Adaptive Control of Robot Manipulator

Sabyasachi Kundu
Control and Automation,Electrical Engineering Department

Delhi, India

Abstract—This document showcases the implementation of

adaptive control on a two-link Robot Manipulator. The control
schemes were simulated using the MATLAB/Simulink environ-
ment, and the corresponding performance analysis with suitable
plots has been presented.

I. SYSTEM MODEL
The dynamics of a two-link robot manipulator are repre-
sented as the Euler-Lagrange system given by:-
    
τ1 p1 + 2p3 c2 p2 + p3 c2 q̈1
=
τ2 p2 + p3 c2 p2 q̈2
| {z }
M (q)
  
−p3 s2 q̇2 −p3 s2 (q̇1 + q̇2 ) q̇1
+
p3 s2 q̇1 0 q̇2 (1)
| {z }
Vm (q,q̇)
  
fd1 0 q̇1
+
0 fd2 q̇2
| {z }
Fd Fig. 1. N-link Robot Manipulator

where q1 (t), q2 (t) denote the angular position (rad) of It is desired that the angular positions q1 (t), q2 (t), ..., qn (t)
the robot links and q(t) = [q1 (t), q2 (t)]. The matrices follow certain reference trajectory qd1 (t), qd2 (t), ..., qdn (t). It
M (q), Vm (q, q̇) and Fd denote the inertia, centripetal- Coriolis is assumed that qd , q̇d , q̈d ∈ L∞ . Such systems are known as
and friction terms. The constant robot parameters are given as Euler-Lagrange(E-L) systems and their general form is given
p1 = 3.473 kg, p2 = 0.196 kg −m2 , p3 = 0.242 kg −m2 , fd1 = by:-
5.3 Nm-sec, fd2 = 1.1 Nm-sec, measured using the encoders, M (q)q̈ + Vm (q, q̇)q̇ + F (q̇) + G(q) = τ (3)
c2 = cos(q2 ), s2 = sin(q2 ), τ1 and τ2 represent torque control
inputs in N. Assume q1 (t), q2 (t), q̇1 (t), q̇2 (t) are measurable where,
n×n
signals, however, note that the parameters in M, Vm , Fd are • M (q) ∈ R is called Inertia matrix.
n×1
unknown to the controller. • q(t) ∈ R is called generalized position vector.
• q̇(t) is called generalized velocity vector.
II. T RACKING C ONTROL O BJECTIVE n×n
• Vm (q, q̇) ∈ R is called centripetal-Coriolis matrix.
n×1
• F (q̇) ∈ R
 T is called the Friction Damping Matrix.
Design an adaptive controller τ1 τ2 for the Euler- n×1
• G(q) ∈ R is called the Gravity Vector.
Lagrange system given in (1) to track the desired trajectory n×1
• τ ∈R denotes the controlling torque.
given by:-
    Properties of the above system are:-
q (t) 0.5sin(t)
qd (t) = 1d = (2) 1) Inertia Matrix M(q) is Positive definite and is both upper
q2d (t) 2cos(t/4) and lower bounded.
Verify that Ṁ − 2Vm is skew-symmetric. Also, write the µ1 ∥ζ∥2 ≤ ζ T M (q)ζ ≤ µ2 ∥ζ∥2
expression for the regressor matrix.
where µ1 , µ2 > 0 and ζ ∈ Rn .
This is not valid for all E-L systems. It is only true for
III. ADAPTIVE TRACKING CONTROL OF N -LINK robot systems having robust joints.
ROBOT MANIPULATOR 2) Ṁ − 2Vm is a skew-symmetric matrix.
For an n-link robot arm manipulator is shown below (Fig-1). ∴ ζ T (Ṁ − 2Vm )ζ = 0 ∀ζ ∈ Rn
 3) Dynamics of E-L systems are Linear in Parameters(LIP).
Means (3) can be written as:- From (3), substituting value of M q̈, we have:-
M (q)ṙ = −Vm (q, q̇)q̇ − F (q̇) + τ − M (q)q̈d + M (q)αė (9)
M (q)q̈ + Vm (q, q̇)q̇ + F (q̇) + G(q) = τ (4)
| {z }
YΘ
Equation (9) represents open loop error system. Here G(q) =
n×p
where Y = Y (q̈, q̇, q) ∈ R is known regressor matrix, 0. Adding Vm r − Vm r to R.H.S of (9),
and Θ ∈ Rp×1 vector of unknowns. p is the no. of unknown
parameters in the system. M (q)ṙ = − Vm (q, q̇)q̇ − F (q̇) + τ − M (q)q̈d + M (q)αė
+ Vm r − Vm r
IV. CONTROLLER DESIGN OF TWO-LINK (10)
MANIPULATOR Rearranging the terms,
Objective 1. To design an  Adaptive controller such that M (q)ṙ =(−Vm (q, q̇)q̇ − F (q̇) − M (q)q̈d + M (q)αė
T (11)
angular position(rad) q(t) = q1 (t) q2 (t) of the robot link + Vm r) − Vm r + τ
tracks desired trajectory (2) with minimum error. The position
tracking error is given by:- ∴ M (q)ṙ = Y Θ + τ − Vm r (12)

e ≜ q(t) − qd (t) (5) Here Y Θ represents the LIP form of the term in brackets in
(11). For known dynamics we can write τ as:-
τ = −Y Θ − Kr, K ∈ Rn×n (13)
Here K is a positive definite and symmetric matrix. But since
plant dynamics are unknown to us, we replace Θ by Θ. b So
controlling torque τ is given by:-
τ = −Y Θ
b − Kr (14)
From (12) and (14) we have:-
e − Kr − Vm r
M (q)ṙ = Y Θ (15)
where Θe = Θ − Θ.
b
We can choose a positive definite, radially unbounded and
decrescent Lyapunov function:-
1 T 1 e T −1 e
V = r Mr + Θ Γ Θ (16)
2 2
Fig. 2. 2-link Robot Manipulator where Γ ∈ Rp×p .
To ensure that the filtered error dynamics have a stable
A. Filtered Tracking Error equilibrium point, the derivative of the Lyapunov function V
should be negative semi-definite. To achieve this, appropriate
The filtered tracking error r(t) can be given by:- adjustments to the estimated controller parameters (Θ)b must
r(t) ≜ ė + αe, α > 0, α ∈ Rn×n (6) be made at a suitable rate. However, the stability analysis
  equations leading to this conclusion have not been provided,
α 0 and only the final result is being considered.
Here n=2. So in our example we have, α = 1 where
0 α2 Now since we have V̇ to be negative semidefinite, we have
α1 , α2 > 0 r ∈ L∞ from (23). Further using Barbalet’s Lemma we can
do signal chasing and show that ṙ is bounded ⇒ r is uniformly
continuous. So it means that:-
B. Controller Expressions lim r(t) = 0 (17)
t→∞
From (6) we have, From (6) and (24), we can say that:-
ṙ = ë + αė (7) (e, ė) ∈ L∞ and lim (e, ė) = 0 (18)
t→∞

The rate at which controller parameters should change for a

M (q)ṙ = M (q)ë + M (q)αė (8) stable dynamics is given by:-

= M (q)(q̈ − q̈d (t)) + M (q)αė b˙ = ΓY T r

Θ (19)
  
C. Mathematical expression of Regressor Matrix 0 p3 (2 sin (q2 )q̇1 + sin (q2 )q̇2 )
−p3 (2 sin (q2 )q̇1 + sin (q2 )q̇2 ) 0
From the values of M (q), Vm (q, q̇) and F (q̇) given in (1), Here we can see that,
we will calculate Y . From (11) and (12), we have:-
(Ṁ − 2Vm ) = −(Ṁ − 2Vm )T
Y Θ = −Vm (q, q̇)q̇ − F (q̇) − M (q)q̈d + M (q)αė + Vm r (20) So, (Ṁ − 2Vm ) is a skew-symmetric matrix.
Hence all the assumptions that we have required to use the
From (1) and (16), we have
   n-link manipulator control technique got satisfied.
−p3 s2 q̇2 −p3 s2 (q̇1 + q̇2 ) q̇1
YΘ=− V. SIMULATION RESULTS & DISCUSSIONS
p3 s2 q̇1 0 q̇2
   In this section, the plots obtained for tracking error(e),
f 0 q̇1 control input(τ ), adaptive learning and pure feedback part of
− d1
0 fd2 q̇2 control input, and parameter estimation errors (Θ)
e are shown
  
p1 + 2p3 c2 p2 + p3 c2 q̈1d after simulating in MATLAB/Simulink.
− (21) The values of adaptation gain (Γ), filtered tracking error gain
p2 + p3 c2 p2 q̈2d
    (α), and
 pure feedback gain  (K) are takes as:
p + 2p3 c2 p2 + p3 c2 α1 0 ė1 5 0 0 0 1
+ 1
p2 + p3 c2 p2 0 α2 ė2 0 5 0 0 0     
  2 0 3 0
Γ = 0 0 3 5 0 ; α =
  
−p3 s2 q̇2 −p3 s2 (q̇1 + q̇2 ) r1   ;K=
+ 0 0 5 20 0  0 0.8 0 0.5
p3 s2 q̇1 0 r2
1 0 0 0 10
Expanding the terms in (20) and since they are in  
0
LIP(Linear in parameters) form, separating the manipulator 0
parameters after expansion we obtain Y and Θ as:-  
and initial guess for Θ(0) = 0
b 
  0
p1
   p2  0
y y12 y13 y14 y15
Y = 11
 
& Θ=  p3  (22)

y21 y22 y23 y24 y25 fd1 
fd2
where,
y11 =α1 ė1 − q̈d2
y12 =α2 ė2 − q̈d2
y13 =2c2 α1 ė1 − 2c2 q̈d1 + c2 α2 ė2 − c2 q̈d2 −
s2 q̇2 r1 + 2s2 q̇1 q̇2 − s2 q̇1 r2 − s2 q̇2 r2 + s2 q̇22 Fig. 3. Control Input
y14 = − q̇1
y15 =0 (23)
y21 =0
y22 =α1 ė1 − q̈d1 + α2 ė2 − q̈d2
y23 =c2 α1 ė1 − c2 q̈d1 + s2 q̇1 r1 − s2 q̇12
y24 =0
y25 = − q̇2
Here Y is the regressor matrix and Θ is a vector of manipulator Fig. 4. Adaptive and pure feedback part of Control Input at joint 1
parameters that is unknown to the controller. So we replace Θ
by Θ.
b

D. Skew-symmetricity of (Ṁ − 2Vm ) matrix:

To prove
 Ṁ − 2Vm is a skew-symmetric  matrix,
−2p3 sin (q2 )q̇2 −p3 sin (q2 )q̇2
Ṁ =
−p3 sin (q2 )q̇2
 0 
−2p3 sin (q2 )q̇2 −2p3 sin (q2 )(q̇1 + q̇2 )
2Vm =
−2p3 sin (q2 )q̇1 0
Now,
Ṁ − 2Vm = Fig. 5. Adaptive and pure feedback part of Control Input at joint 2
 Fig. 6. Position error of joint 1 (∆q1 )

Fig. 7. Position error of joint 2 (∆q2 )

Fig. 8. Parameter estimation error (Θ)

e

With the tuned gains, from the plots, we can see that,
• Fig 3. shows the control input to the system, where
control input 1 (τ1 ) is the input applied to joint 1 and
control input 2 (τ2 ) is the input applied to joint 2. Initially,
the control input gets deviated from the required value
but within 10 seconds it settles down to the actual input
required.
• Fig 4 shows the adaptive and pure feedback part of the
control input 1 (τ1 ). The adaptive part of τ1 becomes the
actual input required to joint 1 and the pure feedback
part of τ1 becomes zero as the error goes to zero after
15 seconds.
• Fig 5 shows the adaptive and pure feedback part of the
control input 1 (τ2 ). Here also we can see that the adaptive
part of τ2 becomes the actual input required to joint 1 and
the pure feedback part of τ2 becomes zero as the error
goes to zero after 15 seconds.
• Fig 6 shows the position error (δq1 = q1 −qd1 ) of joint 1.
Initially, the position error of joint 1 (∆q1 ) gets deviated
a bit but it becomes zero within 15 seconds.
• Fig 7 shows the position error (δq2 = q2 − qd2 ) of joint
2. Initially, the position error of joint 2 (∆q2 ) also gets
deviated a bit but it becomes zero within 15 seconds.
• Fig 8 shows the parameter estimation error of the system.
It also converges to zero within 15 seconds.