The 14th IFToMM World Congress, Taipei, Taiwan, October 25-30, 2015

DOI Number: 10.6567/IFToMM.14TH.WC.OS13.026

A Comparative Study of Inverse Dynamics based on Clavel’s Delta robot

J. Brinker 1 B. Corves 2 M. Wahle

RWTH Aachen University RWTH Aachen University RWTH Aachen University
Aachen, Germany Aachen, Germany Aachen, Germany

Abstract— Based on Clavel’s Delta robot, one of the [14] both based on the virtual work principle or the
best-known and most widely spread parallel robots, this Hamilton’s Principle applied to the NUWAR Delta by
paper illustrates the three fundamental approaches for Miller [15], [16], [17]) and mixed approaches that
dynamic modelling (i.e. the Principle of Virtual Work, contain more than one of the aforementioned
the Newton-Euler Formulation, and the Lagrangian formalisms including numerical examples can be found
Formulation). In a direct comparison and using the in [8], [18], [19].
same notation, the differences between the three This paper illustrates the three main approaches for
modelling approaches are emphasized and demon- dynamic modelling in a direct comparison using the
strated clearly. Complete and simplified models are same notation. Thereby, the differences between the
introduced, validated and analyzed concerning their three approaches are emphasized and demonstrated
processing times as one of the key factors for real-time clearly. Sec. II shows a review of dynamic modelling
control. approaches of Delta robots. Then, the inverse kinematic
Keywords: Dynamic modelling, principle of virtual work, Newton- analysis is performed revealing the position, velocity
Euler formulation, Lagrangian formulation, Delta parallel robot
and acceleration of the centers of gravity for each body
of the system (Sec. III). In Sec. IV, the dynamic analysis
I. Introduction is performed using three complete non-simplified
In the 1980s, Clavel invented a parallel robot at the modelling approaches, i.e. the Principle of Virtual Work
École polytechnique fédérale de Lausanne (EPFL) (Sec. IV.A), the Newton-Euler Formulation (Sec. IV.B),
known as Delta robot with three translational degrees of and the Lagrangian Formulation (Sec. IV.C). Ad-
freedom (dof) dedicated to high-speed applications [1], ditionally, two simplified approaches are described (Sec.
[2], [3], [4]. Nowadays, the Delta robot is one of the IV.D). Sec. V gives an overview about the results.
best-known and most widely spread parallel robots.
II. Dynamic Modelling Approaches of Delta Robots
Besides various different designs of the original
architecture, current industrial versions employ ad- The Principle of Virtual Work implies that the work
ditional serial mechanisms in order to obtain rotational due to external forces corresponding to any set of virtual
dof. Bonev [5] gives a detailed overview about the displacement is zero. Applying this principle to analyze
historical, academic and industrial development of the the dynamics of a system does not involve unknown
Delta robot. Research in the general fields of kinematics, constraint forces and moments. It is therefore generally
dynamics, control, singular configurations, workspace, more efficient than the Newton-Euler and Lagrangian
calibration, and mechanical design of Delta robots has method. However, since the calculation of the actuation
been conducted extensively during the last decades. A torques involves the inverse of the transpose of the
decent overview about the main contributions can be Jacobian matrix, the method may become instable for
found in [6]. By solving the inverse dynamic problem singular configurations [8]. Analyses of the Delta robot
the required actuation torques for a given trajectory based on the virtual work principle can be found in
(with known position, velocity, and acceleration) and [20]*, [21]*, [22], where * denotes simplified approa-
thus, the required dimensions of the robot and its ches.
actuators can be determined. On the other hand, With the Newton-Euler Formulation the equations of
dynamic models are needed for control purposes. Hence, motion for each body of the robot are taken into account.
dynamic modelling of parallel robots has been subject Thus, also internal forces can be stated allowing to
of great research interest. The most common analytical introduce frictional effects into the dynamic model. This
approaches for dynamic modelling are [7], [8]: may also improve dimensional syntheses. Clavel [20]
however sees little advantage in these aspects, since the
- the Principle of Virtual Work, spherical joints within the Delta structure inherently
- the Newton-Euler Formulation, and ensure low friction and dimensioning of components is
- the Lagrangian Formulation rather restricted by limitations of production as well as
An overview about these approaches, as well as other stiffness requirements. Moreover, especially for parallel
(e.g. a recursive algorithm by Staicu et al. [9], [10], [11], robots, this model is computationally intensive due to
[12], [13] and a new approach by Wang and Gosselin the large number of equations [8], [23]. Newton-Euler
analyses of the Delta robot are presented in [7]*, [24]*,
1
[email protected] [25]*, [26]*, [27]* and for Delta configurations with
2
[email protected]
more than three dof in [28]. corresponding connecting rods, the proximal links need
An alternative energy-based approach is the Lagran- to withstand additional bending and torsional moments
gian Formulation. The Lagrangian formulation and thus, are usually of stronger dimension. The
describes the equation of motion of a mechanical system following kinematic relations are based on [4].
as a compact function of kinetic and potential energy.
A. Position Analysis
The formulation does not involve the unwanted reaction
The inverse kinematics leads to the relation between
forces and moments and thus, is more efficient than the
the given end-effector position p and the desired
Newton-Euler-Formalism [8]. Three generalized
angle of rotation of the three rotary actuators. The
coordinates are sufficient to describe the dynamics of a
position equation in the global or base coordinate
3-dof Delta robot. However, to avoid difficult
system 0 is given as:
expressions due to the complex geometrical model,
usually the Lagrangian equation of the first type with p  l1i  l2i  ai  bi (1)
redundant coordinates requiring a set of constraint
equations is employed [8], [16], [29], [30]. Lagrangian Figure 1 shows the geometric relations as well as the
based analyses of Delta or Delta-like robots are shown associated vectors and angles.
in [8]*, [27]*, [29], [30]*, [31], [32], [33]*, [34] and for
a 4-dof Delta robot with additional telescopic chain in
[35].
The three approaches are theoretically equivalent, but
the computational intensity may vary [19]. A general
comparison of the computational cost of the different
formulations is presented by Lin and Song [36].
In order to obtain real-time capable models for
control purposes, complicated complete models are
frequently simplified [21], [37]. In these models usually
the rotational inertias of the light-weight distal links are
neglected and their masses are distributed to the
corresponding joints, where 2/3 of the mass are usually
allocated to the tip of the proximal link [19], [21]. Tsai
and Stamper [8], [27] present a simplified Lagrangian
approach for the University of Maryland manipulator, a
Delta-based structure with revolute joints only. Here,
the masses of the connecting rods are distributed evenly
to the tip of the proximal joint and end-effector joint,
respectively (see also [30], [33]). Miller and Clavel [31]
illustrate a simplified and a complete Lagrangian based
approach comparing their accuracy and computational
efficiency.
Neglecting the distal link’s rotational inertia and
changing the direction of the resultant dynamic forces
may cause significant errors (i.e. differences in torque)
of up to 10% depending on the system parameters.
Obtaining similar processing times (about 1,6 times Figure 1: Geometric relations (XZ- and YZ-view)
longer), the authors conclude that the proposed The actuators’ and end-effector’s joints lay on a circle
complete model can also be utilized for real-time con- with radius a = | ai | and b = | bi |, respectively with
trol. Complete non-simplified approaches may also be their circle centers in the base and end-effector
useful for optimization purposes (e.g. analyses of joint coordinate system. The angular division, measured
forces and dimensional syntheses). around the Z-axis, is the same for base and
III. Kinematics end-effector.
The transformation into the local leg-coordinate
Due to the spatial four-link parallelogram (denoted as
systems (cf. Figure 2) is given by the inverse
distal link), the end-effector motion of a Delta robot has
transformation matrix:
three translational and generally no rotational dof [1],
[19]. The connecting rods within the parallelogram are  cos  i sin  i
0
each attached to the proximal links and end-effector,  
i
T0    sin  i cos  i
0 (2)
respectively by spherical joints. Thus, the rods only  0 
need to transmit axial forces allowing for light-weight  0 1
materials and very low inertia. The actuators drive the The following analyses depend on the vectors
rotational motion of the proximal links enabling solely pointing along the links l1i and l2i. For their derivation,
one rotational dof. In contrast to the distal links or the a leg-specific auxiliary vector ri between the position
 of the actuated joint ai and the corresponding joint bi the second auxiliary angle φ2i is given as
connecting the end-effector and the distal link is
defined [38]:  i ri,x 2  i ri,y 2  i ri,z 2  l1i 2  l2i 2 
2i  arccos   (7)
 2  l1i  l2i  sin 3i  
 a  b   
 
i
ri   ai  p  bi   0   iT0  p
i i i
where 0< φ2i <180°. With these angles, the actuation
 0  angle can be derived as:
 
 a  b  px cos( i )  p y sin( i )  (3) l1i i ri,z  l2i s 3i c 2i i ri,z  l2i s 3i s 2i i ri,x
  sin 1i   
  px sin( i )  p y cos( i )  2l1i l2i s 3i c 2i  l1i 2  l2i 2  l2i 2 c2 3i
  (8)
l1i i ri,x  l2i s 3i s 2i i ri,z  l2i s 3i c 2i i ri,x
 p  cos 1i  
 z  2l1i l2i s 3i c 2i  l12  l2i 2  l2i 2 c2 3i

leading to:
 l1i i ri,z  l2i s 3i c 2i i ri,z  l2i s 3i s 2i i ri,x 
X2 1i  atan    (9)
2  l i r  l s s i r  l s c i r 
 1i i,x 2i 3i 2i i,z 2i 3i 2i i,x 
Y
With these angles and Eq. (4) l1i and l2i can be
Y1 calculated.
Y2 a2
B. Velocity Analysis
3 O
Time differentiation of Eq. (1) leads to:
a1
X1 X p  ω1i  l1i  ω2i  l2i
a3 (10)
The angular velocity of the proximal link can be
calculated by:

Y3 ω1i  1i  s Ai (11)

X3 with 1i as absolute velocity of the actuated joint and

sAi denoting the rotation axis of the proximal link and
Figure 2: Local leg-coordinate system
Yi-axis, respectively. The rotation axes can be derived
This vector depends on the given end-effector by transforming the local Yi-axis into the reference
position p and thus, in turn is known as well. Due to coordinate system:
the parallel structure, the auxiliary vector can also be 0
described considering the vectors along the proximal  
s Ai  iT0 -1   1  (12)
and distal links:
0
 
i
ri  i l1i  i l 2i
The angular velocity of the distal link can be
 l1i  cos 1i    l2i  sin 3i   cos 1i  2i   eliminated by dot-multiplying Eq. (10) with the
    (4)
 0  l2i  cos 3i   vector along the distal link l2i:
 l  sin     l  sin    sin     
 1i 1i   2i 3i 1i 2i  p  l2i  φ1i   s Ai  l1i   l2i (13)
The aforementioned vectors and angles are displayed Expanding Eq. (13) for each leg leads to:
in Figure 1.
The local Y-axes correspond to the axes of rotation of  l21T    s A1  l11   l21 0 0   11 
     
the actuators. The actuation angles φ1i relate to the  sA2  l12   l22
 l22T   p   0 0    12  (14)
direction of rotation around the negative Y-axis  T
simplifying the mathematical relations. From Eqs. (3)
 l23 
 

 0 0  sA3  l13   l23   13 
and (4) and the second component of the resulting
relationship, it follows for the auxiliary angle φ3i: The Jacobian matrix JP, relating the given end-ef-
fector velocity to the velocity of the actuated joint,
  px sin i  p y cos i  with:
3i  arccos   (5)
 l2i   J P1
T 
 
 
With the help of ri in Eq. (3): φ1  J P  p   J P2
T
 p (15)
 T 
 J P3 
i 2
rix  i riy2  i riz2  l1i2  l2i2  2l1i l2i sin(3i ) cos(2i ) (6)  
 can then be obtained as: l2i  p  ω1i   ω1i  l1i  ... 
ω2i    (25)
 ...  ω1i  l1i  ω2i   ω2i  l 2i  
   A1 11  21 
2
 s l l -1  l2 
  l21 
0 0 T

   T
 sA2  l12   l22 
-1
JP   0 0    l22  (16) Also, from Eq. (22) the acceleration of the COG of
  
-1   l T  the proximal link can be extracted as:


0 0  sA3  l13   l23    23 
 l  l 
a1i ,COG   ω1i  1i  ω1i   ω1i  1i   (26)
Assume symmetric proximal links with a central  2  2 
point of gravity, the velocity at the end of the link (i.e.
the connecting point to the parallelogram) is given The acceleration of the COG of the distal link can
by: then be derived as:
v1i  ω1i  l1i (17)  2  a1i ,COG .... 
 
with a2i ,COG   l  l  (27)
 ....  ω2i   ω2i  2i   ω2i  2i 
  2  2 
ω1i  1i  s Ai  J Pi
T

 p  s Ai 
Eqs. (22) and (24-27) hold within the local
 l 2i T  (18) link-coordinate system due to their invariance. The
  p   s Ai
 s l l  angular acceleration of the proximal link can be
 Ai 1i 2i 
formulated according to Eq. (11).
Analogously, the velocity of the center of gravity The equations need to be adjusted by changing the
(COG) of the proximal link is: distance vectors in case of unsymmetrical links and
non-central COG, respectively.
ω1i  l1i
v1i ,COG  (19) IV. Dynamics
2
The following sections present the derivation of
The angular velocity of the distal link (i.e. the
complete models for Delta robots using the main
parallelogram) can be calculated by rearranging
representatives of dynamic modelling approaches, i.e.
Eq. (10):
the Principle of Virtual Work (A), the Newton-Euler
l2i   p  ω1i  l1i  Formulation (B), and the Lagrangian Formulation (C).
ω2i  (20)
l2 2
A. Principle of Virtual Work
The velocity of the COG of the distal link is then First, six-dimensional resultants of the applied and
given by: inertia wrenches acting on the COG of each body are
determined (A1). Second, the Link Jacobian Matrices
ω2i  l2i [8], [38], [39] are derived relating the virtual
v2i ,COG  p  (21)
2 displacements of the links to the one of the
end-effector (A2). Finally, the equations of motion
C. Acceleration Analysis
can be obtained applying the principle of virtual work
Differentiation of Eq. (10) with respect to time leads
(A3). The following model is based on the complete
to the acceleration of the end-effector:
model in [22] which in turn is based on Tsai’s virtual
p  ω1i   ω1i  l1i   ω1i  l1i work based dynamic analyses of a Stewart-Gough
(22) manipulator [39]. Similarly, Nefzi [38] applied the
 ω2i   ω2i  l 2i   ω2i  l 2i virtual work approach to a 5-dof parallel robot.
The angular acceleration of the proximal link can be A1. Forces and Moments
calculated in the same manner as Eq. (11): First, the forces and moments (i.e. a six-dimensional
wrench) due to the gravity g and inertial effects need
ω1i  1i  s Ai (23) to be considered for the links and end-effector. Other
Dot-multiplying Eq. (22) with the vector along the external forces and moments as well as frictional
distal link l2i eliminates the influence of the distal link forces are neglected here. Thus, the forces fP and
acceleration and gives the acceleration of the moments nP acting on the COG of the end-effector
actuated proximal link: can be summarized as:

 p  ω1i   ω1i  l1i   ω2i   ω2i  l2i    l2i  f   m  g  p 

1i  (24) QP   P    P  (28)
 s Ai  l1i   l2i  nP   0 

Rearranging Eq. (22) finally gives the angular with mp as end-effector’s mass. Similarly, the
acceleration of the distal link: resultants for the proximal and distal links in the local
coordinate system can be derived as:
  i f1i   m1 iT0  g  i a1i ,COG   
 i
ω2i 
 iT0  0 l2i   iT0  J 1i
    0
p  J 2i,ω  0 p
i
Q1i    (29) (37)
 i n    i I  i ω  i ω  i I  i ω
    l2i  2
 1i   1i 1i 1i 1i 1i 
and Taking advantage of the invariance of the cross-pro-
duct, for the velocity of the COG of the distal link it

i
 i f 2i  
Q2i   
m2 iT0 g  i a2i ,COG   

follows:
(30)
 i n    i I  i ω  i ω  i I  i ω
   ω2i  i l 2i
i
 2i   2i 2i 2i 2i 2i 
i
v 2i ,COG  i p 
2
with m1i and m2i for the masses as well as I1i and I2i i 1 
 T0    iT  0 l 2i   ...
2  0 
for the inertia matrices of the proximal and one of the
 2   2i 
l 
  0 p (38)
two connecting rods of the parallelogram of the distal  
links, respectively. 

...  iT0  0 l 2i   iT0  J 1i

  

A2. Link Jacobian Matrices  J 2i,COG  0 p
The relationship between the end-effector’s velocity,
defined in the global coordinate system, and the A3. Equations of Motion
proximal link’s absolute velocity, within the local From the resultants of the applied and inertial forces
link-coordinate system, can be obtained by the and moments the principle of virtual work can be
derivation of so-called link Jacobian matrices [38], stated as:
[39]:
 φ1T τ   χ TP QP
 l1i T 
 2  sin 1i  J Pi 
3 3 (39)
   i χ1i
T i
 Q1i    i χ T2i  2  iQ2i  0
  0
i
v1i ,COG  0   p  J 1i,COG  p
0
(31) i 1 i 1
 l  with τ as the vector of actuation torques.
  1i  cos 1i  J Pi
T 
 2  Relating the virtual displacements of the links to the
virtual displacement of the end-effector and taking
for their COG and into account the relationships in Eqs. (15), (31), and
(38), it follows [8], [39]:
 l1i  sin 1i  J Pi
T 
  0
i
v1i   0   p  J 1i  p
0
(32)  φ1  J P   χ P (40)
 T   χ1i  J1i*
i
 χP
 l1i  cos 1i  J Pi  (41)
 i χ 2i  J 2i*  χP (42)
for their ends, respectively. For the angular velocity it
follows analogously: where
 0   J 1i,COG 
  J 1i*    and (43)
i
ω1i  1i  s Ai i
  J TPi   0 p  J 1i,ω  0 p (33)  J 1i,ω 
 0 
   J 2i,COG 
*
J 2i   (44)
Formulating Eq. (10) for the local link-coordinate  J 2i,ω 
system gives:
Thus, Eq. (39) gives:
i
T0  p  J1i  p  ω2i  l2i
0 0 i i
(34)
J 
T
P  χP τ   χ TP  QP
i
Cross-multiplying both sides with l2i leads to:
 
3
  J 1i*   χ P
T
i
l2i   T
i
0  
 J 1i  p   l2i   ω2i
0 2 i
(35) i 1
 iQ1i (45)

 
3
  J 2i
T
and after rearranging: *
 χP  2  iQ2i  0

 T  
i 1
i
l2i  i
0  J 1i  0 p
i
ω2i  (36) and hence after factorizing, cancelling δχP, and
 l2i  2
rearranging:

The end-effector’s velocity can be factorized using

τ  J PT

 3
 Q P   
 1i  
 J * T  iQ  ...  
1i 
the cross-product-matrix for iT0  0 l2i denoted by  (46)
 
T
  :  i 1  ... J *  2  i
Q2i 

  2i 
B. Newton-Euler Formulation case of non-symmetrical links, the orientation of the
links cannot be defined due to the spherical joints.
The kinematic relationships describe the state of
However, in industrial applications, usually sym-
motion of the robot’s links and end-effector without
metrical links are used. The Y-axis of the moving
considering the required forces and moments. These
coordinate system is an arbitrarily defined vector
relations can be formulated by the Newton-Euler
perpendicular to the direction of the distal link and in
approach. Therefore, a kinetic chain is cut open at the
this case perpendicular to the global X-axis. The
joints and analyzed in respect to the forces and
corresponding rotation matrix is:
moments of the distal links (B1), the end-effector (B2)
and finally, the proximal links (B3). Action and
reaction forces are assigned to each subsystem which

 l
Q2i   2i
1 0 0   l2i
T

l2i  1 0 0   l2i 
T


 
(48)
finally gives the required actuation torques acting on
 l2i 1 0 0   l2i
T
l2i  1 0 0   l2i 
T
the proximal link. A similar approach for a 5-dof  
parallel robot can be found in [40].
The inertia matrix of the distal link within the base
B1. Forces and Moments of the Distal Link coordinate system can thereby be calculated as:
The basic configuration consists of three proximal
links, six distal links and the end-effector. The I 2i  Q2i  i I 2i  Q2i T
derivation of the equilibria of forces and moments for  d 2,outer 2  d 2,inner 2 
each part leads to a system of equations with 60  0 0 
unknowns [20]. However, in this application, the  8 
equilibrium of forces for the proximal link is not  l2 2  (49)
needed to derive the actuation torques. Figure 3  Q2i  m2i   0 0   Q2i T
 12 
displays the free body diagram of a connecting rod.  2 
l2 
 0 0
F O ,i  12 
Spherical joint 
(Proxim al link)
For the angular velocity and acceleration of the distal
D istal link link, there is no component in direction of the link.
Hence, the X-component, which is dependent on the
l 2i m 2i  g
outer and inner link diameters, is obsolete. The
constant rotation axis of the proximal link is identical
 I 2i  ω 2i  ω 2i   I 2i  ω 2i 
to the Y-axis of the local leg-coordinate system.
 m 2i  a 2i Hence, the rotation matrix for the proximal link is
given as:
 Fi  cos 1i  0  sin 1i  
i -1  
Q1i  T0   0 1 0  (50)
Spherical joint FT ,i  sin   0 cos   
(End-effector)  1i 1i 

Figure 3: Distal link: forces and moments and used to calculate the inertia matrix for the
proximal link analogous to Eq. (49). Here, the motor
The force component acting perpendicularly to the
inertia needs to be added to the Y-component.
direction of the distal link as well as between the
distal link and the end-effector can be computed by
B2. Forces and Moments of the End-Effector
formulating the equilibrium of moments for the distal
The force components acting in the direction of the
links, related to the spherical joints between the
distal link are the unknowns in respect to the
proximal and distal link:
end-effector-equilibrium. Thus, six unknown varia-
 l2i l  bles need to be identified for the equilibria of forces
   m2i  a2i   2i   m2i  g  ...  l2i and moments for the end-effector. The tangential
FT,i   2 2  l 2 (47)
  2i forces FT,i from Eq. (47) are external forces acting on
 ...  I 2i  ω2i  ω2i  2i
 I  ω2i  
the isolated end-effector (cf. Figure 4).
This component is solely dependent on the The tangential force vectors are identical for each
momentary position and state of motion of the distal pair of connecting rods of the same distal link,
link. whereas the forces Fi acting along the rods may be
The locally defined inertia matrix needs to be trans- different. The end-effector-equilibrium is:
formed into the base coordinate system. The moment l
of inertia in direction of the distal link (i.e. the  l2i Fi  mP  g  p    FT,i  0
X-direction of the moving coordinate system) is i 2i i
(51)
presumed to be known. For a symmetrical link, the  
  bi  l2i Fi   rP,S  mP  g  p     bi  FT,i   0
l
components in Y- and Z-direction are identical. In
i  2i  i
 F6
a given state of motion, represented by the Y-com-
F5
ponent of:
l 32 l 32 i
MO,i  Q1iT  MO,i (57)
F1  FT ,3
 FT ,3 F4 M O ,i
b6 b5
l 21
mP  p l 22
b4  FT ,2
b1
 FT ,1 rP ,S O'
A ctuator m 1i  g c i,2
mP  g F3
F2 b3 c i,1  I 1i  ω 1i
b2
 ω 1i   I 1i  ω 1i 
l 21 l 22
 F O ,1  m 1i  a 1i
E nd-effector  F O ,2
 FT ,1  FT ,2
Proxim al link
Figure 4: End-effector: forces and moments Figure 5: Proximal link: forces and moments

By factorizing, a six dimensional system of equations C. Lagrangian Formulation

for the unknowns Fi can be formulated as: First, the kinetic and potential energies of each body
are derived in order to obtain the Lagrangian equat-
Fi  J int -1  Fext (52) ion (C1). Then, the constraint equations are obtained
taking into account, that the distance between the
with joints connecting the distal links to the proximal links
 l21 l21 l22 l22 l23 l23  and end-effector, respectively is equal to length of the
 l l21 l22 l22 l23 l23  connecting rods (C2). In C3, the partial derivatives
J int   
21
(53) are taken to finally calculate the required actuation
 l21 l l l l l  torques (C4). The presented model is based on the
 b1  b2  21 b3  22 b4  22 b5  23 b6  23 
 l21 l21 l22 l22 l23 l23  complete Lagrangian model by Miller and Clavel [31]
which is similarly applied in Ángel et al. [32].
and
C1. Lagrangian Equation
 mP  p  g    FT,i  The Lagrange function is derived by considering the
 i 
Fext   (54) total kinetic energy K and the total potential energy U
 rP,S  mP  p  g     bi  FT,i   of the Delta robot. The total kinetic energy is:
 i 
3
as matrix of coefficients and the vector of external K  K P    K1i  K 2i  (58)
forces, respectively. i 1

where KP is the kinetic energy of the end-effector and

B3. Forces and Moments of the Proximal Link K1i and K2i the kinetic energies of the proximal and
With this and the equilibrium of forces of the distal distal links, respectively. These energies in turn are
link, the joint forces between the proximal and distal given by [31], [32]:
links can be calculated as:

FO,i
l
 FT,i  2i Fi  m2i  a2i  g  (55)
1
2

K P   mP  px2  p 2y  pz2  (59)
l2i
1 1 l2 
K1i  I1i  1i2   I Motor ,i  m1i  1i  1i2 (60)
Finally, the moment acting on the actuator joint of the 2 2  3 
proximal link can be formulated as (cf. Figure 5):
1

K2i   m2i  v2iT ,COG  v2i ,COG  ω2iT  I 2i  ω2i  (61)
MO,i    ci,k  FO,k  
l1i
 m1i   a1i  g  ... 2
k 2 (56) The kinetic energy of the distal links can alternatively
...  I 1i  ω1i  ω1i   I 1i  ω1i  be expressed as [31]:
1 1
   p  v   
T
where ci denote the distance vectors between the K2i   m2i    i p  i v1i i i
p  v1i 
i T i
(62)
1i
rotation points of the proximal links and the con- 2 3 
necting points (i.e. the spherical joints) of the rods.
For a better understanding, here, m2i denotes the mass
Their simple derivation is not shown here. Trans-
of the complete distal link (i.e. both connecting rods).
formation of the moments into the local leg-coordi-
The total potential energy is:
nate system finally results in the actuation torques for
 3 L
U  U P   U1i  U 2i  (63) 0 (74)
i 1 px
L
with UP as the potential energy of the end-effector 0 (75)
and U1i and U2i as the kinetic energies of the proximal p y
and distal links, respectively. These are given by: L  1 
   mP    m21  m22  m23    g and (76)
U P  mP  g  pz (64) pz  2 
1  cos( i )  cos 1i   1i  px  
U1i   m1i  g  l1  sin 1i  (65)  
2  sin( i )  cos 1i   1i  p y  
 
1
2

U 2i   m2i  g  pz  l1  sin 1i   (66)  sin 1i   1i  pz 




 d  L  m2i  l1i
       cos( i )  sin 1i   px   (77)
The Lagrangian function can then be derived as:  dt   1i  6  
 sin( i )  sin 1i   p y  
L  K U (67)  
 cos 1i   pz  
C2. Constraint Equations  
 2  1i  l1i  I1i  1i 
The vector along the distal link l2i within the global  
coordinate system is described as:  cos( i )  cos 1i   1i  px  
 px   a  b  l1i  cos 1i    cos( i )  L m2i  l1i  
     sin( i )  cos 1i   1i  p y  
1i 6  
l 2i   p y   a  b  l1i  cos 1i    sin( i ) 
(78)
(68)  sin 1i   1i  pz 
   

 p z  l1i  sin  1i  
 3  g  cos 1i    m1i  m2i 

Thus, the constraint equations are [27], [32]: for i=1, 2, 3. Differentiation in respect to the six
generalized coordinates of the constraint equations
Γi   l 2 i 
2
 l2i2  0 (69) gives:


 px   a  b  l1i  cos 1i    cos( i )  
2
Γi
 2   px   a  b  l1i  cos 1i    cos(i )  (79)
(69) px
 p   a  b  l  cos    sin( ) 
2

y 1i 1i i Γi
 2   p y   a  b  l1i  cos 1i    sin(i )  (80)
 p  l  sin    l p y
2 2
z 1i 1i 2i
Γi
C3. Partial Derivatives
pz
 
 2  pz  l1i  sin 1i  for i=1, 2, 3 (81)
Partial derivation after the derivatives of the six
Γ1
generalized coordinates, i.e.
11  
 2  l11   a  b  px cos(i )  p y sin(i ) sin 11   pz cos 11 
 (82)
Q   pxpz 11 12 13 
py Γ2
 
(70)
 2  l12   a  b  px cos(i )  py sin(i ) sin 12   pz cos 12  (83)
where px, py, and pz are redundant coordinates, and 12  
subsequent time differentiation lead to: Γ3
13  
 2  l13   a  b  px cos(i )  p y sin(i ) sin 13   pz cos 13 
 (84)
 d   L   1 
      mP    m21  m22  m23   px
 Γ1 Γ1 Γ2 Γ2 Γ3 Γ3
   px  
dt 3       0 (85)
(71) 12 13 11 13 11 12
 
3
1
  m2i  cos(i )  l1i  cos 1i   1i2  sin 1i   1i C4. Actuation Torques
i 1 6
Finally, the Lagrange multipliers λi can be obtained
 d   L   1  for a given state of motion by solving the following
      mP    m21  m22  m23   p y

 dt   p y   3  system of equations:
(72)
 d   L  L 3 Γ
 
3
  i  i
1
  m2i  sin(i )  l1i  cos 1i   1i2  sin 1i   1i      (86)
i 1 6    px  px i 1 px
dt
 d   L   1   d   L  L 3 Γ
      mP    m21  m22  m23   pz       i  i (87)
 dt   pz   3  
 dt   p y  p y i 1 p y
(73)
 
3
1
  m2i  l1i  sin 1i   1i2  cos 1i   1i
i 1 6
 Table 2: Processing times
 d   L  L 3 Γ
       i  i (88)
 dt   pz  pz i 1 pz
Model Time
Principle of Virtual Work 0,73 s
With these multipliers the desired actuation torques τi Newton-Euler Formulation 1,00 s
can be calculated for i=1, 2, 3 as:
Lagrangian Formulation 0,37 s
 d   L  L Γ
i        i  i (89)

 dt   1i  1i 1i

D. Simplified approaches
The introduced models can be simplified by neglect-
ting the rotational inertia of the distal links. Therefore,
the masses of the connecting rods are distributed to
the corresponding connecting joints with the proxi-
mal link (1) and the end-effector (2) by ν1A =2/3 and
ν2A =1/3 [21] or ν1B =1/2 and ν2B =1/2 [27]. Hence,
Eq. (61) needs to be adapted to:
1
   
K2i    1  m2i  l1i2  1i2  2  m2i  px2  p 2y  pz2 
2  
(90)

The constraint equations and their derivatives (cf. C2

and C3) hold equally for the simplification. For the
sake of brevity, the influence of the simplified
approach on the partial derivatives is not shown here.

V. Analyses

To validate the approaches and for further analyses,

the models were implemented in MATLAB® using the
kinematic parameters and mass parameters as shown in Figure 5: Test trajectory and specifications
Table 1 (inertial values omitted).
Table 1: Kinematic and mass parameters

Parameter Value Unit

Radius (base) a 0,200 m
Radius (end-effector) b 0,070 m
Length (proximal link) l1i 0,350 m
Length (distal link) l2i 0,900 m
Mass (proximal link) m1i 0,650 kg
Mass (distal link) m2i 0,375 kg
Mass (end-effector) mP 2,500 kg

The high value of the end-effector mass results from

factoring in a motor attached to it to drive an additional
dof. This is applied in many industrial solutions (36% of
25 commercial portfolios considered). Other companies
usually deploy a central telescopic shaft (72%).
Analyses are based on a line trajectory and the
specification as shown in Figure 5.
Figure 6 displays the corresponding actuation torques
for the three complete models for the proximal links,
distal links and end-effector, respectively. Figure 6 Figure 6: Component-specific torques
bottom right shows the curves of actuation torques for The results demonstrate the advantage of the virtual
the complete robot. All models result in the same work and Lagrangian based approaches over the
solution. Assume a normalized reference time of one computationally intensive Newton-Euler formulation
second as a base for the longest computation time, the with regards to their processing times. In this case, the
processing times for the models can be stated as shown Lagrangian model is more efficient than the virtual
in Table 2. work approach. Merz [23] and Pietsch [41] show similar
results analyzing two prototypes of parallel robots. the IEEE International Conference on Robotics and Automation
(ICRA), pp. 532-537, Nagoya (JP), 1995.
However, it should be noted that the efficiency of a [16] Miller K. Dynamics of the New UWA Robot. In Proc. of the
model is highly dependent on the analyzed mechanical Australian Conf. on Robotics and Automation, Sidney (AU), 2001.
structure (e.g. strong increase in complexity of the [17] Miller K. Optimal Design and Modeling of Spatial Parallel
Manipulators. In The International Journal of Robotics Research,
Lagrangian model for robots with rotational capacity 23(2):127-140, 2004.
[41]), the computational scheme (e.g. recursive methods [18] Guglielmetti P. Model-based Control of Fast Parallel Robots: a
[36]) and the program structure (e.g. number and kind of Global Approach in Operational Space. Ph.D. Thesis, EPFL,
Lausanne (CH), 1994.
mathematical operations). [19] Merlet J.-P. Parallel Robots. Springer Dordrecht (NL), 2006.
[20] Clavel R. Conception d'un robot parallele rapide a 4 degres de liberté.
VI. Conclusions Ph.D. Thesis, EPFL, Lausanne (CH), 1991.
Three fundamental approaches for dynamic model- [21] Codourey A. Dynamic Modelling and Mass Matrix Evaluation of the
DELTA Parallel Robot for Axes Decoupling Control. In Proc. of the
ling were introduced, validated and analyzed based on IEEE/RSJ International Conference on Intelligent Robots and
the widely known Delta robot. The main focus of the Systems (IROS),pp. 1211-1218, Osaka (JP), 1996.
presented analyses is on the processing times of [22] Zhao Y, Yang Z. and Huang T. Inverse Dynamics of Delta Robot
Based on the Principle of Virtual Work. In Transactions of Tianjin
complete and simplified models as one of the key University, 11(4):268-273, 2005.
factors for real-time control. In this study, the Lagran- [23] Merz M. PenTec – ein neues Parallelstruktur-Konzept (eng: PenTec –
gian formalism is most efficient, followed by the virtual a new Concept for Parallel Structures). Diss., TU Braunschweig (DE),
2009.
work based approach. The Newton-Euler formulation is [24] Pierrot et al. High Speed Control of a Parallel Robot. In Proc. of the
inferior due to the large number of equations involved. IEEE International Workshop on Intelligent Robots and Systems
However, internal forces may be useful for design (IROS), Ibaraki (JP), 1990.
[25] Pierrot F., Fournier A. and Dauchez P. Towards a Fully-Parallel 6
optimization purposes. DOF Robot for High-Speed Applications. In Proc. of the IEEE
Future research will focus on analyses of the robust- International Conference on Robotics and Automation (ICRA), pp.
ness of the simplified approaches, primarily in respect to 1288-1293, Sacramento (US), 1991.
varying masses of the distal links and/or of the end-ef- [26] Codourey A. and Burdet E. A Body-oriented Method for Finding a
Linear Form of the Dynamic Equation of Fully Parallel Robots. In
fector. Also, the models will be extended in order to also Proc. of the IEEE International Conference on Robotics and
depict Delta robots with more than three dof as usually Automation (ICRA), pp. 1612-1618, Albuquerque (US), 1997.
deployed in industrial applications. Against this [27] Stamper R. E. A Three Degree of Freedom Parallel Manipulator with
Only Translational Degrees of Freedom. Ph.D. Thesis, University of
background, the extended models will be used for Maryland, Maryland (US), 1997.
task-oriented optimization approaches. [28] Borchert et al. Analysis of the mass distribution of a functionally
extended delta robot. In Robotics and Computer-Integrated
References Manufacturing, 31:111-120, 2015.
[1] Clavel R. Device for the movement and positioning of an element in [29] Ahmadi et al. Inverse Dynamics of Hexa parallel Robot Using
space. Patent, US 4976582, 1990. Lagrangian Dynamics Formulation. In Proc. of the IEEE Int. Conf.
[2] Demaurex, M.-O. The Delta Robot within the Industry. In Parallel on Intelligent Eng. Systems (INES), pp. 145-149, Miami (US), 2008.
Kinematic Machines, pp. 395-399, 1999. [30] Park et al. Dynamics Modeling of a Delta-type Parallel Robot. In
[3] Rey, L. and Clavel, R. The Delta Parallel Robot. In Parallel Proc. of the Int. Symp. on Robotics (ISR), pp. 1-5, Seoul (KR), 2013.
Kinematic Machines, pp. 401-417, 1999. [31] Miller K. and Clavel R. The Lagrange-based model of Delta-4 robot
[4] Wahle M. and Corves, B. Stiffness Analysis of Clavel’s DELTA dynamics. In Robotersysteme, 8:49-54, 1992.
Robot. In Proc. of the 4th Int. Conf. on Intelligent Robotics and [32] Ángel et al. RoboTenis System Part II: Dynamics and Control. In
Applications (ICIRA), Part I, pp. 240-249, Aachen (DE), 2011. Proc. of the IEEE Conf. on Decision and Control-European Control
[5] Bonev I. Delta Parallel Robot - the Story of Success. http://www.par Conf. CDC-ECC, pp. 2030-2034, Sevilla (ES), 2005.
allemic.org/Reviews/Review002p.html, 2001. (20th December 2014) [33] Traslosheros et al. New visual Servoing control strategies in tracking
[6] Zhao Y. Singularity, isotropy, and velocity transmission evaluation of tasks using a PKM. In Mechatronic Systems Simulation Modeling
a three translational degrees-of-freedom parallel robot. In Robotica, and Control, pp. 117-146, InTech, 2010.
31(2):193-202, 2012. [34] Kenmochi et al. Robust Position Control of Delta Parallel
[7] Codourey A. Contribution à la commande des robots rapides et précis: Mechanisms Using Dynamic Model and QFT. In Proc. of the IEEE
application au robot delta à entraînement direct. Ph.D. Thesis, EPFL, International Symposium on Industrial Electronics (ISIE), pp.
Lausanne (CH), 1991. 1256-1261, Istanbul (TR), 2014.
[8] Tsai L.-W. Robot analysis: the mechanics of serial and parallel [35] Tho T. P. and Thinh N. T. Analysis of Kinematics and Dynamics of
manipulators. John Wiley & Sons, NY, 1999. 4-dof Delta Parallel Robot. In Robot Intelligence Technology and
[9] Staicu S. and Carp-Ciocardia D. C. Dynamic Analysis of Clavel's Applications 2, pp. 901-910, 2014.
Delta Parallel Robot. In Proc. of the IEEE International Conf. on [36] Lin Y.-J. and Song S.-M. A Comparative Study of Inverse Dynamics
Robotics & Automation (ICRA), pp. 4116-4121, Taipei (TW), 2003. of Manipulators with Closed-Chain Geometry. In Journal of Robotic
[10] Carp-Ciocardia D. C. and Staicu S. Dynamics of Delta Parallel Robot Systems, 7(4):507-534, 1990.
with Prismatic Actuators. In Proceedings of the IEEE International [37] Codourey A. Dynamic Modeling of Parallel Robots for
Conference on Mechatronics, pp. 870-875, Taipei (TW), 2005. Computed-Torque Control Implementation. In The International
[11] Liu X.-J., Staicu S. and Wang J. Kinematics and Kinetostatics Journal of Robotics Research, 17(12):1325-1336, 1998.
Analysis of the 3-DOF Parallel Cube-Manipulator. In UPB Scientific [38] Nefzi M. Analysis and Optimisation of 4UPS - 1UPU Parallel robots,
Bulletin, Series D: Mechanical Eng., Volume 69 (2):3-18, 2007. Diss., RWTH Aachen, Aachen (DE), 2010.
[12] Chablat D., Wenger P. and Staicu S. Dynamics of the Orthoglide [39] Tsai L.-W. Solving the Inverse Dynamics of a Stewart-Gough
Parallel Robot. In UPB Scientific Bulletin, Series D: Mechanical Manipulator by the Principle of Virtual Work. In Journal of
Engineering, Volume 71 (3):3-16, 2009. Mechanical Design, 122:3-9, 2000.
[13] Staicu S. Recursive modelling in dynamics of Delta parallel robot. In [40] Wahle M. Ein Gesamtkonzept zur Optimierung der Positionier-
Robotica, 27(2):199-207, 2009. genauigkeit von Parallelstrukturen (eng: A global Concept for the
[14] Wang J. and Gosselin C. M. A new Approach for the Dynamic Optimization of the Positioning Accuracy of Parallel Structures).
Analysis of Parallel Manipulators. In Multibody System Dynamics 2, Diss., RWTH Aachen, Aachen (DE), 2014.
pp. 317-334, 1998. [41] Pietsch I. Adaptive Steuerung und Regelung ebener Parallelroboter
[15] Miller K. Experimental Verification of Modeling of DELTA Robot (eng: Adaptive Control of Planar Parallel Robots). Diss., TU
Dynamics by Direct Application of Hamilton's Principle. In Proc. of Braunschweig (DE), 2003.