Y TN ]T ,
B = [ T1 T2 ... TN ]T ,
N is the total number of sampled data points.

Introduction

Related

Contribution

Results

Conclusion

11

Manipulator Dynamics
A 2-Degree-of-Freedom robot manipulator is considered for
simulation as well as experiments.
Manipulator Dynamics:


n11 sin q2 q22 n12 sin q2 q1 q2
N=
,
y
n21 sin q2 q12
l
0

y2

F =
M=

f11 0
0 f22

y1

l1
q1
mass m1 , m2

Related

q1

x1

m11 + m11 cos q2 m12 + m12 cos q2

m21 + m21 cos q2

m22

Introduction

x2

Contribution

x0

Results

Conclusion

12

Experimental Results
0.45

Trajectory
Regressor
Slotine/Li
FAT-Based
Proposed

0.4
0.35
0.3

y[m]

0.25
0.2
0.15
0.1
0.05
0
0

0.1

0.2

0.3

0.4

0.5

x[m]

Experimental results: position of end effectors starting from 0.5b

x + 0b
ym

Introduction

Related

Contribution

Results

Conclusion

13

Simulation results

-0.5

-20

Y1 (q, q,
v, v)
p

0.25

0.5
0.75
Time [s]

Y1 (q, q,
v, v)p

0.25

0.5

0.75
Time [s]

1.25

Error in joint angles for the first

rotation.
10

20
10
0

-10
-20

-1
0

1.25

First joint: Y1 is the 1st row of

regressor matrix

Torque [Nm]

Joint error [rad]

e(1)
e(2)

0.5

Y2 (q, q,
v, v)
p

0.25

0.5
0.75
Time [s]

Y2 (q, q,
v, v)p

Second joint: Y2 is the 2nd row of

regressor matrix

1.25

Joint error [rad/s]

Torque [Nm]

1
20

e(1)

e(2)

5
0
-5

-10
0

0.25

0.5

0.75
Time [s]

1.25

Derivative of error in joint angles for

the first rotation.

Results of Slotine and Lis adaptive control for = 4 rad/s, the

estimated model dynamics do not follow the original model.
Introduction

Related

Contribution

Results

Conclusion

14

Simulation results

Joint error [rad]

Torque [Nm]

1
10

e(1)
e(2)

0.5
0

-0.5

-10

(1) + q(1)

0.25

0.5
0.75
Time [s]

(1)
+ q(1)

-1
0

1.25

0.75
Time [s]

1.25

5
0
-5
(2) + q(2)

0.25

0.5
0.75
Time [s]

(2)
+ q(2)

1.25

Joint error [rad/s]

10

10
Torque [Nm]

0.5

Error in joint angles.

First joint angle

-10

0.25

e(1)

e(2)

5
0
-5

-10
0

0.25

0.5

0.75
Time [s]

1.25

Derivative of error in joint angles.

Second joint angle

Results of proposed adaptive control for = 4 rad/s. Taking a

very low time constant for the update law will confirm the
modeling error to approach zero.
Introduction

Related

Contribution

Results

Conclusion

15

Parameter estimation
Calculated, simulation and experimental estimation of parameters for 2-DOF
manipulator.

Var.
m11

m11
m12

m12
n11
n12
f11
m21

m21
m22
n21
f22

Introduction

Related

Cal.
0.442
0.028
0.009
0.014
0.014
0.030
0.001
0.009
0.014
0.222
0.014
0.001

Sim.
0.442
0.027
0.008
0.015
0.014
0.028
0.004
0.009
0.015
0.222
0.013
0.006

Contribution

Exp.
0.590
0.027
0.006
0.010
0.216
0.041
0.001
0.005
0.009
0.282
0.282
0.001

Results

Conclusion

16

Conclusion






Design of an adaptive controller that requires less tuning

parameters.
Removal of mismatch during feedback linearization.
Parameters identification for a manipulator.
No prior knowledge required about system parameters.
Low computational cost.

Introduction

Related

Contribution

Results

Conclusion

17