ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 1

Motion control of a Pick-and-Place mobile manipulator

Chandrasekar Elankannan and Kishore Nedumaran

Abstract— The purpose is to control a mobile robot manip- high cost to know the regressive matrix of the robot dynamic
ulator to execute the pick-and-place operation by efficiently model. A way to overcome this cost is through applying
reaching the desired target location. A 6-DOF mobile robot neural networks. Single layer neural networks replace the re-
manipulator with omnidirectional mobile base is chosen for
the study. The mobile robot is entitled to reach the target gressive matrix with a vector containing activation functions
position when unanticipated forces act on the end effector, that are not limited to single robot configuration [2]. Hence
which requires control of the manipulator as well as the from the above set of controllers we propose to choose PD
mobile base. The control strategies utilized will be Proportional- with gravity compensation and Classic Adaptive controller to
Derivative control with gravity compensation, and Classical evaluate the performance of a 6DOF mobile manipulator in
Adaptive Controller for the combined control of the mobile
robot manipulator to reach the target position. In order to handling unexpected end effector forces using metrics such
assess the controllers, the performance of the PD controller as tracking error, convergence rate, and robustness.
and Classical Adaptive controller will be compared using the
performance metrics such as tracking error, rate of error
II. M ATERIALS AND M ETHODS
convergence and robustness to uncertain external forces.
The mobile robot manipulator system considered for the
study is shown in the Fig. 1.
I. I NTRODUCTION
The field of robotics has been through numerous advance-
ments ever since GM motors introduced “UNIMATE”, the
first industrial robotic arm in the 1950s. Today, they are an
integral part of many industries to perform various tasks such
as welding, pick and place, painting etc. Robotic manipula-
tors are widely adopted in industries to increase productivity
and precision, while minimizing human error. One of the
main advantages of having a manipulator is its ability to work
in critical conditions where it is risky or fatiguing for humans
while having the advantage to easily reprogram and modify
for new operations. These advanced robotic manipulators
mounted on a mobile platform can be termed as mobile
robot manipulators, combine the advantages of mobile plat-
forms and robotic arms and reduce their drawbacks. The
mobile platform extends the workspace of the arm and wide
opens a whole new range of applications such as mining,
construction, home assistants, space exploration, search and
rescue, and military [1]. In order for these manipulators to
be versatile and useful, it needs the ability to handle various
Fig. 1. 6 DOF Mobile Manipulator System
objects and unanticipated forces while interacting with the
world.
Numerous control algorithms have been developed for
controlling these manipulators such as PD, PD with gravity A. Software and Hardware
compensation, PID, adaptive, fuzzy, and neural network In this project, the steps starting from dynamic modelling
controllers. PD and PID controllers are still widely used to controller implementation is programmed and simulated
in industries primarily due to their simplicity and relatively in MATLAB R2018 a. The 6-DOF kinematic chain of the
satisfactory performances. For horizontal planar robots PD mobile manipulator is extracted (explained in detail in II-
controllers have asymptotic stabilization, but for robots with B.) and modelled as one dimensional link chain to alleviate
rotational DOF tend to have permanent tracking error. Even computation complexity.
though it is simpler to implement, analyzing its stability
can be complex. Fuzzy PID controller can improve the
performance of a linear PID controller, but is more complex B. Kinematic-chain diagram
to implement and analyze its stability. Classical adaptive con- The kinematic chain of the mobile manipulator is devel-
troller enables asymptotic tracking error but has an associated oped according to the Denavit-Hartenberg(DH) procedure.
ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 2

Following conditions are to be satisfied, for employing the • The Y -axis shall be oriented according to the right hand
DH method for kinematic and dynamic modeling: rule
• The robot has to be considered as a manipulator with The DH parameters for the modified kinematic chain
chain of ‘n’ links of the mobile robot manipulator was used to compute the
• The ‘n’ link robot shall have ‘n − 1’ joints with each Manipulator Jacobian J. The DH parameter is tabulated in
of the joints having one DOF. the table shown below.
• Every joint DOF shall be either prismatic or revo-
TABLE I
lute.Thus the joint variable, qi shall either be angle of
DH PARAMETERS
rotation θi , in case of a revolute joint and displacement
of the joint, and di , if the joint is revolute. Joint αi θi ai di
( π
θi if joint i is revolute 1 2
0 0 d1
qi = (1)
di if joint i is prismatic 2 π π
0 d2
2 2

In case of the considered mobile robot manipulator, its 3 0 θ1 L1 0

holonomic mobile base has three states, i.e, the displacement π
in X direction, Y direction and rotation in Z direction (ϕ). 4 2
θ2 L2 L3
These states are considered as the DOFs for the omnidi- 5 0 θ3 L4 0
rectional mobile base.Thus the mobile base is considered
as a PPR serial link manipulator connected in series with 6 0 θ4 L5 0
the robot arm, which is considered as a 3R serial link
manipulator.
The mobile robot manipulator was modified as serial C. Kinematic modelling
link manipulator with the kinematic chain as shown in Fig
2 below[3].. The mobile robot’s freedom for displacement The kinematic equation that relates the joint velocities of
in X and Y directions are considered as prismatic joints the modified kinematic chain of Mobile manipulator with its
and the rotational degree of freedom of the mobile base is end effector velocities is given as [4] [5]:
considered as a revolute joint. This modified form of mobile
manipulator, the first two links L1 and L2 are considered as ẋ = J (q)q̇ (2)
virtual links with zero mass.Thus, the mobile manipulator where, J q is the manipulator Jacobian
has 6 DOFs in total. The DH joint frames are considered as q̇, the vector of robot joint velocities
ẋ, the vector of end effector velocities
The manipulator Jacobian matrix relating ẋ and q̇ is of the
following form:
 
Jv 1 Jv 2 Jv 3 Jv 4 Jv 5 Jv 6
J= (3)
Jω 1 Jω 2 Jω 3 Jω 4 Jω 5 Jω 6
where,
 h i
Ri0 −1 · 0 0 1 T if joint i is prismatic
Jv i = h iT
R0 · 0 0 1 × di if joint i is revolute
i −1 n
(4)

(
0 if joint i is prismatic
Jω i = h i (5)
Ri0 −1 · 0 0 1 T
if joint i is revolute

Ri0 −1 = Rotation matrix for rotation from 0th to i − 1th

joint
d0n = Displacement vector for displacement from nth joint
to 0th joint
Fig. 2. Kinematic chain miagram of the mobile manipulator in the form
of serial link manipulator d0i −1 = Displacement vector for displacement from i − 1th
joint to ith joint
shown in Fig 2 based on the following four DH rules:
0
• All the joints shall have their Z-axis oriented towards
The terms of the form Rm and d0m can be extracted from
the direction of joint actuation or rotation. the coordinate transformation matrix H given as :
 0
Rm d0m

• The X-axis of the joint i shall be perpendicular to the 0
Hm = (6)
Z -axis of ith and i − 1th joint 0 1
ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 3

D. Dynamic modelling And,in this study, the regressor matrix Y is computed using
The dynamic model for any ‘n’ link serial manipulator is the pseudo-inverse of vector θ, as Y = τ · θ† instead of
of the form [5]: using equation(11), to increase the speed of computation.
Note:The parameters of the first two virtual links are not
M(q)q̈ + C(q, q̇)q̇ + g(q) = τ (7) considered as they are assumed to have zero mass.
where,
When external force is acting on the mobile manipulator
τ ∈ Rn is the control torque and forces vector
system, the dynamics have to be modified as:
q ∈ Rn is the vector of joint translations and rotation angles
q̇ ∈ Rn is the joint velocity vector
q̈ ∈ Rn is the vector of joint acceleration M(q)q̈ + C(q, q̇)q̇ + g(q) − J T F = τ (12)
M (q) ∈ Rnxn is the inertial matrix
where F ∈ R3x1 is a vector of force acting on the end
C(q, q̇)∈Rnxn is the matrix of Coriolis and Centrifugal force
effector.
components
g(q) is the vector representing the gravitational forces
In this study, following assumptions and considerations E. PD controller with Gravity Compensation Design
are made, while computing the dynamic model in order to The control law of the PD controller for controlling the
reduce the computational complexity. dynamic model is given as [6] [5] :
• The effect of viscous friction in the dynamics is negli-
gible τ = Kp · q̃ + Kd · q̃˙ (13)
1
• The inertia values, Ii are one dimensional,i.e Ii ∈ R
• Motor mass is neglected and only masses of the links
Where
are considered Center of mass of all the links Li are τ is the torque as a control input
considered to be at the midpoint of the link, Kp ∈ Rn×n is the positive definite matrix for proportional
i.e Lci = Li /2, with the exception of COM of link 3 gain for ’n’ DOFs
considered at the joint 3, as the joint 3 is at the COM Kp ∈ Rn×n is the positive definite matrix for derivative
of the mobile base. gain for ’n’ DOFs
The dynamic model equation (7) is used in this study
As the mobile manipulator is vertically erect, the gravita-
to control the joint space of the robot. The end effector
tional forces might add to the tracking error.Thus the effect of
position is ultimately controlled by converting task space
gravitational forces is compensated by the modified control
configuration i.e the end effector positions to joint space
law of the PD controller as shown in eqn (14).
using inverse kinematics as shown below:
qk = qk−1 + αk J −1 (qk−1 ) (xd − f (qk−1 )) , k = 1, 2, . . . τ = Kp · q̃ + Kd · q̃˙ + g(q) (14)
(8)
where αk > 0 is the step size for the iterations of the where g(q) ∈ Rn is the vector of gravitational forces
equation (12) , till the desired error between the xd and In this project, the Kp and Kd matrices are assumed to
x = f (qk−1 ) is achieved. be a positive diagonal matrix multiplied by a scalar term as
Though the dynamic model explained above in the equa- shown below:
tion (7) can be used for PD joint controllers, the classic Kp = k · I n×n (15)
adaptive controllers need the vector of uncertain parameters
θ ∈ Rm , which have every components as a function of
uncertain link mass, length and inertia. Kd = p · I n×n (16)
For any q,q̇ ∈ Rm , θ ∈ Rm is related with the control force
Where, k ∈ R and p ∈ R
vector τ ,as shown below:
The control schematics of the PD controller with gravity
τ =Yθ (9)
compensation is shown in Fig.3 below.
where, Y ∈ Rnxm is the regressor matrix,whose matrix
elements are given by:
∂τi
Yij = (10)
∂θj
I is the row number of the matrix Y and vector τ , J is the
row number of the vector θj
In this study, 8 uncertain parameters are considered in the
τ vector,given as shown below, as utilized in [1]
Fig. 3. Control scheme of PD + G controller [6]
 T
θ = m1 m2 m3 m4 I1 I2 I3 I4 (11)
ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 4

F. Classic Adaptive Controller Design

The control law of the classic adaptive controller is given
by [6]:

aq = q̈d − Kd (q̇ − q̇d ) − Kp (q − qd ) (17)

where, aq is the control input in terms of joint acceleration

Kd is the derivative gain of the controller

Fig. 4. Control scheme of Classical adaptive controller [6]
Kp is the proportional gain of the controller

ė = Ae + BΦθ̃ (18) G. Simulation procedure

The performance of the PD and classic adaptive control
I n×n
 
0 algorithms are measured in terms of tracking error ,error
where, A = , is a Hurwitz Matrix
−Kp −Kd convergence rate and robustness to uncertain external forces
  acting on the end effector due to the mass of the object
0 picked. The PD and adaptive control algorithms were im-
B= , Φ = M̂ −1 Y (q, q̇, q̈)
I n×n plemented in the dynamic model as shown in the equation
(7) ,when there are no uncertain external forces acting in the
end effector or, in other words no object picked by the end
 
q̃
e= , the error vector effector. Whereas, in case of external uncertain forces, the
q̃˙ algorithms were executed using the dynamic model as shown
in equation(12).

and III. R ESULTS AND D ISCUSSION

θ̃ = θ̂ − θ (19) In this section, the results obtained as plots from the
simulation are discussed namely:
For every timestep with step size α , the θ̂ is updated as
• Joint Torque/Force vs time plots
follows:
˙ • Joint configuration error vs time plots
θ̂ = α · θ̂ (20) • L2 norm of end effector position error, ||xd − x|| vs
˙ time step plots
where θ̂ is given as:
These plots were utilized to compare the effect of propor-
˙
θ̂ = −Γ−1 ΦT B T P e (21) tional gain and unknown external force on the performance
measures of the PD and adaptive controllers.
where, P , a symmetric unique positive definite matrix is
chosen such that it satisfies the Lyapounov function
A. Joint Torque/Force plots
T
A P + P A = −Q (22) The joint torque plot consists of the joint torques computed
for all the 6 joints (2P, 4R) of the mobile manipulator
and Q ∈ Rnxn is a positive definite matrix. varying with respect to the time steps,with each time step
Thus, in this controller, the error vector e is updated as: DT of size 0.01. This torque magnitude corresponds to the
e = α · ė (23) control input fed to the dynamic model to achieve the desired
end effector position.
And in every loop of the iteration, the control aq is fed to The Fig. 5 shows the comparison of the joint torque plots
the dynamic model as shown in the equation below. in different scenarios as listed below to investigate the effect
of proportional gain Kp and derivative gain Kd on the control
τ = M (q) · aq + C(q, q̇) + g(q) (24) inputs for the P D and classical adaptive controllers.
In this study, the following values were chosen arbitrarily: • Kp = 30, Kd = 8
• Kp = 6, Kd = 8

Γ = 20 · I 8×8 No uncertain external force vector F was considered for this

comparison.
Q = I12×12 In the torque plot for PD controller with proportional gain
Kp = 6 and derivative gain Kd = 8, the torque magnitude
The control schematics of the classic adaptive controller for the joint2, joint5 and joint6 were observed to be noisy
is shown in the Fig. 4 below. between time steps 0 and 25. However, no overshoots or
ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 5

Fig. 5. Joint torque plots for (a) Kp = 6, PD +G controller (b) Kp = 30, PD +G controller (c) Kp = 6,Classic Adaptive controller (d) Kp = 30,Classic
Adaptive controller.

undershoots were observed in the amplitude of the joint is better in noise reduction compared to that of a P D
torques and joint forces in this case. controller with gravity compensation.
When the proportional gain is increased to Kp = 30
For the same P D controller, when the proportional gain for the classic adaptive controller, undershoots with large
is increased to Kp = 30, with the same derivative gain oscillations were observed for joint1 force. Joint2 also had
Kd = 8, undershoot of the torque amplitudes occurred for undershoots, but without oscillation of the torque amplitude
joints1 and joint2 and overshoots of different degrees with respect to time. No other joints had overshoots and
were observed for the other four joints. These undershoots undershoots.
and overshoots are due to insufficient derivative gain for Based on the above discussion, it is inferred that choosing
the proportional gain selected. In other words, the damping relevant and optimal Kd for Kp is needed to avoid overshoots
ratio would have to be increased or decreased to avoid the and over damping. However, selecting different Kp and Kd
overshoots and undershoots respectively. However,by doing values for different joints is needed to solve the overshoots
so, the overshoots and undershoots for all the joint torques and undershoots all at once, which demands automated PD
could not be avoided, as the Kp and Kd for all the joints tuning. The prevention of these overshoots and undershoots
are the same, according to equation 15 , 16. This shows of the joint torque as a control input is crucial, as the same
that each of the joints should have different proportional would lead to oscillation of the control response.
and derivative gains to prevent overshoot and undershoots Other than the above observations and inference, increase
of the in Kp is found to improve the error convergence rate. This
could be observed when comparing Fig. 5(a) and Fig.5 (b),
where the torque magnitude of most of the joints reach zero
joint torques for all the joints. In the case of classic adaptive faster for Kp = 30, when compared to Kp =6.
controllers, proportional gain Kp = 6 and derivative gain
Kd = 8 lead to smooth curves of torque with respect to time
steps and no noises were observed for this case. A very B. Joint Configuration Error Plots
mild undershoot of the joint 1 torque amplitude was noted. In this subsection, the effect of Kp and Kd on the joint
When comparing the torque plot results of classic adaptive angle error for the P D(gravity compensated) and classic
controller with PD controller Kp = 30, it is observed that adaptive controllers are investigated.The comparison of the
the initial torque input noises of joint1, joint5 and joint6 joint configuration errors for the scenarios as listed below
are eliminated. This infers that the classic adaptive control are shown in the Fig. 6:
ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 6

Fig. 6. Joint configuration error plots for (a) Kp = 6, PD +G controller (b) Kp = 30, PD +G controller (c) Kp = 6, Classic Adaptive controller (d) Kp
= 30,Classic Adaptive controller.

• Kp = 6, Kd = 8 The joint configuration error plot for the classic adaptive

• Kp = 30, Kd = 8 controller with Kp =6 and Kd =8 has smooth exponential
Similar to joint torque plots, the effect of Kp and Kd on error decay, as shown in Fig. 6 (c).No oscillations and noises
the joint configuration error response was investigated for were observed in this plot. When the proportional gain is
the P D and classical adaptive controllers. The plots shown increased to = 30, oscillations of the joint configuration error
in the Fig. 6 has the joint configuration error, i.e, the error were observed for the classic adaptive controller,as shown
between the desired and actual joint displacement or angle in Fig. 6 (d).However the amplitude and frequency of the
was plotted in the Y -axes against the time steps in the X- oscillations for classic adaptive controller with Kp = 30 were
axis. These joint configuration error plots were computed for observed to be lower than the P D controller with Kp = 30
the case with no uncertain external forces acting on the end as in Fig. 6 (b).
effector. In the following figure, the end effector position tracking
errors are compared for:
In the joint configuration error plot of the PD controller
with Kp =6 and Kd =8, the joint configuration error curves • Kp gains

were smooth without much of the noises. However mild • uncertain external forces to assess the performance of

noises were observed for the joints5 and joint6 , for the the chosen controllers in terms of tracking error
time steps from 0 to 10.This plot is shown in Fig. 6 (a). • error convergence rate of controllers

The noise in the torque of these joints (as discussed in the • robustness to uncertain external forces

subsection III-A) could be the reason for these mild noises. The plot has log||xd − x|| as Y -axis ,which varies in
Increase in proportional gain to Kp = 30 for the same P D accordance with the time steps in the X -axis.
controller results in joint configuration error oscillations as The end effector position error plots were compared be-
shown in 6 (b). This is attributed by the increase in Kp with tween the PD controller and classical adaptive controller for
insufficient Kd . However, the convergence rate was observed four scenarios as mentioned below:
T
to be quicker than the PD controller with Kp = 6.This is

• Kp = 30, F = 0 0 0
based on the observation that the joint configuration error 
• Kp = 30, F = 0 0 −4.95 N
T
reaches closer to zero in the 150th time step for Kp = 30  T
• Kp = 30, F = 0 0 0
as shown in Fig. 6 (b), whereas the error was not converged  T
at the same time step for PD controller with Kp = 06 ,as • Kp = 60, F = 0 0 −4.95 N

depicted in Fig. 6 (a) . For all the four plots shown in 7, the tracking error of
ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 7

Fig. 7. End effector error plot for (a) Kp = 30, uncertain mass =0 kg (b) Kp = 60, uncertain mass =0 kg (c) Kp = 30, uncertain mass =0.5 kg (d) Kp
= 60, uncertain mass =0.5 kg

the PD controller (gravity compensated) is observed to be of the robustness of the controller considered is shown in
lower than the classical adaptive controller. Whereas, the equation (25).
convergence rate of the classic adaptive controller in all the
four cases is higher than that of the P D controller with (xc − xuc )
R= (25)
gravity compensation. Moreover, the P D controller is found xc
to have greater oscillations of the tracking error,whereas the
where,
classic adaptive controller has fewer oscillations and curves
R is the robustness measure
flatten sooner.
xc is the L2 norm error in certain conditions (when the
It is also observed that for an increase in external uncertain uncertain force F = [0 0 0]T )
force from 0 to 4.95 N in −Z direction, there is an increase xuc is the L2 norm error in uncertain conditions(when the
in permanent tracking error irrespective of the controller uncertain force acting is not a zero vector)
type.Similar observation was found in one of the reference
papers[2], where the controllers had a permanent tracking Based on the above equation robustness for the classical
error due to effect of gravitational force. Thus, this result adaptive and P D controllers for different proportional gains
once again acknowledges that the external forces contribute are computed and plotted in the Fig. 8 shown below.The
to permanent tracking error of the PD and classic adaptive figure has two plots and each of the plots are the robustness
controllers. for uncertain external force magnitudes of 2.45 N and 4.9
Moreover, the increase in proportional gain Kp is found N respectively. For the cases with no uncertain forces or
to reduce the tracking error. The same is observed while disturbances, the robustness of the controllers are considered
comparing the Fig. 7 (c) with Fig. 7 (d) , wherein the tracking to be 100%.
error for adaptive controllers was found to be approximately From the plots for robustness of the controllers as shown
-2.5 for Kp = 30 and -3 for Kp = 60. in Fig. 8 (a) and Fig. 8(b), the robustness is found to increase
The robustness of the P D with gravity compensation linearly with increase in the proportional gain. Moreover, the
and classical adaptive controllers are compared using the robustness decreases with increase in the external uncertain
robustness plots in the Fig. 8 as shown below. input or disturbances, which is the unknown object mass in
For this study, robustness is considered as a quantitative the end effector.
measure, which is defined as a ratio of change in tracking Based on the above plots, the robustness of the P D
error to the tracking error without uncertainty. The equation controller with gravity compensation is higher than that of
ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 8

Fig. 8. Robustness plot between PD+G and adaptive for (a) Uncertain Mass = 0.25 kg, (b) Uncertain Mass = 0.5 kg

the classical adaptive controller.Though, the classic adaptive controller for the 6-DOF mobile manipulator chosen in this
controller was expected to perform better than the P D study.
controller, the result obtained is counter-intuitive . In order to
TABLE II
investigate this , one extra simulation test was run changing
S UMMARY OF CONTROLLER PERFORMANCES
the value of the regressor matrix Y and applying the same
in the classic adaptive controller. For the chosen uncertain
Controller Tracking Error Robustness
parameters, the regressor matrix Y was computed as shown error Conver-
in equation(10), and compared with computation using the gence
pseudo-inverse of the vector θ . The tracking error obtained is PD+G controller Low Low High
found to vary drastically, according to the change in the mag- Classical Adaptive Controller High High Low
nitude of the components in the regressor matrix. Moreover,
it was computationally complex to find the regressor matrix
Y for the chosen theta parameters, for this 6-DOF robot. IV. C ONCLUSION
Thus, in this study, the inaccuracy of the regressor matrix
In this study, the performance of the Proportional Deriva-
parameters is found to contribute to the classic adaptive
tive controller with gravity compensation is compared with
controller being less robust than the PD controller (gravity
the performance of the classic adaptive controller for a pick-
compensated).
The Fig. 9 below shows the trajectory followed by the end and-place 6-DOF mobile manipulator, to study the effect of
effector to reach the desired end effector position. Though the uncertain mass of the object picked on the controller
the tracking error and convergence rate of the classic adaptive performances.
controller and the P D controllers vary between each other, Both the PD + G controller and classic adaptive controllers
a similar trajectory was followed by both of the controllers. were found to have a constant tracking error, which is due
to the uncertain mass acting as a weight on the end effector
in −Z direction. PD + G controller was simpler and faster
to implement than the classic adaptive controller. This is
attributed to the computation time of the regressor matrix
Y for a 6-DOF mobile manipulator. Though the rate of error
convergence was faster for the classic adaptive controller, the
PD + G controller has smaller tracking error when compared
to the classic adaptive controller, which was quite a counter-
intuitive result. However, the high sensitivity of the tracking
error for smaller inaccuracies in the Y regressor matrix is
the reason for a higher tracking error in the classic adaptive
controller.
Since robustness is considered as the relative measure of
error in uncertain conditions with respect to the error in
Fig. 9. Trajectory Followed by the End Effector and the Mobile Base to
reach the desired End Effector Position. certain conditions, PD + G controller found to be more
robust than the classic adaptive controller , in this study.
The table shown below summarizes the performance com- Apart from the above conclusions,increasing the proportional
parison of the adaptive and P D controller (gravity compen- gain reduced the tracking error, and increased the error
sated) relative to each other.Based on the following summary, convergence rate and robustness of the MRM irrespective
the P D controller performs better than the classical adaptive of the controllers chosen. However, a wise choice of the
ECE 780, Robot Dynamics and Control, University of Waterloo, Spring 2022 9

derivative gain is required to prevent the oscillations of the

control response.
In the future work, the performance of the PID controllers
as well as the automatic tuning of the PD parameters could
be explored.

R EFERENCES
[1] M. Boukattaya, T. Damak, M. Jallouli, Robust adaptive control for
mobile manipulators, International Journal of Automation and Com-
puting 8 (1) (2011) 8–13. doi:10.1007/s11633-010-0548-y.
URL http://link.springer.com/10.1007/
s11633-010-0548-y
[2] T. Žilić, J. Kasać, M. Essert, B. Novaković, Ž. Šitum, Performance
comparison of different control algorithms for robot manipulators,
Strojarstvo: časopis za teoriju i praksu u strojarstvu 54 (5) (2012) 399–
407.
[3] C. López-Franco, J. Hernández-Barragán, A. Y. Alanis, N. Arana-
Daniel, M. López-Franco, Inverse kinematics of mobile manipulators
based on differential evolution, International Journal of Advanced
Robotic Systems 15 (1) (2018) 1729881417752738.
[4] B. Siciliano, L. Sciavicco, L. Villani, G. Oriolo, Differential kinematics
and statics, Robotics: Modelling, Planning and Control (2009) 105–160.
[5] M. W. Spong, S. Hutchinson, M. Vidyasagar, Robot modeling and
control, John Wiley & Sons, 2020.
[6] B. Siciliano, L. Sciavicco, L. Villani, G. Oriolo, Motion control,
Robotics: Modelling, Planning and Control (2009) 303–361.