Meccanica (2022) 57:473–506

https://doi.org/10.1007/s11012-021-01464-7 (0123456789().,-volV)
( 01234567
89().,-volV)

Dynamics analysis, offline–online tuning and identification

of base inertia parameters for the 3-DOF Delta parallel
robot under insufficient excitations
Faraz Abed Azad . Saeed Ansari Rad . Mohammad Reza Hairi Yazdi .
Mehdi Tale Masouleh . Ahmad Kalhor

Received: 19 November 2020 / Accepted: 15 November 2021 / Published online: 3 January 2022
Ó Springer Nature B.V. 2022

Abstract In this paper, the dynamics model of the bandwidth. Additionally, this paper presents an algo-
3-Degrees-Of-Freedom (DOF) Delta parallel manip- rithm to address the singularity problem by mapping
ulator is elaborated by means of a Screw-based the regressors to an orthogonal space and obtaining an
formulation of the virtual works method in a linear optimal set of base parameters in an online manner,
regression form. Moreover, a reduced dynamics model namely the online base inertial parameter tuning
is obtained by exploiting the singular value decom- method. Thereafter, through a set of simulations, the
position, leading to the determination of the base effectiveness of the proposed method has been illus-
inertial parameters. This method is an offline tuning trated for different paths. Finally, a change in the value
approach for which an optimized path containing of an inertial parameter of the system, which can be
several harmonics is required to establish the desired regarded as a disturbance, is exerted and the obtained
results reveal that the proposed method can compen-
sate for extrinsic mass effect. In summary, it can be
F. Abed Azad deduced that the the proposed approach is also able to
Human and Robot Interaction Laboratory, School of identify the changed model under insufficient excita-
Mechanical Engineering, University of Tehran, Tehran,
Iran
tions for real-time purposes.
e-mail: [email protected]
Keywords 3-DOF Delta parallel robot  Dynamics
S. Ansari Rad  M. Tale Masouleh (&) analysis  The virtual works method  Insufficient
Human and Robot Interaction Laboratory, School of
Electrical and Computer Engineering, University of
excitations  Online identification  Online tuning of
Tehran, Tehran, Iran base parameters
e-mail: [email protected]
S. Ansari Rad
e-mail: [email protected]
1 Introduction
M. R. Hairi Yazdi
School of Mechanical Engineering, University of Tehran,
Tehran, Iran The idea of the Delta parallel robot is accredited to
e-mail: [email protected] Prof. Clavel [18] which was the results of his PhD
thesis [19]. The widespread of application of the Delta
A. Kalhor
School of Electrical and Computer Engineering,
parallel robot makes this robot, beside the Gough-
University of Tehran, Tehran, Iran Stewart platform [84], as one of the commercial and
e-mail: [email protected]

123
474 Meccanica (2022) 57:473–506

widely used parallel robots in the industry. This links of the Delta parallel robot has been neglected and
manipulator performs three transitional Degrees-Of- then, the Newton-Euler method was used. However,
Freedoms (DOF) by means of three identical limbs. due to the fact that almost all of the constraint forces
Topology of the 3-DOF Delta parallel robot results in and torques appear in the calculation, the Newtonian
better rigidity, leading to higher acceleration and method is a suitable strategy when it is desired to study
precision capabilities as well as reducing the weight of the manipulator’s structural design or selecting the
each robotic chain, which makes this mechanism very most appropriate joints or actuators [35, 41].
applicable in the industrial automation [61], especially Another alternative is the Lagrangian method,
in those that precision and speed are of great essence; which is based on the energy of the mechanism;
ranging from delicate surgeries [8, 53] and jewelry or therefore, in contrast to Newtonian method, there is no
precise machining [46, 93] to high-speed pick-and- need to directly compute internal forces in the
place applications, whether by the original Delta manipulators’ kinematic chains. However, since in
parallel manipulator [13, 63], or other variations of the parallel manipulators the constraint of closed loops
Delta-like parallel configurations [56, 62]. Ultimately, should be satisfied for all the limbs and the robot as a
as an alternative application of a Delta-like parallel whole; thus, these constraints will increase the number
mechanism, in [59], the inertial properties of a rigid of Lagrange multipliers to be determined. The latter
body have been determined by using a 3-URU parallel issue results in higher volume of computations,
mechanism. leading to higher values of errors in the identification
There has been a vast literature on the kinematic of dynamics parameters and torque computations, as
analysis of the Delta parallel robot. In [34], the screw well as longer process times. Therefore, in a multitude
theory is proposed as a tool to determine the of researches, the geometry and inertia distributions of
kinematics of the Delta parallel robot; however, the the manipulator have undergone some simplifying
final formulations have not been presented. In the assumptions which provokes some errors in the
context of dynamic analysis of the Delta parallel robot, derived dynamics equation of motion [87, 96]. Addi-
from the best knowledge of the authors, there is a gap tionally, in other works, some researchers opted the
on applying the screw-based formulation of the virtual Kane’s method for dynamics analysis of parallel
works approach to determine a compact dynamics manipulators [17], while some resorted to other
model of this manipulator. The screw theory is a improved formulations for parallel manipulators
powerful tool in order to derive parallel manipulators’ [2, 42, 90].
kinematics and dynamics equations of motion conve- Ultimately, due to the complexity of the topology of
niently in a compact form [3, 32, 75], to address their parallel manipulators and their closed-loop constraint
singularity analysis [48], or to facilitate stiffness equations, the virtual works method is an advisable
analysis by deriving the screw-based formulations method especially for this class of manipulators
[15]. In [14, 64], the authors resorted to the screw [11, 21]. The virtual works method is opted by quite
theory to evaluate kinematic and dynamic perfor- a few researchers [30, 33, 37, 91] in analysis of diverse
mance of various structures of the Delta robot, and classes of parallel manipulators. This method, akin to
another study [12] carried out the kinematics and the method of Lagrange, is based on the energy of the
workspace analysis of the prismatic Delta parallel manipulator and does not stipulate derivation of
robot. internal forces and torques. Over and above, there
Accuracy and compactness of the dynamics model would be no need to calculate the energy of the
of the parallel robots are of great importance [80]; mechanism and the corresponding Lagrange Multipli-
because, this kind of robots can achieve high speeds ers. Moreover, this method, compared to the Newto-
and accelerations with high precision [55, 81]. For this nian and Lagrangian approaches, is almost
end, their dynamics model was studied via different straightforward and the computations volume does
methods. First and most general of them is the not highly depend on the choice of generalized
Newton-Euler method. The Newtonian formulation coordinates. By the same token, this reduction in the
demands higher volume of computations and it is not volume of computations allows one to incorporate
practically applicable in real-time control. In order to complementary methods to the principle of virtual
mitigate this problem, in [22] the rotational inertia of works in order to improve the accuracy of calculations

123
Meccanica (2022) 57:473–506 475

in the determination of the dynamics models [44], of manipulators [27]. Besides, some parameters,
such as implementation of various automatic control regardless of the planned trajectory or its excitation
schemes with higher complexities [10, 92]. number, do not affect the dynamics response at all
Several approaches have been proposed in the [23]. While on the other hand, some parameters have
literature for the identification of dynamic parameters, respectively similar effects in such a way that the
namely, direct measurement, CAD software, and response can be written in linear combination of these
mathematical approach. The first one, which is based parameters [27]. Thus, only a set of parameters are
on direct mechanical inspection, requires to disassem- identifiable, which for reducing the computational
ble the mechanism which in many cases may not be time of the identification procedure, the minimal set of
possible. As for the second case, CAD models have these parameters or base inertial parameters, are
been shown to be relatively inaccurate compared to the determined whether directly in symbolic form via
experimental results [82]. Moreover, attaining these the dynamics equation of motion [43, 49, 72, 73], or
parameters with high precision from CAD models or first numerically and then converting it into symbolic
experimental results gives the ideal or nominal value results [1, 27, 38, 39, 82]. This formulation will
of the parameters regardless of the manufacturing culminate in the objective of reducing the model and
processes used in fabricating the manipulator. There- makes the identification of parameters possible.
fore, using identification techniques on the dynamics Nevertheless, for the purpose of accurate robot
parameters is a crucial step for deriving an accurate identification, an imperative challenge arises where
dynamics model. Since the subject of this paper is one should identify the optimal trajectory [16] as well
concerned with dynamics identification of a 3-DOF as a set of sufficient excitation [85]. To do so, it is
Delta parallel robot, in what follows, some works have required to first determine the desired bandwidth for
been included in this regard. In [45], dynamics which the actuators of the manipulator are intended to
identification and model feed-forward control of a perform, and also the total number of inertial param-
3-DOF Falcon device have been performed. In eters. To this end, the optimal path proposed in [85] is
[78, 79], authors investigated direct and indirect considered in form of finite Fourier series for each of
dynamics identification for an over-constrained the actuators. In the latter study, the number of
3-DOF decoupled parallel manipulator, namely, the harmonic frequencies of the path is chosen based on
Tripteron parallel robot; and in [80], dynamics iden- the required excitation number. However, most of the
tification and control of the 3-DOF Tripteron parallel study conducted on the identifications of base inertial
manipulator have been addressed in presence of a parameters are based on offline tuning approach.
variable friction and feedback delay. In [74], authors Nonetheless, in the case that external loads are applied
performed dynamics control for a 3-RPS parallel robot or sudden changes in the inertial parameters occurs,
by algebraic parameters identification. In [66], the online identification would be an inevitable approach.
dynamics identification of a 3-DOF Delta parallel The RLS method plays a significant role in
robot is addressed along with the identification of the parameters estimation and control of time-varying
actuator dynamics in a serial form. systems which is suitable for time-varying systems
Consistently, referring to the third approach men- with sufficient excitation. Otherwise, when excitation
tioned above, in order to perform inertial parameters is insufficient [4, 5, 68, 69], the singularity problem
identification, the developed dynamics model is which is referred to as estimation windup (also known
reformulated in a linear regression form, attaining as estimation blowup) may occur. In this case, the
the matrix of regressors and parameters vector. In estimation gain increases exponentially and outputs of
[50, 51], authors show that the dynamic equation of the estimator become sensitive to any disturbances in
manipulators is linear w.r.t. (with respect to) their the regressors. A few methods were proposed to
inertial parameters. Consequently, these parameters attenuate the estimation windup [9, 70, 77]. These
can be identified by the so-called Recursive Least methods, however, do not provide any systematic
Squares (RLS) method. Nevertheless, in parallel solutions. In [28], a RLS scheme with convex
manipulators, inertial parameters and some rows of regularization is investigated. In [89], an algorithm
the observation matrix are dependent on each other based on the LS method with smoothing conditions
due to the constraint equations of motions in this kind and adaptive regularization of parameters is proposed

123
476 Meccanica (2022) 57:473–506

which offered a suitable solution for numerical conditioned situations is proposed. Moreover, in
instabilities in the classic LS. However, the method [29, 47, 52, 86] different approaches for the RLS
is impractical for many control applications since a method are introduced for ill-conditioned situations.
recursive formulation cannot be derived for the The main contributions of this paper consist in
algorithm. determining a compact formulation for dynamics
In [54], a method known as Damped Least Squares modeling of the 3-DOF Delta parallel robot which is
(DLS) was proposed which modified the classical RLS convenient for online identification. The foregoing
method to cope with parameters’ variations. The idea contribution incorporates two concepts, namely,
of the foregoing method was based on optimizing a establishing the base inertial parameters of the under
quadratic cost function. Similar approaches were study parallel manipulator and proposing a novel
adopted in [28, 31, 76, 83, 98]. Convergence and concept of online base inertial parameters tuning and
stability of recursive DLS algorithm, as two main identification under insufficient excitations. In fact,
criteria, is discussed in [98]. A generalized version of online inertial parameter identification is a definite
DLS, called GDLS, was proposed in [97]. The latter asset in practice; since deriving the accurate online
algorithm systematically refined the estimation dynamics parameters is an undeniable prerequisite of
windup in a closed-loop control strategy. The new implementing a precise and fast control scheme. In
definition of the cost function resolved the singularity this regard, the manipulator may be under influence of
problem in steady state which makes GDLS a various unanticipated disturbances in its working
powerful alternative approach for tracking systems environment, which may cause changes in the dynam-
with slowly changing parameters [6, 67, 100]. ics model’s parameters.
In [6], a new Auxiliary-Model-Based Damped More specifically, in this paper, first, kinematic
Recursive Least Squares (AMBDRLS), which could analysis of the robot was performed by means of the
simultaneously identify the system model and the screw theory followed by dynamics analysis via a
noise model, was developed in which the heart rate formulation of the virtual works principle compatible
dynamics could be identified in a real-time manner. In with the screw theory. Hence, the linear dynamics
[40], another technique using the filtering technique equations of motion of the parallel manipulator has
and the auxiliary model for simultaneously identifying been derived in a compact form which allows one to
the system model and the noise model is proposed. In readily and rapidly extract the torques of each actuator
fact, GDLS recursively uses the unitary matrix for while maintaining high accuracy and without simpli-
avoiding the possible singularity which leads to the fication of the model. Furthermore, this formulation
same problem in the regularized constant-trace [9], enables one to perform online and offline identifica-
leading to an increase in the output estimation error. tion procedures and to conveniently obtain the actu-
Therefore, GDLS’s control applications are limited to ator torques in a recursive form. Reaching this step, in
a few adaptive schemes. The Singular Value Decom- order to perform inertial parameter identification for
position (SVD) is used in [65, 94, 95] to address the parallel manipulators, the base inertial parameters
singularity problem in the LS method. However, a should be obtained as the first step. However, in order
systematic solution for the estimation windup caused to determine the base inertial parameters and perform
by singularities is not provided. Some numerical offline identification, a path with sufficient excitation
methods are also proposed to solve the singularity number is required. In parallel manipulators, there are
problem. In [25], a Self-Adaptive Iterative Algorithm a relatively high number of inertial parameters to be
(SAIA) is introduced for processing ill-conditioned determined compared to their serial manipulator
matrices emerging in the RLS technique. The SAIA counterparts. This fact results in a large number of
refines the speed of convergence and the accuracy of minimum sufficient excitation number. In order to
calculations by avoiding inverting matrices with derive a path with sufficient excitation number for the
Cholesky decomposition, where parameters are esti- identification purpose, a set of cumbersome and time-
mated based on the rate of residual error. In ill- consuming multi-variable optimizations would be
conditioned situations, the LS method may result in required. Even though one can obtain the perfect path
inaccurate estimations. In [71], an adjustable approach with low enough excitation number for a given parallel
for increasing the accuracy of the LS method in ill- manipulator, this path only enables offline base inertial

123
Meccanica (2022) 57:473–506 477

parameters identification. As mentioned above, the dedicated to carry out some studies to simulate the
manipulator may be under influence of various probable problems associated with the implementa-
unforeseen disturbances in its working environment, tion of the method on the robot. In this regard, an
causing changes in the dynamics model’s parameters. investigation has been carried out to illustrate the
In these cases, online identification is inevitable to capability of the method to compensate and identify
attain precise and fast dynamics control. For instance, the base inertial parameters with high precision and
as a simple application, the fast pick-and-place performance in the case of implementation of an
applications where the weight on the end-effector extrinsic mass to the parallel manipulator in the midst
may vary on each cycle can be regarded as an example of the simulation, where the challenge of considering
for the latter issue. Nevertheless, in order to perform insufficient excitation number still exists. Also, in
online parameter identification, unfortunately the order to extend the investigations for practical issues,
sufficient excitation number requirement still holds sensors’ measurement noise have been taken into
where the desired path may hardly fulfill this require- account while implementing the proposed method
ment to achieve precise parameter identification. In under insufficient excitations. This investigation is a
these cases, estimation windup problem occurs; which prerequisite for the ongoing work of conjointly online
causes the estimation gain to increase exponentially base parameter identification and implementation of
and outputs of the estimator become sensitive to any precise dynamics control schemes on the manipulator
disturbances in regressors, leading to inaccurate in practice.
estimations or even instability of the system in cases The remainder of the paper has the following
of very low excitation numbers compared to the structure. First, in Sect. 2, the structure of the under
sufficient one. In order to address the aforementioned study manipulator is presented and the prerequisite
dilemmas, merging the ideas arisen from DLS and kinematic parameters, comprised of position vectors
SVD, a novel algorithm is introduced which resolves and joint variables, are elucidated. In Sect. 3, the
the aforementioned identification windup problem dynamics equation of motion of the Delta parallel
under insufficient excitations. The SVD damped-LS manipulator is extracted by means of the virtual works
(SVD-DLS) approach searches through the orthogonal principle, where, a compact linear regression formu-
space and obtains the optimal estimation in the RLS. lation for the dynamics model of the manipulator has
The latter leads to alleviate the windup problem and been established. In Sect. 4, the under study 3-DOF
the capability to damp parameter oscillations. The Delta parallel robot has been modeled in MATLAB’s
proposed method is applied for both tuning and Simscape and a simulation study has been carried out
identifying the base inertial parameters of the Delta to verify the validity of the elaborated kinematics and
parallel robot in an online manner. The obtained dynamics equations. In Sect. 5, the SVD-based
results reveal that the online tuning of base inertial damped RLS method is explained. The proposed
parameters along with their online identification has method eludes the wind-up problem where the system
been performed successfully with high precision in the has insufficient excitations. With the aim of solving
cases that the excitation number is far less than what is this dilemma, Sect. 6 is devoted to determine the base
considered as a sufficient excitation. inertial parameters by the SVD method and optimizing
Finally, it should be mentioned that as the ongoing the training path with sufficient excitations via the
work of this manuscript, the proposed method will be genetic algorithm. After testing the computed base
used to perform online tuning and identification of the inertial parameters, namely the offline tuning method
base inertial parameters in order to implement precise of base parameters, the proposed online tuning method
automatic control schemes on the manipulator. There- of base inertial parameters has been used to compute
fore, as a limitation of this paper and a prerequisite of the actuator torques while identifying and tuning the
the ongoing works, an online tuning and identification base parameters in an online manner. Then, perfor-
of the base inertial parameters method has been mance of the proposed method has been inspected by
proposed in this paper and its performance has been comparing its resultant torques with the reduced
investigated in simulations. However, in order to dynamics model, which is based on the offline tuning
assimilate the simulations to the implementations in of base parameters, and the full dynamics model.
practice, a section at the end of the paper has been

123
478 Meccanica (2022) 57:473–506

Finally, the paper concludes with some hints and

remarks as ongoing works in Sect. 7.

2 Kinematic analysis of the Delta parallel robot

The first step to derive the dynamics equations of

motion consists in determining the governing kine-
matic relations. Figure 1 schematically depicts the
architecture of the Delta parallel manipulator under
study. The foregoing parallel manipulator consists in
three limbs or closed loop chains which connects the
end-effector to the fixed base plate. As represented in
Fig. 2, each limb has two links, namely, the actuated
link and the parallelogram link. In each limb, the
actuated link is connected to the base platform by an
actuated revolute joint and the parallelogram link
connects the actuated links with the end-effector via
two universal joints located in each side of it. The Fig. 2 Representation of one limb of the Delta parallel
actuated links are represented by vector l1i and the manipulator with its kinematic parameters
parallelogram links are addressed by vector l2i which
connects the center of universal joints to each other. Initially, the global reference frame is attached to the
As demonstrated in Fig. 2, the angle of actuated links’ center of the fixed platform for which the x-axis
revolute joint is addressed by h1 and the universal joint intersects with one of the active joints and the z-axis
variables are h2=1 and u which respectively describe points downward vertically. For the sake of simplify-
the planar angle between upper and lower links (i.e., ing the computations, the global reference frame has
h2 ¼ h1 þ h2=1 ) and the angle of the parallelogram. been rotated by 120 along its z-axis, creating three
Additionally, in Fig. 3, the normal view of the additional reference frames defined at each of the
manipulator’s base platform with the requisite coor- active revolute joints on the fixed platform’s plane.
dinate frames is provided. Alike global reference frame, these frames are fixed
In what follows, prior to delving into the kinematics and they are regarded as the limb coordinate frames.
analysis, the coordinates used throughout the kine- With the aim of performing dynamics analysis, a
matics and dynamics modeling have been elaborated. local coordinate system should be defined for each of

Fig. 1 Schematic of the

Delta parallel manipulator

123
Meccanica (2022) 57:473–506 479

screws are represented by bold handwritten English

letters and symbol ‘‘ $ ’’. Meanwhile, hat accent (‘‘b’’)
on lowercase letters stands for unit vectors or unit
screws, and for uppercase letters, it represents a skew-
symmetric matrix. In addition, ‘‘’’ and ‘‘’’ indicate
cross product and dot product of two vectors, respec-
tively. Furthermore, the reciprocal product of two
screws is illustrated by the operator ‘‘’’.

2.1 Position analysis

As the first step toward obtaining the kinematic

relations of the Delta parallel robot, one ought to find
a solution for its position problem. In what follows, the
geometric relations of the robot have been elaborated.
Fig. 3 Fixed base platform of the Delta parallel manipulator A loop-closure equation is written for each of the
with end-effector’s kinematic parameters limbs in the corresponding limb coordinate system to
establish the governing kinematic equations of the
the links in such a way that they are attached to the robot:
corresponding link and move concurrently with it. For ! ! ! ! !
the upper links or the actuated links, the latter OAi þ Ai Ci þ Ci Bi þ Bi P ¼ OP i ¼ 1; 2; 3
coordinate system’s center is located on the actuated ð1Þ
revolute joint which connects the limb to the base
platform. In the parallelogram links, the center of the In Eq. (1), the accent ‘‘!’’ is adopted to display vector
coordinate system is located on the revolute joint connecting the first point to the next one. Now, by
connecting the limb to the end-effector platform. The expanding Eq. (1) algebraically and representing them
x-axis of the latter coordinate frames are along the in limb coordinate systems, one has:
direction of the line connecting two joints of the L1 cos h1 þ L2 sin u cosðh1 þ h2 Þ ¼ qx ð2Þ
corresponding link, or l1 and l2 in other words; thus,
other axes would be determined by transforming the L2 cos u ¼ qy ð3Þ
center and x-axis of the limb coordinate frame to the
aforementioned configuration. There are two consid- L1 sin h1 þ L2 sin u sinðh1 þ h2 Þ ¼ qz : ð4Þ
erable benefits for selecting these points, first and
foremost, the location of this centers are fixed on the In the above expressions, q ¼ ½qx ; qy ; qz T represents
link, thus makes the identification process convenient position vector of the parallelogram links’ passive
and avoids the necessity of changing hefty number of joint which connects it to the end-effector, w.r.t. the
parameters in each iteration. Secondly, the link corresponding limb coordinate systems center or
Jacobian matrices would be simplified to a significant namely the active links’ upper revolute joint; in other
degree, since the linear velocity part of the matrices, !
words, Ai Bi i ¼ 1; 2; 3. Additionally, L1 and L2
which involves more terms and parameters than their denote the length of the upper links and lower links,
angular velocity counterpart, is simplified conspicu- respectively. From Eq. (3), one can directly calculate
ously. These facts will be more evident in Sects. 3.1 u as:
and 2.3.  
For the sake of unifying the representation of u ¼ cos1 qy =L2 : ð5Þ
mathematical parameters used throughout this paper, a
It should be noted that the domain of u is a subspace
defined notation is adopted which is explained in what
of ð0; pÞ [36, 58]; where for u ¼ p2, the corresponding
follows. Scalars are presented with non-bold alphabet,
vectors are demonstrated by lowercase bold alphabet, link’s motion is planar. Therefore, the solution for u in
matrices are depicted by Bold capital alphabet and Eq. (5) is unique. As the next step, by summing the

123
480 Meccanica (2022) 57:473–506

squares of Eqs. (2), (3), and (4), and by performing for h1 in Eq. (14) is unique. Also, without loss of
simplifications by algebraic manipulation, one has: generality, according to the manipulators workspace
[24, 36, 58], the value of h2=1 is positive; since, the
2L1 L2 sin u cos h2 þ L21 þ L22 ¼ q2x þ q2y þ q2z ð6Þ domain of h2 ¼ h1 þ h2=1 is a subspace of ð0; pÞ.
where the angle h2 can be obtained as follows: Consequently, in regard to the determination of h1 , h2 ,
  and u, all of the corresponding position vectors
h2 ¼ cos1 ðq2x þ q2y þ q2z  L21  L22 Þ=2L1 L2 sin u : including l1 and l2 can be computed, as calculated
later in Eq. (26).
ð7Þ
On the other hand, isolating the terms containing the 2.2 Velocity analysis
variable h2 in Eqs. (2) and (4), and then summing the
squares of them results in: As the next step, in order to execute kinematic
analysis, it is required to derive the velocity equations
2L1 ðqx cos h1 þ qz sin h1 Þ ¼ q2x þ q2z þ L21  ðL2 sin uÞ2 : for each joint variable. To this end, the screw theory is
ð8Þ employed to first obtain the governing relation
between twist of the end-effector and velocity of the
Hence, in order to solve Eq. (8) for determining angle joint variables, and consequently acquire the rate at
h1 , it is intended to transform the foregoing equation which each of the joint variables is changing. For this
into a polynomial expression and define the half-angle purpose, one may consider each limb as an open-loop
tangent substitution as what follows: kinematic chain comprised of 5 single DOF joints
w ¼ tanðh1 =2Þ: ð9Þ connecting the end-effector to the fixed base platform
[88]. Therefore, the instantaneous twist of the end-
Substitution of Eq. (9) into Eq. (8) and performing effector can be described for each limb in its
some simplifications leads to: corresponding limb coordinate frame as follows:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 b þ b2  4ac $p ¼ h_ 1 $^5 þ h_ 2 $^3 þ u_ $^4 ð15Þ
aw þ bw þ c ¼ 0; w¼
2a
In the preceding equation, $^5 is the screw associated
ð10Þ
with h1 , which corresponds to the revolute joint of
where: upper or the actuated link. Moreover, rotation of the
universal joint is composed of two screws, denoted by
a ¼ ðqx þ L1 Þ2 þq2z  ðL2 sin uÞ2 ð11Þ $3 and $4 , which are the first and second rotation of the
universal joint respectively. Subsequently, the screw
b ¼ 4L1 qz ð12Þ $p denotes the twist of the end-effector.
Reaching this step, the concept of reciprocity is
c ¼ ðqx  L1 Þ2 þq2z  ðL2 sin uÞ2 ð13Þ applied to determine each of the joint variable’s
Therefore, one has the following for h1 : velocity. First, by locking the actuated joint in each
limb, a screw reciprocal to all of the remaining screws
h1 ¼ 2 tan1 w: ð14Þ – passive joints – except the active joint screw can be
determined. Then by applying the reciprocal product
In a previous research carried out at the University of to the both sides of Eq. (15), the passive joint screws
Tehran’s Human and Robot Interaction Laboratory are eliminated:
[24], workspace of the under study 3-DOF Delta  
parallel manipulator has been derived. In this regard, $r  $p ¼ $r  h_ 1 $^5 þ h_ 2 $^3 þ u_ $^4 ð16Þ
without loss of generality, the domain of h1 has been
   
determined as a subspace of  p2 ; p2 ; where for
$Tr  $p ¼ $Tr  h_ 1 $^5 ð17Þ
h1 ¼ 0, the corresponding link will be in horizontal
configuration, or in other words, the upper link will be in which:
parallel to the base platform. Therefore, the solution

123
Meccanica (2022) 57:473–506 481

 
l1 xp noting that twist of a link consists in linear velocity of
$r ¼ ; $p ¼ : ð18Þ a point on the link and angular velocity of the rigid
0 vp
body about the corresponding point [3].
By expanding the reciprocal product, it follows that: The actuation Jacobian matrix establishes the
  velocity relation between input and output—actuators
xp s^1
T
0 l1 T  T
¼ 0 l1  T h_ 1 : and the end-effector—of the system. Ergo, regarding
vp ðl2 þ l1 Þ  s^1 to the results of previous section, one has:
ð19Þ 2 3
lT21 T
0
Thus, one has: 2 3 6 T
6 l11  ðl21  s^11 Þ
7
7
  h_ 11 6 7  
6_ 7 6 7
T vp vp
lT11 vp 6 l 22 T7
h_ 1 ¼ T 0T  : ð20Þ 4 h12 5 ¼ 6 T 0 7 ¼ Ja
l21  ðl21  s^11 Þ xp l
6 12 22 ðl  ^
s 12 Þ 7 x p xp
h_ 13 6 7
4 lT23 5
Thereafter, by locking the joint corresponding to angle 0T
u, the reciprocal screw to the remaining screw system lT13  ðl23  s^13 Þ
is determined as: ð24Þ
 The matrix Ja is known as input-output Jacobian
^ s^1
$r;u ¼ $1 ¼ ð21Þ matrix or the actuation Jacobian. Since the 3-DOF
0
Delta parallel robot’s end-effector could only possess
which leads to: translational DOF (i.e., xp ¼ 0), the last three
  columns of Ja does not have any contribution in h_ 1 ,
s^T1 vp
u_ ¼ T 0T : ð22Þ and they may be eliminated. Therefore, according to
l1  ðs^2  s^1 Þ xp
the derived inverse kinematics equations and the
Now by substituting Eqs. (20) and (22) into Eq. (15), actuation Jacobian matrix, Eq. (24) can be reformu-
h_ 2 would be: lated as:
20  T 1T 3 2 31
s^2  ðl2  l1 Þ lT21
6 Bl1 þ T s^1 7 v 6 lT  ðl  s^ Þ 7 2
6 l2  ðs^2  s^1 Þ C T7 6 11 21 11 7 3 2 3
h_ 2 ¼ 6 B C p
7 : 6 7 h_ 11 h_ 11
4B C 0 6 T 7 6
@ lT1  ðl2  s^1 Þ A 5 xp l22
vp ¼ 6 7 _ 7 1 6 _ 7
6 lT  ðl  s^ Þ 7 4 h12 5 ¼ Ja  4 h12 5
6 12 22 12 7
6 7 h_ 13 h_ 13
ð23Þ 4 lT23 5
lT13  ðl23  s^13 Þ
2.3 Jacobian matrices ð25Þ
In which:
In order to formulate the dynamics model of a system
by means of the virtual works principle, Jacobian s^T1i ¼ ½0 ; 1 ; 0 ; lT1i ¼ L1 ½cosðh1 Þ ; 0 ; sinðh1 Þ
matrices should be determined. For the purposes of s^T2i ¼ ½sinðh1 þ h2 Þ ; 0 ;  cosðh1 þ h2 Þ
this paper, two requisite type of Jacobian matrices are
lT2i ¼ L2 ½sinðui Þ cosðh1 þ h2 Þ ; cosðui Þ ; sinðui Þ
obtained, namely, the actuation Jacobian and the link
Jacobian matrix. The actuation Jacobian maps twist of cosðh1 þ h2 Þ
the actuators into the twist of the end-effector. This ð26Þ
matrix is an asset for forward kinematics analysis and
As described above, the link Jacobian matrix formu-
path optimizations, along with implementation of
lates the relation between twist of the end-effector and
various dynamics control algorithms. In turn, the link
twist of each link; although, according to the screw
Jacobian matrix relates the end-effector’s twist into
theory, one can consider any point on the correspond-
the twist of the corresponding link, and therefore, each
ing link to establish the twist for, and the screw-basis
link does have a link Jacobian matrix. It is worth

123
482 Meccanica (2022) 57:473–506

   2
Jacobian matrix could be greatly simplified if its
v_ p ¼h€1 s^1  l1 þ h€2 s^1 þ u^
€s2  l2  h_ 1 l1
reference frame is defined at specific points [88].  2 
Generally, gravitational center of each link is selected  h_ 2 þ u_ 2 l2
as reference point to calculate the twist for; neverthe-  
less, the choice of link coordinate frames’ center for þ h_ 2 u_ ðs^1  ðs^2  l2 Þ þ s^2  ðs^1  l2 ÞÞ
calculating the twist would simplify the link Jacobian
ð29Þ
matrices to a significant degree [11]. It should be noted
that each link’s twist and the link Jacobian matrix Following the same reasoning as the velocity analysis,
formulated based on it, have been defined in the by taking the dot product from both sides of the above
corresponding limb’s coordinate frame. equation by l2 to find h€1 and by s^1 to determine u.€
As a result of the foregoing discussion, the link € €
Consequently, h2 is obtained through substitution of h1
Jacobian matrices for the upper and lower links can be and u€ into Eq. (29). Therefore:
written w.r.t. the corresponding link’s coordinate    2 
2
system as follows: v_ p þ h_ 1 l1  l2  h_ 2 ul
_ 2  ðs^1  ðs^2  l2 Þ þ s^2  ðs^1  l2 ÞÞ þ h_ 2 þ u_ 2
h€1 ¼
2 3 ðl2  l1 Þ  s^1
 033  
vl1 vp vp ð30Þ
¼ 4 T
s1  l2
^ 5 ¼ Jl1
xl1 x p 0 31   2   2 
s^T1  ðl2  l1 Þ v_ p þ h_ 2 þ u_ 2 l2 þ h_ 1 l1  s^1  h_ u^
_ s1  ðs^2  ðs^1  l2 ÞÞ
u€ ¼
ð27Þ ðs^1  s^2 Þ  l2
  ð31Þ
vl2 vp
¼ Jl2   2  
xl2 xp v_ p þ h_ 2 þ u_ 2 l2 þ h_ u_ ðs^1  ðs^2 l2 Þþ s^2  ðs^1 l2 ÞÞ l1 þ u
€ðl1 l2 Þ s^2
2 3 h€2 ¼
133 ðl2 l1 Þ s^1
6 0 " # 1T 7
6 s^T2  ðl2  l1 Þ 7 ð32Þ
6 l1 þ T   s^1 C 7
6 B 7
¼6 s^2  s^T1 B l1  ^l2  s^1 C 7
6 þ ^
s 
1 B
B C 7
6 lT  ðs^2  s^1 Þ lT1  ðl2  s^1 Þ C 7
4 1 @ A 5


3 Dynamics analysis
vp
031 In the literature, various virtual works formulations
ð28Þ has been opted in order to carry out the dynamics
modeling of parallel manipulators [7, 8, 57], and, more
In the above, 133 stands for a three-by-three identity precisely, the 3-DOF Delta parallel robot
matrix. [14, 20, 60, 99]. However, since the ultimate goal of
this paper is to accomplish online inertial parameter
2.4 Acceleration analysis identification, it is desired to determine a compact and
computationally efficient dynamics model for the
In order to develop the dynamics equations of motion, Delta parallel robot. In some researches, screw-based
the second-order kinematics relations are required. virtual works methods have been used in order to
Therefore, by applying second order derivative from determine a compact formulation of the dynamics
loop-closure equation, one has: equations of motion of parallel manipulators which
would be suitable for online identification purposes in
order to employ more precise model-based automatic
control schemes [23, 33]. Therefore, in order to enable
convenient online parameter identification of the 3-
DOF Delta parallel robot, in this section, dynamics
analysis of this manipulator has been addressed by a
screw-based virtual works formulation. In Sect. 2, all
of the governing kinematics relations have been

123
Meccanica (2022) 57:473–506 483

determined in screw-based and compact form. At this 3.1 Extraction of inertial wrenches
moment, the inverse dynamics analysis of the 3-DOF
Delta parallel robot can be asserted by the means of Conducive to making parameter identification meth-
d’Alembert’s formulation of the virtual works princi- ods compatible with this approach, for the purpose of
ple. By assuming that all links are rigid, one has: online identification of the inertial parameters, coor-
 dinate systems are attached to each link in such a way
XN
mi v_ Gi  Fi
d$Tp  JTi ¼ 0: ð33Þ that their center is located on the joints instead of the
i¼1 Ii x_ i þ xi  Ii xi  Ti center of mass of each one. Inertial forces acting on
center of mass of each body are written as:
where d$Tp denotes the variation of the end-effector’s 
twist. As delineated in Sect. 2.3, the link Jacobian mv_ Gi
Finertia
Gi ¼ : ð37Þ
matrix establishes the relation between the twist of the IGi x_ i þ x  IGi x
corresponding link and the twist of the end-effector. Since the ultimate goal of this paper consists in
Twist of links can be regarded as: performing online identification and minimizing the
  calculations’ delay, the center of each link’s coordi-
vi vp
v_i ¼ ¼ Ji : ð34Þ nate frame have been considered for formulating its
xi xp
wrench. It should be noted that these centers could be
Referring to Eq. (34) and since d$Tp 6¼ 0 holds within placed on either side of the links’ joints; nonetheless,
the manipulator’s workspace, with the knowledge that in consideration of simplifying the link Jacobian
matrices, these centers have been situated on the
d$Tp constitutes an independent set, Eq. (33) can be
actuated links’ upper joints and lower joints of the
reformulated as follows:
parallelogram links, as mentioned before. In virtue of
XN  this choice, linear velocities of the upper links’ body
T mi v_ Gi  Fi
Ji ¼ 0: ð35Þ fixed frames would be zero and linear velocities of the
i¼1 Ii x_ i þ xi  Ii xi  Ti
lower links’ body fixed frames would be equal to end-
In order to categorize and simplify the above equation, effector’s velocity; consequently, the link Jacobian
the wrench notation is opted: matrices defined w.r.t. the centers of these coordinate
systems will be simpler and more compact.
X
N   XN
According to Appendix 1, all the terms related to
applied
JTi Finertia
i  F i ¼ JTi
i¼1 i¼1
the center of mass of the links in Eq. (37) have been
"  # ð36Þ eliminated, thereby, it simplifies as:
FInertia
i  Fact
i þ Fext
i 
Inertia
 act  : mðv_ i þ xi  ðxi  ri Þ þ x_ i  ri Þ
Ti  Ti þ Text
i Finertia ¼ :
i
Ii x_ i þ xi  Ii xi þ mðri  v_ i Þ
Reaching this step, each of the wrenches should be ð38Þ
computed in order to determine the dynamics model of
the manipulator. Discernibly, Eq. (36) is comprised of By factorizing the inertia parameters, Eq. (95) can be
two major screws; inertial wrenches and the screw of written in a linear form w.r.t. its inertial parameters:
external forces and torques. The latter type consists of 2 3
" # mi
actuator torques, which is applied by the actuators on v_i D 0 36 6 7
each of the active joints, and gravitational forces plus Finertia
i ¼ _ 4 m i ri 5 ð39Þ
b
031  V i C
externally exerted forces and torques, in form of either li
additive mass, disturbances, or frictional screws. In
in which:
what follows, extraction process of each type is
described. bi X b_ i
bi þ X ð40Þ
D¼X

bi X
C¼X e_ i
ei þ X ð41Þ

where:

123
484 Meccanica (2022) 57:473–506

2 3
0 xz xy corresponding link, in which it is identified as a whole
b ¼6
X 4 xz 0
7
xx 5 ð42Þ by one of the online identification methods. This can
be construed from the fact that an external load is a
xy xx 0
screw; thus, it can be transferred to the center of link
2 3 coordinate frame via the screw transformation matrix.
xx xy xz 0 0 0
6 7 The contribution of this external wrench will be added
e
X¼4 0 xx 0 xy xz 0 5 ð43Þ to that link’s dynamics matrix; therefore, it can be
0 0 xx 0 xy xz deemed as an additional inertial vector added to that
link’s inertial vector, while the dynamics matrix of the
li ¼ ½ Ixx Ixy Ixz Iyy Iyz Izz T ð44Þ corresponding link is regarded as fixed. The latter
capability of the proposed dynamics model, adjoined
It is worth noting that the choice of the link coordinate
by the proposed online base parameters tuning and
frames and their centers has simplified v_ i and V b_ i in anti-windup identification under insufficient excita-
Eq. (39) further, as v_ i ¼ 0 for the upper links, and tions, will be exhibited in Sect. 6.2.
v_ i ¼ v_ p for the lower links. Back to the main point, there are still some
simplifications which lead to obtain a more compact
3.2 Wrenches of the actuators and external loads formulation of the dynamics equation of the Delta
parallel robot. In fact, the contribution of Fext i can be
In this section, actuator torques and external loads considered in the term Fi inertia
, as it possesses the form
acting on each link of the 3-DOF parallel robot are of an inertia. It is assumed that each link has its own
determined. By aligning the y-axis of the limb acceleration plus the acceleration due to the gravita-
coordinate frame (frame attached to the active joint tional field. Hence, by combining the latter two
or the joint with an actuator) with the motor axis, one parameters, one has:
can write motor’s actuation wrench as follows: " #
v_ i  gi D 036
Fact 0 T inertia ext  
i ¼ ½0 0 0 0 si Fi  Fi ¼ b_ i  G
bi
ð45Þ 031  V C
¼ ½0 0 0 0 1 0 T si ¼ $^act
i si 2 3
mi
In the above expression, si is the scalar value of ith 6 7
4 mi ri 5 ¼ Pi Pi
motor’s torque and $^act
i is its corresponding unit screw. li
Therefore, contribution of the manipulator’s actuator
ð48Þ
wrenches in Eq. (36) can be written as:
2 3 In the rearmost equation, gi is the gravitational
X N h i s1 acceleration vector expressed in the ith link’s coordi-
JTi Fact T ^act JT $^act JT $^act 6 7
i ¼ J1 $1 2 2 3 3 4 s2 5 nate frame.
i¼1
¼ ½ c1 c2 c3  s: s3
ð46Þ 3.3 Dynamics equation of motion in linear
regression form
where s is a vector composed of actuator torques.
In order to extract the linear form of the under study Subsequent to derivation of the dynamics model’s
parallel robot’s dynamics equation of motion, it is terms separately, one should merge them to form the
assumed that the wrenches of external loads are only dynamics equation of motion of the 3-DOF Delta
composed of gravitational forces, in other words: parallel robot. As noted in Sect. 3.1, Eq. (48) has been
 derived w.r.t. ith link’s coordinate frame, also, as
ext m i gi
Fi ¼ : ð47Þ discussed in Sect. 2.3, all of the Jacobian matrices
0
have been derived in the limb coordinate systems.
Otherwise, in the case of presence of other external Nevertheless, in order to develop the dynamics
loads, their contribution may be recognized as an equation of the manipulator in form of Eq. (35), these
addition to the inertial parameters vector of the terms ought to be defined in the base coordinate

123
Meccanica (2022) 57:473–506 485

system. To do so, three rotation matrices have been s ¼ J1

s ðJ ðP PÞÞ: ð57Þ
defined; namely, rotation matrix of the base to limb
coordinate systems (B R0i ), rotation matrix of limb Equation (57) represents the dynamics equation of
coordinate frames to the upper links’ (0 R1i ), and motion of the 3-DOF Delta parallel robot in a compact
rotation matrix of limb coordinate frames to the lower matrix form, and linear w.r.t. the manipulator’s inertial
links’ coordinate frames (0 R2i ). These matrices have parameters. It is worth noting that in order to make
been comprehended in Appendix 2. By taking the identification process more convenient and faster, the
foregoing arguments into consideration, Eq. (35) inertial parameters have been described w.r.t. a fixed
would result in the form below: point on the concomitant link, namely the center of
link coordinate frames, and they are to be measured
X
N   XN
directly w.r.t. the foregoing coordinate frames’ axes.
B
RT0i JTi Finertia
i  Fext
i ¼ B T T act
R0i Ji Fi :
With this in mind, the dynamics equation could be
i¼1 i¼1
written in the following form:
ð49Þ
Upon computing each part of Eq. (49), one can write a s ¼ J1 T
s JP P ¼ D P: ð58Þ
compact linear form of the dynamics equation of the In this equation, s is the vector of actuator torques
Delta parallel robot as follows: derived for a single point on the trajectory, DT denotes
  the matrix of regressors for this point and, finally, P
J P PT ¼ J s s ð50Þ
represents the vector of inertial parameters to be
where: identified. Thereafter, vector sT and matrix DTT have
been formed by concatenation of torque vectors and
Js ¼ B
RT01 c1 B
RT02 c2 B
RT03 c3 ð51Þ
matrices of regressors during the identification pro-
cess. Finally, the ultimate form is as follows:

h i
J¼ B
RT01 JT1 B
RT02 JT2 B
RT03 JT3 B
RT01 JT4 B
RT02 JT5 B
RT03 JT6 JTp ð52Þ

h T 0 T 0 T 0 T 0 T 0 T iT
P¼ 0
RT11 P1 RT12 P2 RT13 P3 RT21 P4 RT22 P5 RT23 P6 Pp T ð53Þ

h iT
P ¼ PT1 PT2 . . . PTp ð54Þ sT ¼ DTT P: ð59Þ

" #
v_ i  gi D 036
Pi ¼   ð55Þ
031  Vb_ i  G
bi C 4 Simulation and verification of equations

T In this section, kinematics and dynamics equations are

Pi ¼ m i mi rTi lTi ð56Þ
verified as a preceding to the base parameters analysis.
It is worth mentioning that by concatenation of the Thence, model of the 3-DOF Delta parallel manipu-
inertial parameters vectors for each link, Pi , the lator has been constructed in MATLAB’s Simscape
inertial parameters vectors of the robot, P, have been environment. Moreover, kinematics and dynamics
formulated. Therefore, the final form of the inverse parameters of the Delta parallel robot under study are
dynamics model will be: given in Table 1 in which mcon and mrod denote the
mass of each connector and the mass of each rod of the
parallelogram links, respectively. These parameters

123
486 Meccanica (2022) 57:473–506

Table 1 Upper and lower link’s geometric and inertial of position, velocity, and acceleration of the manip-
parameters of the 3-DOF Delta parallel robot ulator’s actuating joints have been read from Sim-
Actuated link Parallelogram link scape’s ‘‘Rigid Transform Sensor’’ in every time-step.
Subsequently, the position, velocity, and acceleration
Parameter Value and unit Parameter Value and unit
of the end-effector have been derived according the
L1 309 mm L2 595 mm kinematics equations in Sect. 2. Then, according to the
mcon 0:259 kg dynamics model derived in Sect. 3, actuators’ torques
m1 0:781 kg mrod 0:574 kg have been derived and been compared to their
m2 1:666 kg counterparts read from the Simscape model.
Ixx1 3:0e3 kg m 2 Ixx2 5:2e2 kg m2 In what follows, a simulation study is performed to
Iyy1 2:5e2 kg m 2 Iyy2 2:6e1 kg m2 verify the kinematics and dynamics equations of the
Izz1 2:7e2 kg m 2 Izz2 2:2e1 kg m2 extracted model. Without loosing the generality, the
Ixy1 8:8e4 kg m2 Ixy2 2:0e2 kg m2
following joint space trajectory is used:
2
Ixz1 2:0e3 kg m Ixz2 2:8e2 kg m2 h11 ðtÞ ¼ 0:5 sinð3p=4Þ; h12 ðtÞ ¼ 0:4 sinð6p=5Þ;
Iyz1 4:0e3 kg m2 Iyz2 1:2e2 kg m2 h13 ðtÞ ¼ 0:3 sinð5p=3Þ
Base End-effector ð60Þ
R 140 mm mp 0:637 kg which is considered as an input to the simulated
model. In the derived model, first the joint space
inputs’ (actuators) position, velocity and acceleration
were used to calculate the corresponding actuator
were derived from the CAD model of the manipulator
torques as well as the end-effector’s position, velocity
and consequently are used in the analytical and the
and acceleration via the proposed inverse dynamics
Simscape model. An overview of the Simscape model
formulation. Thereafter, all joint variables of the Delta
of the 3-DOF Delta parallel robot has been shown in
parallel robot along their velocity and accelerations
Fig. 4. It is worth mentioning that in order to perform
are compared to the results of the proposed compact
the simulations, first, a feasible joint-space path has
model of the robot. It should be noted that difference
been selected for the actuators. Afterwards, the values
between the foregoing parameters in the derived

Fig. 4 Simscape model of the 3-DOF Delta parallel manipulator

123
Meccanica (2022) 57:473–506 487

Fig. 5 Comparison of the computed torques and simulation results of the 3-DOF Delta parallel robot in the trajectory given in Eq. (60)

equations and the Simscape model is less than 1e-12 insufficient excitation conditions. Equation (57) is
(m). Finally, in Fig. 5, torques of each of the actuators, rewritten as the following linear form:
resulted from the proposed linear form of the virtual
st ¼ DTt Pt þ gt ð61Þ
works principle method, is compared with the corre-
sponding values extracted from MATLAB’s Sim- where st 2 Rk denotes the output of the system, Pt 2
scape. The difference between the corresponding Rn denotes the vector of parameters, and Dt 2 Rnk
torques is not more than 1e-5 (N.m). represents the matrix of regressors comprised of
J1
s J P. It is worth mentioning that measuring the
output vector of the system in the identification
5 SVD-Based damped RLS method
procedure entails a measurement noise in the model,
As discussed in Sect. 3, in order to find the precise full which is represented by gt 2 Rk at time t. The standard
dynamics model formulated in Eq. (57), the inertial RLS method minimizes the following quadratic cost
parameters should be identified. Various approaches function to estimate the parameters:
have been adopted in which the LS method is the basis X
t
T
of the identification procedure. In this regard, most of Jt ¼ ktT ðsT  DTT P^t Þ ðsT  DTT P^t Þ ð62Þ
the studies carried out on the identification of base T¼t0

inertial parameters are based on offline tuning methods. where k and t0 denote the forgetting factor and initial
In the case that external loads or sudden inertial time, respectively. The estimation of parameters P^t 2
parameters changes are applied to the robot, online
Rn is calculated as P^t ¼ R1
t Ft , where
identification is inevitable. In the so-called RLS
method, with the purpose of online identification, X
t
sufficiency of input-output regressors, i.e., having Rt ¼ ktT DT DTT 2 Rnn ð63Þ
T¼t0
bounded condition number of the matrix of regressors,
plays a significant role in parameters estimation of the is the online covariance matrix of regressors, and
dynamics model. In the case of having abundant inertial
parameters, the RLS often faces with insufficient
excitation problem, leading to the estimation windup
of identified inertial parameters. In this section, upon
explaining the novel approach, the online identification
of inertial parameters becomes possible under

123
488 Meccanica (2022) 57:473–506

X
t p p p T
tT n Rt ¼ Vt Kt Vpt ð67Þ
Ft ¼ k D T sT 2 R ð64Þ
T¼t0 p
where Rt is the non-zero part of Rt and the nullity
is the online mutual dependency vector of input-output vector of Rt is the solution of Rt Vnul ¼ 0. Since,
regressors. The recursive format of these equations
X
t
can be written as follows: Rt Vnul ¼ ktT DT DTT Vnul ¼ 0;
T¼t0 ð68Þ
Rt ¼ kRt1 þ DT DTT ; Ft ¼ kFt1 þ DT sT : ð65Þ
! DTT Vnul ¼ 0; 8T ¼ t0 ; . . .; t
Singular value decomposition is one of the most
convenient methods in analyzing the nullity subspace nullity vectors of DT DTT can lead to nullity of the
of matrices in the identification approaches. In fact, covariance matrix if DT DTT ; 8T ¼ t0 ; . . .; t have at least
SVD not only determines the condition number of a one common nullity vector in [t0 ; t]. Therefore, nullity
matrix but also extracts the subspace of the nullity. vectors of the covariance matrix should be constant in
This can be achieved by demonstrating the parts of the this interval to cause rank deficiency. Using the matrix
observation matrix lacking the adequate measurement of eigenvectors Vt 2 Rnn as a transform matrix,
of data and as the result are not utilized in precise regressors are mapped from the current space to an
identification of the model. In another words, in the orthogonal space:
dynamics model of the robot, the linear dependencies " # 
between rows of the observation matrix, which VTnul DT 0
ZT ¼ VTt DT ¼ T
¼ red ; 8T ¼ t0 ; . . .; t
instigate nullity of the covariance matrix and estima- Vpt DT ZT
tion wind-up in the RLS method, are elicited. By so ð69Þ
doing, the identification of the base parameters
becomes practicable. By singular value decomposition where Zt 2 Rnk denotes the matrix of orthogonal
of the covariance matrix Rt , which is a positive semi- regressors and Ztred 2 Rnpk denotes the matrix of
definite matrix, can be formulated as: orthogonal reduced regressors. The cost function can
" #" T # be rewritten based on orthogonal-space variables as
h i K Vnul
p nul 0 follows:
Rt ¼ Vnul Vt p T
|fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl} 0 Kt Vpt ð66Þ
Vt |fflfflfflfflffl{zfflfflfflfflffl} X
t
T
VTt Jt ¼ ktT ðsT  DTT Vt VTt P^t Þ ðsT  DTT Vt VTt P^t Þ
T¼t0
where, the diagonal matrix Knul 2 Rnwnw contains
the first nw -zero eigenvalues of Rt , and the diagonal X
t
T
¼ ktT ðsT  ZTT f^t Þ ðsT  ZTT f^t Þ
matrix Knul 2 Rnpnp contains the rest of eigenvalues T¼t0
(kj [ 0; j ¼ nw þ 1; . . .; n). ð70Þ
Assumption A1: Suppose Rt has nw zero eigenval-
ues. It is assumed that nullity of Rt occurs in the where, the vector of estimated parameters in orthog-
interval [t0 ; t]. That is, Rt0 is a full rank matrix and Rt onal space is denoted by f^t 2 Rn . Based on the
is a matrix with rank deficiency of nw . In the interval definition of the matrix of orthogonal reduced regres-
[t0 ; t], Rt is not invertible and can be written as sors in Eq. (69), the cost function could be simplified
follows: as follows:

123
Meccanica (2022) 57:473–506 489

0 2 31T  
i f^0t
h EtZ red ¼ Zred st  Zred
T
red
f^t1
X
t
T t
ktT @sT  0 Zred 4 5A
T
Jt ¼ T  
^red T ð75Þ
T¼t0 ft ¼ Vpt Dt st  DTt P^t1
0 2 31 |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl}
h i f^0t Et
@sT  0 ZredT 4 5A
T In order to negate the effects of transformations in
f^red t
estimating the output s^t , one should have P^t ¼ Vt f^t .
X
t  T  
T
^red redT ^red Therefore, by remapping into the original space, the
¼ ktT sT  Zred
T f t s T  Z T f t
T¼t0
parameters estimation is calculated as follows:
" #
ð71Þ b
P^t ¼ Vt f^t ¼ Vnul Vt
b p
where the vector of reduced estimated parameters is f^red
t ð76Þ
denoted by f^red 2 Rnp . The solution of minimization
t
¼ Vnul b þ Vpt f^tred ¼ Vnul b þ P^t
could be calculated through the RLS procedure:
In an ideal situation, all estimated parameters should
f^tred ¼ RtZ red FtZ red
1
ð72Þ result in the same estimated output and selecting
where RtZ red 2 Rnpnp denotes the covariance matrix of different values for b should not change the identifi-
cation error. Therefore b can be considered as a design
reduced regressors and FtZ red 2 Rnp denotes the mutual
parameter. By using Vpt as the transformation matrix to
vector of input-output reduced regressors given by the
map the orthogonal space to the original space,
following expressions.
following equations for recursive estimation of
X
t
T
parameters are obtained:
RtZ red ¼ ktT Zred
T ZT
red
 
P^t ¼ Vpt f^tred ¼ Vpt Vpt T P^t1 þ Vpt Rpt 1 Et
T
T¼t0

X
t   ð77Þ
p p
¼ P^t1 þ Rpt 1 Dt st  DTt P^t1 :
T T
¼ ktT Vpt DT DTT Vt ¼ Vpt Rpt Vt ;
T¼t0
ð73Þ Based on the study conducted in [73], the following
X
t
FtZ red ¼ k tT
Zred equation is written by applying a permutation matrix
T sT
T¼t0 to Eq. (76):
X
t 2 01b 3 " # 2 01 3
T T 01
¼ ktT Vpt DT sT ¼ Vpt Ft P^ V P^ t
4 t 5¼ nul
b þ 4 02 5
T¼t0 02b 02
P^t V nul P^ t ð78Þ
The recursive equation of estimated parameters in the |fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl}
0
|fflfflffl{zfflfflffl}
0
0b Vnul P^t
orthogonal space based on the original space param- P^t
eters can be obtained as follows: 02
where V nul 2 Rnwnw is the invertible part of the
f^tred ¼ f^t1
red þ Rt 1 Et ¼ f^red þ
Z red Z red t1
0
permuted nullity matrix, Vnul 2 Rnnw . It is worth
 T 1 T ð74Þ mentioning that the prime symbol ( 0 ) on a parameter
Vpt Rpt Vpt red þ VpT Rp 1 E
Vpt Et ¼ f^t1 t t t denotes the permuted parameter. Consequently, the
following optimal estimation is acquired for the
where permuted inertial parameters in the offline procedure:
2 01 1 02 01
3
0 ^ ¼ V0 1 V0 2
P ^ t þ P^ t
P
P^offline ¼ 4 t nul nul
02
5 ð79Þ
^
Pt ¼ 0
Therefore, one can attain the following matrix:

123
490 Meccanica (2022) 57:473–506

" #
02 1 Additionally, in order to implement the method
H¼ 1 V nul ð80Þ proposed in this paper on other parallel manipulators,
0 0
first, the governing kinematics relations of the manip-
whose rows demonstrate the relations governed ulator should be determined. Then, following the
between original inertial parameters. Also, I 2 selection of link coordinate systems appropriate for
Rnpnp denotes the identity matrix. Nonetheless, after the identification process, the link Jacobian matrices
the offline identification of parameters, autonomous should be determined. Afterwards, one can derive the
devices are obliged to conform to the surrounding matrice of regressors and the inertial parameters
varying conditions. Hereafter, the tuning of parame- vector according to the equations of Sect. 3, leading
ters should be adopted as a necessary online approach. to derivation of the manipulators dynamics model. It is
Despite other methods, the aim here consists in worth mentioning that one can conveniently use
damping oscillations of estimated parameters. These equations presented in Sect. 3 and substitute the
oscillations can cause instability when using least- corresponding kinematics values for their under study
squares method for identification. In order to select an parallel manipulator. It is worth mentioning that the
optimal value for b, the cost function Jb is minimized: only equation that changes in Sect. 3, is the equation
 T   of the actuator twist. In this regard, the unit screw $^act
i
Jb ¼ P^bt  P^t1 P^bt  P^t1 ð81Þ in Eq. (45) should be changed according the type of
the manipulator’s actuators, namely, prismatic actua-
By selecting
tor or revolute actuators.
 1   Reaching this step, an approximate set for offline
b ¼  VTnul Vnul VTnul Rpt 1 Dt st  DTt P^t1
base inertial parameters should be determined. For this
ð82Þ end, first, a path with sufficient excitations should be
determined. It should be noted that it is not required for
as the optimal value for the design parameter, the best the path to have a very low condition number, i.e., be
estimated parameters in order to damp oscillations impeccably optimal; since, as it will be shown in the
under insufficient excitations could be formulated as results presented in Sect. 6, the online base inertial
follows. parameters tuning and identification algorithm can
 1  accomplish the identification procedure successfully

P^online ¼ P^t1 þ I  Vnul VTnul Vnul VTnul without an accurate offline base inertial parameters
  which is determined for a specific path bandwidth.
Rpt 1 Dt st  DTt P^t1
Afterwards, the path should be used to determine the
ð83Þ approximate offline base inertial parameters following
In this methodology, there exists no constraint on the the steps delineated in Sect. 6. Subsequently, using the
degree of persistent excitations, variations of the calculated offline base inertial parameters as an initial
system parameters, and number of estimated param- condition, one can implement the online tuning and
eters. While excitation is sufficient, the nullity of the identification of base inertial parameters method on
covariance matrix remains zero and all estimated the under study parallel manipulator following the
parameters converge to the system parameters. How- steps delineated in Algorithm 1 for any desired path
ever, when the nullity of the covariance matrix with insufficient excitations.
becomes non-zero, Eq. (83) still estimates the optimal
parameters with damped oscillations. Equation (83)
tries to damp oscillations of estimated parameters in
presence of all identification challenges such as slowly
variation of system parameters, disturbances, and
noise. In order to identify statistical properties of the
SVD-DLS, the consistency and efficiency of the
proposed method in the presence of noise are studied
in Appendix 3.

123
Meccanica (2022) 57:473–506 491

6 Base inertial parameters of the dynamic dynamics model of the 3-DOF Delta parallel robot
formulation of the 3-DOF Delta parallel robot derived directly by the means of the screw formulation
of the virtual works method without neglecting any
Linear dynamics equation of motion consists in mechanical bodies or simplifying assumptions. Due to
k equations with n inertial parameters. In order to the topology of over-constraint parallel manipulators,
fully identify this system, at least n equations are or more specifically the Delta robot, some inertial
required for describing n parameters. Therefore, for a parameters have similar effects on the dynamics
prescribed path, a set of data is obtained in form of a behaviour of the manipulator. For the sake of param-
sequence of the actuated joints’ positions, velocities, eter identification, either offline identification or
accelerations, and actuating torques from experi- online identification, the reduced dynamics model
ment—by the Simscape’s model. This data set is then should be derived in order to eliminate these depen-
used to formulate stacked matrix of regressors and dencies and similarities to make the parameter iden-
outputs (torques of the actuators) by the derived tification possible.
dynamics model of the manipulator.
The dilemma which arises here is that the condition 6.1 Offline tuning of base inertial parameters
number of the matrix of regressors should be mini-
mized since it is a measure of dependencies of rows of Base parameters identification falls into two cate-
the regressor matrix w.r.t. the path under study; this gories, offline tuning and online tuning. In the offline
means that the path’s sequence of data is not unique tuning method, the optimal path is considered tanta-
and does not fully excite regressors relating all of the mount to the one proposed in [85], albeit not the same.
parameters. Thus, the condition number of the This path is in form of finite Fourier series which the
normalized matrix of regressors is the criterion to number of its harmonic frequencies is chosen based on
design the best experimental path. the required excitation number (Table 2). Although, in
Prior to delving into the base inertial parameters this research, the harmonic frequencies have been
identification, a sum-up of problem definition and optimized by the genetic algorithm instead of the
differences between dynamics models has been pro- domains, since they represent the excitation and
vided. The ‘‘full dynamics model’’ regards to the

123
492 Meccanica (2022) 57:473–506

Table 2 Frequencies of the optimal path for base inertial similar impact on the output torque, and they can be
parameters determination along their counterparts at the testing grouped to reduce count of the requisite inertial
path
parameters. The LS method can be regarded as a
Actuator 1 (Hz) Actuator 2 (Hz) Actuator 3 (Hz) proper approach in determining the inertial parame-
ters. In this regard, based on the aggregated matrices of
Training path
regressors, or the so-called observation matrix, the
1 1.320, 5 2.783, 8 7.265, 5
covariance matrix Rt is constructed in Eq. (66).
2 3.628, 5 0.877, 7 5.382, 7
Thereafter, the nullity matrix is extracted from the
3 4.905, 1 9.560, 7 2.494, 4
singular value decomposition of the covariance
4 0.609, 5 0.481, 8 3.602, 4
matrix, instigating the base inertial parameters deter-
5 0.365, 5 2.792, 1 3.200, 2
mination. Applying a permutation matrix, V01 nul and
6 0.052, 3 5.692, 6 3.893, 5 02 02 nwnw
Vnul are acquired in such a way that Vnul 2 R is
7 2.899, 2 4.303, 6 4.465, 3
the invertible part of the permuted nullity matrix. As
Testing path
delineated in Sect. 5, one can attain the matrix H from
1 2.356, 2 3.769, 9 5.236, 0
Eq. (80). The rows of the matrix H illustrate the
relations governed between original inertial parame-
ters, implicating the linear combination of them
bandwidth of the path. Equation of the active joint forging the base inertial parameters. Eliminating and
trajectories is as shown below: grouping the parameters will allow one to use
identification methods to determine indispensable
X
N
ak   bk   inertial parameters to demonstrate input torques of
hi ð t Þ ¼ i
k
sin xki t  ik cos xki t ð84Þ
x
k¼1 i
xi the system. In addition, this tedious algebraic manip-
ulation is done with faster pace since the matrix of
X
N     regressors and parameters’ vector are much smaller
h_i ðtÞ ¼ aki cos xki t þ bki sin xki t ð85Þ and more compact.
k¼1
In Tables 3, 4, and 5, base inertial parameters of
X
N     the 3-DOF Delta parallel robot have been represented.
h€i ðtÞ ¼ aki xki sin xki t þ bki xki cos xki t ð86Þ By extraction of the concomitant reduced dynamics
k¼1 model of the manipulator, one can calculate the
In which, xki is an element of a N  1 vector, actuators’ torques. As depicted in Fig. 6, in the offline
containing ith link’s excitation frequencies. Magni-
tude of the frequency xi as well as number of
harmonic frequencies determines the bandwidth of Table 3 Base inertial parameters of limb 1 of the Delta par-
the trajectory. This means that the dynamic response allel robot
of the system will be accurate in the predefined Base Inertial Parameter Combination of Original Parameters
frequencies based on the resultant base parameters.
Pb101 P1 þ P4 þ P14
By determining the optimal path, it follows that not
Pb102 P2
all of the parameters will directly affect the dynamics
Pb103 P3
response of the model, which was obvious from the
Pb104 P6
outset. For instance, zero columns of the dynamics
Pb105 0:5P4 þ 0:5P5 þ P7 þ 0:5P14
matrix corresponds to parameters that do not have
influence on the dynamics response of the manipula- Pb106 P8
tor. For example, the actuated links are connected to Pb107 0:5P4 þ 0:5P5 þ P9 þ 0:5P14
the motor by a revolute joint, ergo its inertia about its Pb108 0:5P4 þ 0:5P5 þ P10 þ 0:5P14
longitude axis does not influence the dynamics Pb109 0:5P4  0:5P5 þ P11  0:5P14
response. Moreover, some group of columns of the Pb110 0:5P4  0:5P5 þ P12  0:5P14
matrix of regressors are linearly dependent; conse- Pb111 0:5P4 þ 0:5P5 þ P13 þ 0:5P14
quently, the parameters associated to them have a

123
Meccanica (2022) 57:473–506 493

Table 4 Base inertial parameters of limb 2 of the Delta par- tuning path, the resultant torques of the actuators have
allel robot been compared to the full dynamics model.
Base Inertial Parameter Combination of Original Parameters
Following extraction of the base inertial parameters
Pb201 P1 þ P4 þ P14 and comparing the resultant torques of the actuators in
Pb202 P2 offline tuning method with the full dynamics model, a
Pb203 P3 þ 0:5P4 þ 0:5P14 simple trajectory with low excitation number is
Pb204 P4 þ P5 þ P14 utilized in order to assess the performance of the
Pb205 P6 offline tuning method. The input for each actuator’s
Pb206 P7 position is a single sine function. As presented in
Pb207 P8 þ P9 Fig. 7, the accuracy of the reduced dynamics model’s
Pb208 P10 torques w.r.t. the full model’s torques are highly
Pb209 0:5P4 þ P11  0:5P14 dependent on the frequency of the corresponding input
Pb210 P12 sine function; Motor 3 had the best result, and motors
Pb211 0:5P4 þ P13  0:5P14 one and two, respectively, had less accurate results.
This fact can be discerned by inspecting the frequen-
cies provided in Table 2. Also, it should be mentioned
Table 5 Base inertial parameters of limb 3 of the Delta par- that the value for the condition number of the matrix of
allel robot regressors in Eq. (61) has been derived as 1:499e þ 03
Base Inertial Parameter Combination of Original Parameters for this path.
There are two major aspects which one can infer
Pb301 P1 þ P4 þ P14
from the offline tuning actuator torques figures: firstly,
Pb302 P2
the desired path is invariable and known. This method
Pb303 P3 þ 0:5P4 þ 0:5P14
is appropriate for the cases in which a desired path is
Pb304 P6
considered and the base parameters tuning stage is
Pb305 0:5P4 þ 0:5P5 þ P7 þ P14
performed by determining an analogous path with
Pb306 P8
sufficient excitation number and bandwidth. The
Pb307 P9
second conclusion from aforementioned figures is that
Pb308 P10 if the excitation number of the path is not sufficient,
Pb309 0:5P4 þ 0:5P5 þ P11 þ 0:5P14 the identification results will not converge to the finite
Pb310 P12 inertial parameter’s value; however, the reduced
Pb311 0:5P4 þ 0:5P5 þ P13 þ 0:5P14 dynamics model would require excitation number of

Fig. 6 Torques computed

by the reduced dynamics
model and the full dynamics
model in the offline base
parameters’ training stage

123
494 Meccanica (2022) 57:473–506

Fig. 7 Torques Computed by the reduced dynamics model and the full dynamics model in the test stage of offline base parameters
determination

33 instead of 70. Furthermore, its computation time is Figure 8 depicts the actuator torques computed
reduced by about 30 percent due to the reduction based on the tuned ORP and their online identification
occurred on the size of the matrices which eases the with the commensurate reduced dynamics model of
implementation of more complex and computationally the 3-DOF Delta parallel manipulator. The desired
demanding online control algorithms. path on this figure is considered as the finite Fourier
series used in the offline tuning method and the
6.2 Online base parameters tuning aforementioned torques are compared to the fully
determined (without identification), full dynamics
The base parameters derived in the offline tuning model of the Delta parallel manipulator.
method force the user to implement a path with Furthermore, the resultant torques, based on the
sufficient excitation number; nevertheless, the desired ORP of the test stage, were compared to their full
paths often do not possess this property. Bearing this in dynamics model counterpart. In Fig. 9, it is clear that
mind and for achieving the goal of online identifica- in contrast to the results of offline base parameters
tion in general desired paths with insufficient excita- tuning, which determined the torques with consider-
tion numbers, the performance of the active base able error value in the test stage, actuator torques
parameters tuning method which was introduced and converge to the real values faster and with higher
interpreted in Sect. 5 is inspected here. It should be precision. Unlike the offline base parameters, the
noted that in the online tuning method, base inertia online base parameters tuning and identification,
parameters formulated in the offline tuning method is regardless of the excitation number of inputs, can
used as an initial condition for the online base inertial successfully identify the essential inertial parameters.
parameters tuning and identification, or in other It should be noted that the offline base parameters
words, the determination of Online Relevant Param- tuning method can be used in the situations that the
eters (ORP) [26]. In order to shed light on the matter, it desired path is identical to the training path. In these
should be pointed out that the reduced dynamics scenarios, the actuator torques computed by the
model and reduced parameters’ combinations are used reduced dynamics model are equal to the full dynam-
without considering their real values for initial con- ics model. As aforementioned, there are two promi-
dition of the online identification. Based on the nent advantages in this method compared to the full
recursive formulation described in Eq. (83), the opti- dynamics model. Firstly, since the training path and

mal ORP, P^online , are estimated in which P^t is the desired path are analogous with each other, it
calculated recursively based on Eq. (77). allows one to implement online identification methods

123
Meccanica (2022) 57:473–506 495

Fig. 8 Torques computed conjointly with online identification and online base parameters tuning in contrast to the non-identified full
dynamics model in the path with sufficient excitation number

Fig. 9 Torques Computed conjointly with online identification and online Base Parameters tuning in contrast to the non-identified full
dynamics model in the path with insufficient excitation number

and keep track of the changes in base inertial 6.3 Performance evaluation of the proposed
parameters of the manipulator resulted by external method compared to the other in-literature
loads and faults. On the other hand, tedious compu- methodologies with insufficient excitations
tations for determining the required torques would be
reduced dramatically since the reduced dynamics In order to investigate the effectiveness of the
model has less than half of the full model’s dynamics proposed online base inertial parameters tuning and
terms and corresponding inertial parameters. identification approach, its performance has been
compared to those of the other techniques. For this
end, performance of two prominent approaches which
can be used to perform online parameter identification
under insufficient excitations have been investigated.

123
496 Meccanica (2022) 57:473–506

Fig. 10 Torques identified

with the Regularization
algorithm in contrast to the
non-identified full dynamics
model in the path with
insufficient excitation
number

Fig. 11 Torques identified with the GDLS algorithm in contrast to the non-identified full dynamics model in the path with insufficient
excitation number

The first approach is the algorithm proposed in [89], an In what follows, the regularization algorithm and
algorithm based on the LS method with smoothing the GDLS algorithm have been implemented in
conditions and adaptive regularization of parameters. simulations to carry out online parameter identifica-
In this paper, this method is regarded as the Regular- tion under insufficient excitations. In order to conve-
ization method. The second method is a generalized niently compare the results with the proposed
version of DLS, called GDLS, which was proposed in method’s results, the path selected for performing
[97]. This algorithm systematically refines the esti- the simulations with the Regularization and the GDLS
mation windup in a closed-loop control strategy method is identical with the one used in the test stage
[6, 67, 100]. More details about these algorithms can of the proposed method in this paper. In Figs. 10 and
be found in the Introduction section; alternatively, the 11, performance of the identification by the Regular-
reader can refer to the original papers for more in- ization algorithm and the GDLS algorithm, respec-
depth details. tively, in contrast to the non-identified full dynamics
model have been depicted. Additionally, in order to

123
Meccanica (2022) 57:473–506 497

Table 6 Results of the RMSE for different methods

Offline base inertial parameters Regularization GDLS Online base inertial parameters tuning and
identification method method method identification method

J RMSE 0.6327 0.3896 0.1821 0.0315

establish an explicit comparison between the identi- been studied. First, in the midst of online tuning and
fication performances, a quantitative index has been identification of base inertial parameters under insuf-
utilized. This index has been defined as follows: ficient excitations, an extrinsic mass has been applied
X  on the Delta parallel robot. Secondly, the performance
J RMSE ¼ siden  sref 2 ð87Þ
T T of the proposed algorithm has been investigated in
in which sT have been derived according to Eq. 59; presence of noise effects while under insufficient
also, the superscript ‘‘iden’’ or ‘‘ref’’ indicate the excitations.
corresponding identified torque versus the actuator
torques derived by the full dynamics model of the 6.4.1 Performance of the online base parameters
manipulator. The RMSE values for different methods tuning and identification method in presence
have been represented in Table 6. of extrinsic mass
According to Table 6, and by comparing the results
presented in Figs. 9, 10, and 11, it can be observed that In this section, an extrinsic mass has been imple-
the online base inertial parameters tuning and identi- mented on the Delta parallel robot. The extrinsic mass
fication method achieved the best performance among can be implemented on any link of the manipulator or
the other methods. The reason for this results is that the its end-effector, causing a change in the value of
online base inertial parameters tuning and identifica- inertial parameters as well as the base inertial param-
tion method derives a set of base inertial parameters in eters’ formulation. In this investigation, the extrinsic
every iteration, namely the ORP, according to the mass has been implemented on one of the parallelo-
online identification’s path. Therefore, the selected gram links of the Delta parallel robot; since, imple-
ORP would be an appropriate set of parameters for that mentation of the extrinsic mass on the parallelogram
specific identification path, and the identification links affects a higher number of base inertial param-
process can be performed more conveniently. Also, eters of the system than to be implemented on the
this will result in a reduced number of base parameters actuated links or the manipulator’s end-effector. It is
(or the ORP) to be identified, leading to reducing the worth mentioning that the extrinsic masses have been
required number for the excitations to be considered as implemented on L32 of the parallelogram links and at the
sufficient. On the other hand, the other methods seek to instant which the simulation reaches the half of its
provide a sufficient excitation number for the identi- total time.
fication process by ameliorating the path’s excitation Performance of the proposed method is limned in
number via different techniques; however, accom- Fig. 12. The online base parameters tuning and
plishing this goal might be an arduous task in cases identification method’s torques vary in the moment
where the excitation number of the online identifica- of implementing an extrinsic mass and they converge
tion’s path is far less than what is considered as a to the unidentified full model of dynamics model with
sufficient excitation one. high accuracy, in spite of deficiency of the excitation.
In addition to online identification capabilities, the
6.4 Investigation of the effectiveness of the online aforementioned property would allow one to use this
base parameters tuning and identification method in fault tolerant control schemes. Addition of
method under external disturbances external load to the manipulator can be regarded as
fault in the dynamics parameters, which the results
As the next step, in order to examine the proposed showed that the proposed method can compensate the
method’s capabilities, two different scenarios have

123
498 Meccanica (2022) 57:473–506

Fig. 12 Torques computed conjointly with online identification and online Base Parameters tuning in contrast to the non-identified full
dynamics model in presence of an extrinsic mass, implemented in the midst of simulation

fault and can track the required actuator torques while vector for the Nth iteration can be shown as: s ¼
maintaining minimal jump compensation. PN
i¼N9 si for i ¼ 1; 2; 3; where, si is the unfiltered
identified vector of actuator torques at the ith iteration.
6.4.2 Performance of the online base parameters Figure 14 represents the identified torque results
tuning and identification method in presence without the moving-average low-pass filter. Resultant
of sensor noise torque diagrams have been depicted in Fig. 15; where,
one can see that the algorithm remained stable and
In practice, the sensor measurements’ readings may be converged to the desired torque values while under
noisy and inaccurate. Therefore, in this section,
insufficient excitations with J RMSE ¼ 1:3167.
measurement noise have been considered for the
It should be noted that due to the dynamics
sensor readings and the online base inertial parameters
equations of motion of the manipulator, the effects
identification and tuning has been carried out in
of zero-mean white noises on the position, velocity,
presence of the measurement noise. As previously
and acceleration sensors are amounts to a nonzero-
mentioned in Sect. 4, position, velocity, acceleration,
mean colored noise on the derived torques. The
and actuation torque values of the actuated joints have
colored noise maintains the independent and identi-
been obtained from the sensors in an online manner.
cally distributed (iid) characteristic of the sensors’
Four different white noises have been considered for
white noise, although do not conserve their zero-mean
these readings; where the position, velocity, acceler-
property. Since the white noise on the aforementioned
ation, and torque sensors’ noise ranged between
  deg sensors results in a colored noise on the calculated
1:2ðdegÞ, 2:9 deg s , 14:5 s2 and 0:25ðN:mÞ, torques, the RMSE value of the proposed algorithm
respectively. under the noise effects has been larger than without
Figure 13a–c represent, respectively, the position, noise. Additionally, as shown in the Appendix 3, the
velocity, and acceleration of the actuated joints in white noise on the torque sensors (system outputs)
presence of the measurement noise. Subsequently, the does not affect the estimation and the estimation of
online base inertial parameter tuning and identification deterministic parameter vector has been obtained,
has been carried out in presence of the aforementioned which is equivalent with the noise-free estimation of
noises, and a simple moving-average low-pass filter parameters.
with a moving window of 10 iterations has been
implemented on the resultant torques. The formula of
the moving-average low-pass filtered identified torque

123
Meccanica (2022) 57:473–506 499

Fig. 13 Noisy sensor

readings of the position,
velocity, and acceleration of
the manipulator’s actuated
joints

123
500 Meccanica (2022) 57:473–506

Fig. 14 Unfiltered actuator torques identified with online Base Parameters tuning method in contrast to the non-identified full dynamics
model in presence of sensor noise on position, velocity, acceleration, and torque sensors

Fig. 15 Filtered actuator

torques identified with
online base parameters
tuning method in contrast to
the non-identified full
dynamics model in presence
of sensor noise on position,
velocity, acceleration, and
torque sensors

7 Conclusion utilizing the screw theory. Then, the screw-based link

Jacobian matrices of each link, establishing the
In the presented paper, firstly, a linear dynamics relation between twist of the end-effector and the
equation of motion of the 3-DOF Delta parallel robot corresponding link’s twist, were determined conve-
was elaborated w.r.t. (with respect to) its inertial niently. Finally, the dynamics model of the Delta
parameters with the goal of enabling online identifi- parallel manipulator was extracted w.r.t. each link’s
cation. The dynamics analysis was based on the screw coordinate frame, making the inertial parameters’
theory, and the dynamics modeling has been carried identification possible. Afterwards, the concept of
out by resorting to a screw-based formulation of the base inertial parameters was examined and two
virtual works principle. In order to sum-up the methods for tuning the base parameters were delin-
delineated method, following the screw analysis of eated. At the start, offline tuning of the base param-
the manipulator’s geometry, kinematic analysis of the eters was performed and its results were analyzed.
Delta robot was performed in a compact form by Reaching this step, the merits and drawbacks of the

123
Meccanica (2022) 57:473–506 501

aforementioned method were inspected, leading to in practice, performance of the proposed method has
proposing a solution to avoid windup in recursive been investigated in a series of simulations. However,
parameter estimation in periods of insufficient excita- from the the presented results derived from the
tions. Unlike the classic DLS or GDLS, the proposed simulations in the last section of the manuscript, the
method solved the problem of matrix singularity by effectiveness of the proposed method under sensors’
mapping the covariance matrix to an orthogonal space measurement noise or unanticipated external masses,
and removing the zero eigenvalues. Remapping to the while lacking sufficient excitations for the identifica-
original space, the optimal solution to damp oscilla- tion process can be inferred.
tions of recursively estimated parameters was
Declarations
selected. By considering the results of offline base
parameters tuning as initial conditions, the introduced Conflict of interest The authors declare that they have no
method was used to tune the base parameters in an conflicts of interest, neither financial nor non-financial, relevant
online manner. Such a novel estimation method is also to the contents of this article.
applicable to any online estimation procedure under
the insufficient excitation conditions. As the final
stage, simulation results revealed the efficacy of the Appendix
proposed method to perform online base inertial
parameters identification under insufficient excita- Wrenches of Inertia with respect to the link
tions. Also, the performance of the online base inertial coordinate frames
parameters tuning and identification method has been
investigated in comparison with the other online Inertial wrenches of a body, defined in Eq. (36), have
identification techniques proposed to address the been determined in Eq. (37) w.r.t. the center of gravity
insufficient excitations problem. The effectiveness of of the corresponding link. However, from the theory of
the proposed algorithm can be deduced from the second order infinitesimal rotations, one can write
comparisons’ results. It should be noted that the acceleration of the center of mass of each link relative
dynamics analysis and base inertial parameters’ online to another point on the corresponding link. Further-
tuning and identification of the 3-DOF Delta parallel more, from Euler’s rotation equation of motion, it can
robot have been carried out regardless of the type and be written relative to an arbitrary point on the rigid
model of the actuators. In this regard, the actuator body. Therefore, taking these concepts into account,
torques acting on the actuated links of the manipulator Eq. (37) would be:
are considered as the input torques to the robot, and the 
mðv_ i þ xi  ðxi  ri Þ þ x_ i  ri Þ
identification process has been performed on the Finertia
Gi ¼ :
IGi x_ i þ xi  IGi xi þ mðri  v_ Gi Þ
foregoing system. The actuators may be considered as
a system, where its dynamics model could be identi- ð88Þ
fied to establish the governing relations between input As it is can be observed from the above, the two
or desired torques and the output or the torques exerted equations constitute the wrenches of inertia; first one
on the manipulator’s actuated links. As ongoing as Newton’s second law of motion and the second one
works, the compact dynamics equations open an as Euler’s rotation equation; both defined w.r.t. an
avenue to implement nonlinear-adaptive control arbitrary location, namely, center of the links’ coor-
schemes on the robot, followed by extending the dinate frames. However, the second equation contains
proposed online base parameters determination and some terms regarding the center of mass which should
tuning method to fault-tolerant control where, on this be eliminated. Thus, the inertial tensor of each link has
regard, some hints were provided at the end of the been written w.r.t. the center of its own link coordinate
paper. Moreover, it is worth mentioning that this paper frame:
concerns with proposing the method of online tuning    
and identification of base inertial parameters under Ii ¼ IGi þ mi rTi  ri 133  ri  rTi : ð89Þ
insufficient excitations. Therefore, as a limitation of
Substitution of Eq. (89) into second equation of
this paper and a prerequisite of the implementation of
Eq. (88) leads to:
the method with precise automatic control algorithms

123
502 Meccanica (2022) 57:473–506

      2 3
IGi x_ i ¼ Ii  mi rTi  ri 133  ri  rTi x_l cos h1i 0  sin h1i
0 crd 6 7
ð90Þ R1i ðh1i Þ ¼ 4 0 1 0 5 ð97Þ
   sin h1i 0 cos h1i
xi  IGi xi ¼ xi  Ii  mi rTi  ri 133
   ð91Þ 2 3
 ri  rTi xi cos / sin / 0
0 6 7
Rcrd
2i ðh2i ; /i Þ ¼ 4  sin / cos / 05
ri  v_ Gi ¼ ri  ðv_ i þ xi  ðxi  ri Þ þ x_ i  ri Þ 0 0 1
2 3 ð98Þ
ð92Þ cos h2i 0  sin h2i
6 7
Moreover, from linear algebra one has: 4 0 1 0 5
  sin h2i 0 cos h2i
a  ðB  aÞ ¼ aT a 133  a aT B ð93Þ
  According to the coordinate rotation matrices, the
a  ðB  ðB  aÞÞ ¼ B  aT a 133  a aT B Screw rotation matrices have been derived as:
ð94Þ " #
0 crd
B R1i 033
R0i ¼ ð99Þ
where a is an arbitrary 3-by-1 vector, and B denotes a 033 0 Rcrd
1i
3-by-3 arbitrary matrix.
" #
Substitution of Eqs. (90) to (92) into Eq. (88) and 0
Rcrd 033
0 1i
Considering Eqs. (93) and (94), the analogues terms R1i ¼ 0 crd
ð100Þ
033 R1i
with opposite sign appear in Eq. (88) that by elimi-
nating them one can rewrite Eq. (37) as follows: " #
 0
0
Rcrd
2i 033
mðv_ i þ xi  ðxi  ri Þ þ x_ i  ri Þ R2i ¼ 0 crd
ð101Þ
inertia 033 R2i
Fi ¼ :
Ii x_ i þ xi  Ii xi þ mðri  v_ i Þ
ð95Þ
Statistical Properties of SVD-DLS
In this equation, the terms regarding the center of
gravity have been eliminated and it is more convenient In this section, the statistical properties of the
to use it in the identification process. proposed SVD-DLS method and classic LS are
compared. To the end of investigating the effects of
Screw rotation matrices used in this paper noise, the system output, st , is assumed to be
contaminated with additive zero-mean white noise
In this Appendix the rotation matrices used in this vector, gt , with variance r. The noise vector, denoted
paper have been illustrated. As discussed in Sect. 3.3, with Nt , contains kt independent and identically
there are three types of coordinate frames, namely, the distributed (iid) random variables. Therefore, the
limb coordinate frames, actuated links’ coordinate outputs of system can be modeled by:
frames, and parallelogram links’ coordinate frames. In
what follows, the rotation matrices associated with st ¼ st þ gt ¼ DTt Pt þ gt ð102Þ
these coordinate frames have been elaborated, after-
where st denotes the deterministic part of the output.
wards, the Screw rotation matrices have been defined.
The non-recursive parameter estimation in LS method
The coordinate rotation matrices are as follows:
2 3 is calculated as follows:
cos w sin w 0  1
B crd 6
R0i ðwÞ ¼ 4  sin w cos w 0 5
7
ð96Þ P^t ¼ Xt T Xt Xt T Yt ð103Þ

0 0 1 where Xt 2 Rktn shows the augmented matrix of

input regressors and is not affected by the additive
noise. Yt 2 Rkt denotes the augmented vector of
output measurements. Using the deterministic part of

123
Meccanica (2022) 57:473–506 503

the output (Y  t ), the estimation of deterministic 2. Abo-Shanab RF (2019) Dynamic modeling of parallel
 1 manipulators based on Lagrange–d’Alembert formulation
parameter vector, Pt ¼ Xt T Xt Xt T Y  t , is obtained and Jacobian/Hessian matrices. Multibody Syst Dyn
which is equivalent with the noise-free estimation of 48:403–426
parameters. Applying the proposed method leads to 3. Angeles J (2002) Fundamentals of robotic mechanical
systems. Springer, Berlin
the following estimation for the deterministic part of 4. Ansari-Rad S, Jahandari S, Kalhor A, Araabi BN (2018)
parameters: Identification and control of mimo linear systems under
  sufficient and insufficient excitation. In: 2018 Annual
p p T 1 T  American Control Conference (ACC). IEEE,
P^bt ¼ Vt Kt Vpt Xt Yt þ Vnul b
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} pp 1108–1113
5. Ansari-Rad S, Kalhor A, Araabi BN (2019) Partial iden-
Pbt ð104Þ tification and control of mimo systems via switching linear
  reduced-order models under weak stimulations. Evol Syst
p p T 1 T
þ Vt Kt Vpt Xt Nt 10(2):111–128
6. Argha A, Ye L, Cao K, Su SW, Celler BG (2017) Real-
time identification of heart rate responses via auxiliary-
The expected value of P^bt could be calculated as
model-based damped RLS scheme. In: 2017 39th annual
follows: international conference of the IEEE Engineering in
  Medicine and Biology Society (EMBC). IEEE,
p p T 1 T
EfP^bt jXt g ¼ Pbt þ Ef Vt Kt Vpt Xt Nt g ¼ Pbt pp 1312–1315
7. Arian A, Danaei B, Masouleh MT, Kalhor A (2017)
ð105Þ Dynamic modeling and base inertial parameters determi-
nation of 3-DOF planar parallel manipulator. In: 2017 5th
Since the noise is assumed to be zero-mean and iid, RSI International Conference on Robotics and Mecha-
SVD-DLS leads to an unbiased estimation for any b. tronics (ICRoM). IEEE, pp 546–551
8. Asgari M, Ardestani MA (2015) Dynamics and improved
Thereafter, computing the covariance matrix of the computed torque control of a novel medical parallel
estimated vector leads to: manipulator: applied to chest compressions to assist in
  cardiopulmonary resuscitation. J Mech Med Biol
p p T 1 T
covfP^bt jXt g ¼ Vt Kt Vpt Xt EfNt Nt T g 15(04):1550051
9. Åström KJ, Wittenmark B (2013) Adaptive control.
    Courier Corporation, Chelmsford
p p T 1 p p T 1
Xt Vt Kt Vpt ¼ r Vt Kt Vpt : 10. Azad FA, Rahimi S, Yazdi MRH, Masouleh MT (2020)
Design and evaluation of adaptive and sliding mode con-
ð106Þ trol for a 3-DOF delta parallel robot. In: 2020 28th Iranian
p p pT Conference on Electrical Engineering (ICEE). IEEE,
It follows from Eq. (66) that Rt ¼ Vt Kt Vt . There- pp 1–7
fore, when invasive measurements are available, the 11. Azad FA, Yazdi MRH, Masouleh MT (2019) Kinematic
p and dynamic analysis of 3-DOF delta parallel robot based
eigenvalues of inverse of Vt are decreased and as a on the screw theory and principle of virtual work. In: 2019
result, SVD-DLS estimations remain consistent. 5th conference on Knowledge Based Engineering and
A similar procedure can be conducted to show that Innovation (KBEI). IEEE, pp 717–724
the covariance matrix of estimated parameters is equal 12. Baran EA, Ozen O, Bilgili D, Sabanovic A (2019) Unified
kinematics of prismatically actuated parallel delta robots.
to the Cramer-Rao lower bound. It follows that the Robotica 37:1513–1532
proposed method is efficient when the additive noise is 13. Barreto JP, Corves B (2019) Resonant delta robot for pick-
white Gaussian. Consequently, the SVD-DLS method and-place operations. In: IFToMM World Congress on
has similar facets of the classic LS from the statistical Mechanism and Machine Science. Springer,
pp 2309–2318
point of view, and in addition, solves problems of 14. Brinker J, Corves B, Takeda Y (2018) Kinematic perfor-
estimation windup and insufficient excitations. mance evaluation of high-speed delta parallel robots based
on motion/force transmission indices. Mech Mach Theory
125:111–125
15. Cao W, Yang D, Ding H (2018) A method for stiffness
analysis of overconstrained parallel robotic mechanisms
References with Scara motion. Robot Comput Integr Manuf
49:426–435
1. Abeywardena S, Chen C (2017) Inverse dynamic mod- 16. Chen CT, Liao TT (2016) Trajectory planning of parallel
elling of a three-legged six-degree-of-freedom parallel kinematic manipulators for the maximum dynamic load-
mechanism. Multibody Syst Dyn 41(1):1–24 carrying capacity. Meccanica 51(8):1653–1674

123
504 Meccanica (2022) 57:473–506

17. Cheng G, Shan X (2012) Dynamics analysis of a parallel 35. Ghaedrahmati R, Raoofian A, Kamali A, Taghvaeipour A
hip joint simulator with four degree of freedoms (3r1t). (2019) An enhanced inverse dynamic and joint force
Nonlinear Dyn 70(4):2475–2486 analysis of multibody systems using constraint matrices.
18. Clavel R (1988) Delta, a fast robot with parallel geometry, Multibody Syst Dyn 46(4):329–353
vol 26–28 36. Gharahsofloo A, Rahmani A (2015) An efficient algorithm
19. Clavel R (1990) Device for the movement and positioning for workspace generation of delta robot
of an element in space 37. Gherman B, Pisla D, Vaida C, Plitea N (2012) Develop-
20. Codourey A (1998) Dynamic modeling of parallel robots ment of inverse dynamic model for a surgical hybrid par-
for computed-torque control implementation. Int J Robot allel robot with equivalent lumped masses. Robot Comput
Res 17(12):1325–1336 Integr Manuf 28(3):402–415
21. Codourey A, Burdet E (1997) A body-oriented method for 38. Goldenberg AA, He X, Ananthanarayanan S (1992)
finding a linear form of the dynamic equation of fully Identification of inertial parameters of a manipulator with
parallel robots. In: Proceedings of international conference closed kinematic chains. IEEE Trans Syst Man Cybern
on robotics and automation, vol 2. IEEE, pp 1612–1618 22(4):799–805
22. Codourey A, Clavel R, Burckhardt C (1992) Control 39. Grotjahn M, Heimann B, Abdellatif H (2004) Identifica-
algorithm and controller for the direct drive delta robot. In: tion of friction and rigid-body dynamics of parallel kine-
Robot control 1991. Elsevier, pp 543–549 matic structures for model-based control. Multibody Syst
23. Danaei B, Arian A, Masouleh MT, Kalhor A (2017) Dyn 11(3):273–294
Dynamic modeling and base inertial parameters determi- 40. Ha QP, Rye DC, Durrant-Whyte HF (1999) Fuzzy moving
nation of a 2-DOF spherical parallel mechanism. Multi- sliding mode control with application to robotic manipu-
body Syst Dyn 41(4):367–390 lators. Automatica 35(4):607–616
24. Dastjerdi AH, Sheikhi MM, Masouleh MT (2020) A 41. Hong J, Yamamoto M (2009) A calculation method of the
complete analytical solution for the dimensional synthesis reaction force and moment for a delta-type parallel link
of 3-DOF delta parallel robot for a prescribed workspace. robot fixed with a frame. Robotica 27(4):579–587
Mech Mach Theory 153:103991 42. Hui J, Pan M, Zhao R, Luo L, Wu L (2018) The closed-
25. Deng X, Yin L, Peng S, Ding M (2015) An iterative form motion equation of redundant actuation parallel robot
algorithm for solving ill-conditioned linear least squares with joint friction: an application of the Udwadia-Kalaba
problems. Geodesy Geodyn 6(6):453–459 approach. Nonlinear Dyn 93(2):689–703
26. Diaz-Rodriguez M, Valera A, Mata V, Valles M (2012) 43. Iriarte X, Ros J, Mata V, Aginaga J (2017) Determination
Model-based control of a 3-DOF parallel robot based on of the symbolic base inertial parameters of planar mech-
identified relevant parameters. IEEE/ASME Trans anisms. Eur J Mech A Solids 61:82–91
Mechatron 18(6):1737–1744 44. Kalani H, Rezaei A, Akbarzadeh A (2016) Improved
27. Ebrahimi S, Kövecses J (2010) Unit homogenization for general solution for the dynamic modeling of Gough-
estimation of inertial parameters of multibody mechanical Stewart platform based on principle of virtual work.
systems. Mech Mach Theory 45(3):438–453 Nonlinear Dyn 83(4):2393–2418
28. Eksioglu EM, Tanc AK (2011) RLS algorithm with con- 45. Karbasizadeh N, Zarei M, Aflakian A, Masouleh MT,
vex regularization. IEEE Signal Process Lett Kalhor A (2018) Experimental dynamic identification and
18(8):470–473 model feed-forward control of Novint Falcon haptic
29. Eldén L (1984) An efficient algorithm for the regulariza- device. Mechatronics 51:19–30
tion of ill-conditioned least squares problems with trian- 46. Kelaiaia R (2017) Improving the pose accuracy of the delta
gular Toeplitz matrix. SIAM J Sci Stat Comput robot in machining operations. Int J Adv Manuf Technol
5(1):229–236 91(5–8):2205–2215
30. Enferadi J, Tootoonchi AA (2010) Inverse dynamics 47. Kelley C, Ipsen I, Pope S (2010) Rank-deficient and ill-
analysis of a general spherical star-triangle parallel conditioned nonlinear least squares problems. In: Pro-
manipulator using principle of virtual work. Nonlinear ceedings of 2010 East Asian SIAM Conf
Dyn 61(3):419–434 48. Khalifa A, Fanni M, Mohamed AM (2017) Geometri-
31. Evestedt M, Medvedev A (2006) Stationary behavior of an cal/analytical approach for reciprocal screw-based singu-
anti-windup scheme for recursive parameter estimation larity analysis of a novel dexterous minimally invasive
under lack of excitation. Automatica 42(1):151–157 manipulator. Robot Auton Syst 98:56–66
32. Gallardo-Alvarado J (2020) A Gough-Stewart parallel 49. Khalil W, Bennis F (1995) Symbolic calculation of the
manipulator with configurable platform and multiple end- base inertial parameters of closed-loop robots. Int J Robot
effectors. Meccanica 55(3):597–613 Res 14(2):112–128
33. Gallardo-Alvarado J, Aguilar-Nájera C, Casique-Rosas L, 50. Khalil W, Kleinfinger JF (1987) Minimum operations and
Pérez-González L, Rico-Martı́nez J (2008) Solving the minimum parameters of the dynamic models of tree
kinematics and dynamics of a modular spatial hyper-re- structure robots. IEEE J Robot Autom 3(6):517–526
dundant manipulator by means of screw theory. Multibody 51. Khosla PK (1986) Real-time control and identification of
Syst Dyn 20(4):307–325 direct-drive manipulators (robotics)
34. Gallardo-Alvarado J, Balmaceda-Santamarı́a AL, Cas- 52. Kubin G (1988) Stabilization of the RLS algorithm in the
tillo-Castaneda E (2014) An application of screw theory to absence of persistent excitation. In: 1988 International
the kinematic analysis of a delta-type robot. J Mech Sci Conference on Acoustics, Speech, and Signal Processing,
Technol 28(9):3785–3792 1988. ICASSP-88. IEEE, pp 1369–1372

123
Meccanica (2022) 57:473–506 505

53. Kumar KK, Srinath A, Siddhartha B, Subhash LV (2015) 72. Ros J, Iriarte X, Mata V (2012) 3d inertia transfer concept
Simulation and analysis of parallel manipulator for and symbolic determination of the base inertial parame-
manoeuvring laparoscopic camera-CAD based approach. ters. Mech Mach Theory 49:284–297
Int J Eng Technol 7(1):294–302 73. Ros J, Plaza A, Iriarte X, Aginaga J (2015) Inertia transfer
54. Lambert E (1987) Process control applications of long- concept based general method for the determination of the
range prediction. Ph.D. thesis, University of Oxford base inertial parameters. Multibody Syst Dyn
55. Lara-Molina FA, Dumur D (2021) Robust multi-objective 34(4):327–347
optimization of parallel manipulators. Meccanica 74. Ruiz-Hidalgo N, Blanco Ortega A, Abúndez Pliego A,
56:2843–2860 Colin-Ocampo J, Alcocer Rosado W (2019) Dynamic
56. Laski PA, Takosoglu JE, Blasiak S (2015) Design of a analysis and control of a three-revolute-prismatic-spheri-
3-DOF tripod electro-pneumatic parallel manipulator. cal parallel robot by algebraic parameters identification.
Robot Auton Syst 72:59–70 Int J Adv Robot Syst 16(3):1729881419841533
57. Li Y, Xu Q (2005) Dynamic analysis of a modified delta 75. Saafi H, Laribi MA, Zeghloul S (2017) Optimal torque
parallel robot for cardiopulmonary resuscitation. In: 2005 distribution for a redundant 3-RRR spherical parallel
IEEE/RSJ international conference on intelligent robots manipulator used as a haptic medical device. Robot Auton
and systems. IEEE, pp 233–238 Syst 89:40–50
58. Liu XJ, Wang J, Oh KK, Kim J (2004) A new approach to 76. Sadikovic R, Korba P, Andersson G (2006) Self-tuning
the design of a delta robot with a desired workspace. J In- controller for damping of power system oscillations with
tell Robot Syst 39(2):209–225 facts devices. In: Power Engineering Society General
59. Liu Y, Liang B, Xu W, Wang X (2019) A method for Meeting, 2006. IEEE, p 6
measuring the inertia properties of a rigid body using 77. Shah SL, Cluett WR (1991) Recursive least squares based
3-URU parallel mechanism. Mech Syst Signal Process estimation schemes for self-tuning control. Can J Chem
123:174–191 Eng 69(1):89–96
60. Lu XG, Liu M, Liu JX (2017) Design and optimization of 78. Sharifzadeh M, Arian A, Salimi A, Masouleh MT, Kalhor
interval type-2 fuzzy logic controller for delta parallel A (2017) An experimental study on the direct & indirect
robot trajectory control. Int J Fuzzy Syst 19(1):190–206 dynamic identification of an over-constrained 3-DOF
61. Merlet J (2012) Parallel robots, vol 74. Springer, Berlin decoupled parallel mechanism. Mech Mach Theory
62. Mo J, Shao ZF, Guan L, Xie F, Tang X (2017) Dynamic 116:178–202
performance analysis of the x4 high-speed pick-and-place 79. Sharifzadeh M, Masouleh MT, Kalhor A (2017) On
parallel robot. Robot Comput Integr Manuf 46:48–57 human-robot interaction of a 3-DOF decoupled parallel
63. Mokled E, Chartouni G, Kassis C, Rizk R (2019) Parallel mechanism based on the design and construction of a novel
robot integration and synchronization in a waste sorting and low-cost 3-DOF force sensor. Meccanica
system. In: Mechanism, machine, robotics and mecha- 52(10):2471–2489
tronics sciences. Springer, pp 171–187 80. Sharifzadeh M, Masouleh MT, Kalhor A, Shahverdi P
64. Müller A (2022) Dynamics of parallel manipulators with (2018) An experimental dynamic identification & control
hybrid complex limbs—-modular modeling and parallel of an overconstrained 3-DOF parallel mechanism in
computing. Mech Mach Theory 167:104549 presence of variable friction and feedback delay. Robot
65. Nguyen VS, Im N (2014) Development of computer pro- Auton Syst 102:27–43
gram for solving astronomical ship position based on circle 81. Shen H, Meng Q, Li J, Deng J, Wu G (2021) Kinematic
of equal altitude equation and svd-least square algorithm. sensitivity, parameter identification and calibration of a
J Navig Port Res 38(2):89–96 non-fully symmetric parallel delta robot. Mech Mach
66. Olsson A (2009) Modeling and control of a delta-3 robot. Theory 161:104311
MSc Theses 82. Shome SS, Beale DG, Wang D (1998) A general method
67. Padilla A (2017) Recursive identification of continuous- for estimating dynamic parameters of spatial mechanisms.
time systems with time-varying parameters. Ph.D. thesis, Nonlinear Dyn 16(4):349–368
Université de Lorraine 83. Stenlund B, Gustafsson F (2002) Avoiding windup in
68. Rad SA, Tamizi MG, Azmoun M, Masouleh MT, Kalhor A recursive parameter estimation. Preprints of reglermöte
(2020) Experimental study on robust adaptive control with 2002, pp 148–153
insufficient excitation of a 3-DOF spherical parallel robot 84. Stewart D (1965) A platform with six degrees of freedom.
for stabilization purposes. Mech Mach Theory 153:104026 Proc Inst Mech Eng 180(1):371–386
69. Rad SA, Tamizi MG, Mirfakhar A, Masouleh MT, Kalhor 85. Swevers J, Ganseman C, De Schutter J, Van Brussel H
A (2021) Control of a two-DOF parallel robot with (1996) Experimental robot identification using optimised
unknown parameters using a novel robust adaptive periodic trajectories. Mech Syst Signal Process
approach. ISA Trans 10(5):561–577
70. Rao Sripada N, Grant Fisher D (1987) Improved least 86. Thirumalai S, Gallivan K, Van Dooren P. Algorithms for
squares identification. Int J Control 46(6):1889–1913 rank-deficient and ill-conditioned Toeplitz least-squares
71. Rong Q, Shi J, Ceglarek D (2001) Adjusted least squares and QR factorization
approach for diagnosis of ill-conditioned compliant 87. Tho TP, Thinh NT (2014) Analysis of kinematics and
assemblies. J Manuf Sci Eng 123(3):453–461 dynamics of 4-DOF delta parallel robot. In: Robot intel-
ligence technology and applications, vol 2. Springer,
pp 901–910

123
506 Meccanica (2022) 57:473–506

88. Tsai LW (1999) Robot analysis: the mechanics of serial 95. Yan Y, Cai Y (2006) Precision PEP-II optics measurement
and parallel manipulators. Wiley, Hoboken with an SVD-enhanced least-square fitting. Nucl Instrum
89. Tu Y, Wernsdorfer A, Honda S, Tomita Y (1997) Esti- Methods Phys Res Sect A 558(1):336–339
mation of conduction velocity distribution by regularized- 96. Yang C, Han J, Zheng S, Peter OO (2012) Dynamic
least-squares method. IEEE Trans Biomed Eng modeling and computational efficiency analysis for a
44(11):1102–1106 spatial 6-DOF parallel motion system. Nonlinear Dyn
90. Vieira HL, Beck AT, da Silva MM (2021) Combined 67(2):1007–1022
interval analysis-Monte Carlo simulation approach for the 97. Yoo CK, Sung SW, Lee IB (2003) Generalized damped
analysis of uncertainties in parallel manipulators. Mecca- least squares algorithm. Comput Chem Eng
nica 1–15 27(3):423–431
91. Wang D, Wang L, Wu J (2021) Physics-based mecha- 98. Zengqiang C, Maoqiong L, Zhuzhi Y (2000) Convergence
tronics modeling and application of an industrial-grade and stability of recursive damped least square algorithm.
parallel tool head. Mech Syst Signal Process 148:107158 Appl Math Mech 21(2):237–242
92. Wen S, Qin G, Zhang B, Lam HK, Zhao Y, Wang H (2016) 99. Zhao Y (2013) Dynamic optimum design of a three
The study of model predictive control algorithm based on translational degrees of freedom parallel robot while
the force/position control scheme of the 5-DOF redundant considering anisotropic property. Robot Comput Integr
actuation parallel robot. Robot Auton Syst 79:12–25 Manuf 29(4):100–112
93. Xu P, Li B, Chueng CF (2017) Dynamic analysis of a 100. Zolghadrit A, Cieslakt J, Goupil P, Dayre R (2017)
linear delta robot in hybrid polishing machine based on the Turning theory to practice in model-based FDI: successful
principle of virtual work. In: 2017 18th International application to new generation airbus aircraft. IFAC-
Conference on Advanced Robotics (ICAR). IEEE, PapersOnLine 50(1):12773–12778
pp 379–384
94. Yaacoub F, Abche A, Karam E, Hamam Y (2008) MRI
Publisher’s Note Springer Nature remains neutral with
reconstruction using SVD in the least square sense. In: 21st
regard to jurisdictional claims in published maps and
IEEE international symposium on Computer-Based Med-
institutional affiliations.
ical Systems, 2008. CBMS’08. IEEE, pp 47–49

123