Ref - 8 An Improved Kinematic Model For Calibration of Serial Robots Having Closed-Chain Mechanisms

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Robotica

http://journals.cambridge.org/ROB

Additional services for Robotica:

Email alerts: Click here


Subscriptions: Click here
Commercial reprints: Click here
Terms of use : Click here

An improved kinematic model for calibration of serial robots having


closed-chain mechanisms

Minh To and Phil Webb

Robotica / Volume 30 / Issue 06 / October 2012, pp 963 - 971


DOI: 10.1017/S0263574711001184, Published online: 17 November 2011

Link to this article: http://journals.cambridge.org/abstract_S0263574711001184

How to cite this article:


Minh To and Phil Webb (2012). An improved kinematic model for calibration of serial robots having closed-chain
mechanisms. Robotica, 30, pp 963-971 doi:10.1017/S0263574711001184

Request Permissions : Click here

Downloaded from http://journals.cambridge.org/ROB, IP address: 169.230.243.252 on 11 Apr 2015


Robotica (2012) volume 30, pp. 963–971. © Cambridge University Press 2011
doi:10.1017/S0263574711001184

An improved kinematic model for calibration of serial robots


having closed-chain mechanisms
Minh To∗ and Phil Webb
School of Engineering, Cranfield University, Cranfield, Bedfordshire, UK
(Accepted October 13, 2011. First published online: November 17, 2011)

SUMMARY chains, i.e., the parallelogram mechanism (Fig. 1) in


Many industrial robots employ closed-loop actuating current ABB’s IRB4400, IRB6660; KUKA’s KR 30,50-
elements such as the parallelogram mechanism for increased PA., KR 700-PA robots, and Comau’s Smart H, NJ, NX
stiffness. Modeling these manipulators for the purpose of series. When errors in the closed-loop actuating elements
calibration presents a challenge due to complex nonlinear are considered in kinematic calibration of these robots,
couplings between parameters of the chains. The modeling parameter redundancies are unavoidable due to the closure
method presented in this paper involves the integration of the constraints imposed by the loops, and hence, the modeling
open- and closed-loop elements whose errors can be resolved conventions for open-loop structures introduced earlier
as linear functions of their parameters. As a result, the model cannot be employed directly. To continue using them, it
is similar to that of a serial-link robot, which makes it possible is required to resolve the constraint equations explicitly
to use existing well-defined calibration techniques in the and merge the result with the existing manipulator’s
area. Simulation and experimental studies on an industrial global model. In their paper, Schröer et.al.7 considered a
robot for verifying the correctness and effectiveness of the degenerated parallelogram, a planar four-bar linkage, of
proposed model are also described. which the position constraints were mathematically solved
for its passive joint angles with respect to actuated joint
KEYWORDS: Kinematic identification; Closed-chain angles and actual link lengths. A similar approach was
mechanism; Parallelogram structure. proposed in the work of Marie et.al.,8 where the solution
was further differentiated to obtain the loops’ error model.
Though such derivations are necessary, difficulties may
1. Introduction arise due to the fact that solutions of the loop equations
Kinematic calibration for geometric parameters of serial usually are highly nonlinear and complicated as well as
manipulators is an intensive topic in robotic research. Several redundant parameters may still remain. In contrast, the model
comprehensive reviews and textbooks covering essential suggested by Alici et.al.9 is too simple because the essential
aspects regarding modeling, measurement, identification, relation between errors in passive joint angles and other
and correction of the technique are available.1–5 Modeling, loop parameters was not provided. Ananthanarayanan et.al.10
or how to represent the observable tool errors as identifiable introduced an experimental method to investigate link length
functions of internal robot errors, has received most attention. errors of a parallelogram mechanism. As this method is
Various models were already introduced, many of which only dedicated to identifying geometric parameters, their
were derived from the well-known Denavit–Hartenberg (DH) calibration results are prone to other unaccounted effects,
convention while others were specially developed for the i.e., the compliance due to robot’s gravitational loading. It is
purpose of calibration (see, e.g., [17–24]). Nevertheless, the thus desirable to have a simpler yet more accurate model for
four-parameter DH model with Hayati’s modification for manipulators of this type.
handling the cases of successively parallel joint axes (Hayati, This paper presents an improved kinematic model for
1985) has been intuitively accepted and become the most serial manipulators having closed chains (the parallelogram
popular convention in kinematic calibration owing to its mechanism in particular as this is the most popular type).
“user-friendly” form. Schröer et.al. have further proposed Error propagation in the parallelogram is modeled from
that this combined model is only a subset of their DH-based the linearization of the constraint equations, rather than
“complete, minimal, and continuous” kinematic models.6 solving the equations directly to avoid the complication
The contribution of their research is a systematic rule mentioned above. After being merged with the resulting
of setting up DH/Hayati frames for different types and loop’s error model, the manipulator’s global model becomes
configurations of joints in an open-loop structure. Once similar to that of a serial-link robot which was already
followed, the work of geometric modeling to serial-link well-defined and treated in literature. The analytic form
manipulators becomes formulaic. of the Jacobian matrix is also given, based on which the
It is worth noting that many serial manipulators are remaining redundant parameters, due to kinematic design of
not purely open-loop structures but contain closed-loop the parallelogram structure, were eliminated. A simulation
study was performed to validate the model which was
* Corresponding author. Email: m.h.to@cranfield.ac.uk then compared with the one proposed in ref. [8] in terms

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252


964 Calibration of serial robots having closed-chain mechanisms
y5 Table I. Nominal DH parameters of the Smart H4 robot.
θ6

z5,6
Link i θi (◦ ) di (m) ai (m) αi (◦ )
z4 θ5 y4
x6 x3
θ4 L1 θ1 1 0.2 −90
L∗2 θ2 0 1.05 0
z3 L2 θ2 0 0.45 0
−83 θ4'

°
x2 L3 θ3 0 1.05 0
x3'
x4' L4 θ4 0 0.45 0
Parallelogram
L3 −83 0 0.25 −90
structure L4 θ4 1.1395 0 −60
L5 θ5 1.159 0 60
x1 θ2 L6 θ6 0.2176 0 0
θ 3'
θ 2'
(∗ : link does not belong to the chain 12’. . .6)
x2'
z0
θ1
x0
resolutions of the closure constraints imposed by the loop.
Fig. 1. Schematic diagram of the Smart H4 robot with DH frame The orientation and position constraints in this case, i.e., the
assignments (passive joints are marked with gray colors). z-axes of frames 2 and 4 align and the origins of frames 2
and 4 coincide. Examining these constraints in the common
frame 1, one will find out that only two out of possible six
constraints’ equations need to be satisfied: those are position
of identification, accuracy, and convergence. Experiments constraints in x- and y-axes:11
with a Comau Smart H4 robot have shown that significant
improvement in accuracy can be made with the model a2 cθ2 + a3 cθ2 3 + a4 cθ2 3 4 − a2 cθ2 = 0,
compared with the one that simply regarded the robot as (4)
an open-link structure. a2 sθ2 + a3 sθ2 3 + a4 sθ2 3 4 − a2 sθ2 = 0,

where cθi , sθi , and θij are short for cosθi , sinθi , and θi + θj ,
2. Kinematic Model respectively.
Since a4 = a2 and a3 = a2 , Eq. (4) becomes:
2.1. Nominal forward kinematics
The nominal kinematics of the Smart H4 manipulator is a2 (cθ2 + cθ2 3 4 ) + a2 (cθ2 3 − cθ2 ) = 0,
(5)
computed following the method described in (ref. [11], a2 (sθ2 + sθ2 3 4 ) + a2 (sθ2 3 − sθ2 ) = 0,
Section 2.9.2). Joint 3 of the robot is virtually cut open,
allowing link frames that can be assigned with the DH which leads to the following solutions, given arbitrary choice
convention (Fig. 1). Joints 3 , 4 , and 3 are passive joints of a2 and a2 :
driven by two actuated and coaxial joints 2 and 2 via the
parallelogram structure formed by links 2 , 3 , 4 , and 2. θ3 = θ2 − θ2 ,
Notice that there are two frames at the cut joint 3: frame (6)
θ4 = π − θ2 + θ2 .
2 describes the relation between links 2 and 3 whereas frame
4 between links 4 and 3. Nominal parameters of the robot
are given in Table I. 2.2. Error modeling
The location of the tool with respect to (w.r.t.) the base Regarding kinematic modeling for calibration, each
frame is represented as: transformation in Eq. (1) is constructed with the modified
DH convention which has the general form of:
T60 = T10 .T31 .T43 .T54 .T65 (1)
Tii−1 = Rot(z, θi )Tran(z, di )Tran(x, ai )Rot(x, αi )
where T31 can be expressed either with the branch (12 3 4 3):
Rot(y, βi ), (7)
  
T31 = T21 .T32 .T43 .T34 (2) where βi is the additional Hayati parameter for handling the
or, equivalently, with the branch (123): case zi−1 /zi .
With this convention, the initial error model of the open
T31 = T21 .T32 . (3) chain (12 . . . 3 . . . 6) composed of eight links presumably
driven by all actuated revolute joints will have 4 × 8 =
In this paper, Eq. (1) was computed using Eq. (2) 32 parameters.4,6 It can be obtained through the use of
to account for the parameters in the parallelogram differential homogenous transformations14 of Eq. (1) in the
  
mechanism. T21 , T32 , T43 , T34 in Eq. (2) are with joint angles base frame:
(θ2 , θ3 , θ4 , θ3 = −83◦ × (π/180◦ )), where the passive joint
variables θ3 , θ4 must be computed from θ2 and θ2 via x = J g, (8)

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252


Calibration of serial robots having closed-chain mechanisms 965

θ4 '
θ4' first items on the right-hand side of Eq. (12) are errors due to
x4 ' small variations of θ2 , θ2 only and thus can be obtained
a4' from derivatives of Eq. (6):
a2
a3'
θ3 = θ2 − θ2 ,
(13)
θ2
θ 3' θ3'
θ4 = −θ2 + θ2 .
x1 a2 '
θ 2' x2 ' Likewise, the second items on the right-hand side of Eq. (12)
are joint angle errors when link lengths slightly deviate
Fig. 2. A degenerated parallelogram.
from their nominal values. Considering ai = a i + ai , the


position constraints (4) become:


where x is a (6 × 1) vector made up three differential ori-
entations and three differential positions of the end-effector.  
(a 2 + a2 )cθ 2 + (a 3 + a3 )c(θ 2 3 + θ3 )
 

g is a (32 × 1) concatenated vector of geometric errors, 


+ (a 4 + a4 )c(θ 2 3 4 + θ3 + θ4 )

g = (θ, d, a, α, β)T where θ, a, α are (8 × 
1) vectors such that θ = (θ1 , θ2 , . . . , θ6 )T etc, d = − (a 2 + a2 )cθ 2 = 0,


(d1 , d3 , . . . , d6 )T , and β = (β2 , β3 , β4 )T ,   (14)


(a 2 + a2 )s θ 2 + (a 3 + a3 )s(θ 2 3 + θ3 )
 

where J is the (6 × 32) matrix relating x and g. Each 


+ (a 4 + a4 )s(θ 2 3 4 + θ3 + θ4

column in J represents a Jacobian of the tool’s pose error 
− (a 2 + a2 )s θ 2 = 0,

x with regards to a particular parameter in g, i.e.,
Ja2 is the Jacobian related to parameter a2 and so on.
The formulations of Jθi , Jdi , Jai , Jαi , Jβi , i = 1 . . . n, for a where:
general n-link serial manipulator are given as:5,12
a 2 = a 3 , a 4 = a 2 ,
   
     
zi−1 0 0     
Jθi = , Jdi = , Jai = , θ 2 3 = θ 2 + (θ 2 − θ 2 ) = θ 2 ,
zi−1 × pi−1 zi−1 xi     
    θ 2 3 4 = θ 2 + θ 3 + π − θ 3 = π + θ 2 .
xi yi
Jαi = , Jβi = , (9)
xi × pi yi × pi Simplifying Eq. (14) by ignoring high-order terms and using
the linearization forms of cosine and sine functions:
where xi , yi , zi are directional vectors and pi is the position
  
of frame i in the base frame. Obviously, the model (8) is not c(θ i + θi ) = cθ i − s θ i θi ,
yet complete due to parameter redundancies between frames    (15)
s(θ i + θi ) = s θ i + cθ i θi ,
2 , 3 , 4 , 3 and thus needs subsequent modifications.
When there are errors in parameters of the parallelogram
structure (Fig. 2), the actual passive joint angles θ3 , θ4 devi- and we finally obtain:
 
ate from their nominal values θ 3 , θ 4 computed in Eq. (6) as: θ3 = m11 (a3 − a2 ) + m12 (a4 − a2 ),
(16)
 θ4 = m21 (a3 − a2 ) + m22 (a4 − a2 ),
θ3 = θ 3 + θ3 ,
 (10)
θ4 = θ 4 + θ4 . where:
 
In fact, they relate to other parameters of the mechanism as 1 c(θ 2 − θ 2 ) 1 1
follows: m11 =  .  
, m12 = −  .  
,
a 2 s(θ 2 − θ 2 ) a 2 s(θ 2 − θ 2 )
θ3 = f (θ, a), 1 1
 
1 c(θ 2 − θ 2 ) (17)
(11) m21 =  .  m 22 = −  . 
θ4 = h(θ, a),  
a 2 s(θ 2 − θ 2 ) a 2 s(θ 2 − θ 2 )
where θ = (θ2 , θ2 ), a = (a2 , a2 , a3 , a4 ), and f and h −m11 , −m12
are position functions of a general four-bar linkage (see
Appendix A1). Combining Eqs. (13) and (16) yields:
With small θ and a, errors in passive joint angles θ3 ,
θ4 can be derived from the linearization of Eq. (11): θ3 = θ2 − θ2 + m11 (a3 − a2 )
  + m12 (a4 − a2 ),
∂f  ∂f  θ4 = θ2 − θ2 + m21 (a3 − a2 )
(18)
θ3 = θ + a.
∂θ a=0 ∂a θ=0 + m22 (a4 − a2 ),
(12)
∂h  ∂h 
θ4 = θ + a.
∂θ a=0 ∂a θ=0 with mij given in Eq. (17). Replacing θ3 and θ4 in
Eq. (8) by Eq. (18) and then rearranging the resulting
This equation can be obtained without actually linearizing matrix equation, we obtain the desirable kinematic model
f and h, which are highly nonlinear and complicated. The for serial manipulators having the parallelogram mechanism.

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252


966 Calibration of serial robots having closed-chain mechanisms
Table II. Identifiable parameters.
y5
Link i gi Link i gi
d6 θ6
l0∗ a0 , b0 , α0 , β0 l3 d3 , a3 , α3 z7, z5
l1 θ1 , d1 , a1 , α1 l4 θ4 , d4 , a4 , α4
d7 a6
l2∗∗ θ2 , a2 l5 θ5 , d5 , a5 , α5
l2∗∗ θ2 , a2 , α2 , β2 l6 θ6 , d6 , a6 , α6 θ 7
α 6
l3∗∗ a3 , α3 , β3 l7∗ θ7 , d7
l4∗∗ (none) Total 34 z6, y7 x6,x7 x5
(∗ : parameters of the base and tool transformations, ∗∗ : parameters
of the parallelogram mechanism). Fig. 3. The six parameters to compensate for errors in tool
transformation.
The Jacobians for corresponding parameters in Eq. (18) are:
With reference to Eq. (7), one might see that of the
Jθ2 = Jθ3 − Jθ4 , six parameters used to model errors in BASE, only
Jθ2 = Jθ2 − Jθ3 + Jθ4 , a0 , b0 , α0 , β0 are independent while θ0 , d0 are
Ja2 = Ja2 − m12 Jθ3 − m22 Jθ4 , grouped into θ1 , d1 , respectively. Their Jacobians are
(19)
Ja4 = Ja4 + m12 Jθ3 + m22 Jθ4 , given in Appendix A2. To account for errors in TOOL, a
Ja3 = Ja3 + m11 Jθ3 + m21 Jθ4 , virtual DH-based frame 7 is inserted after frame 6 such
Ja2 = −m11 Jθ3 − m21 Jθ4 , that α6 = −90◦ , α7 = 90◦ , and a7 = d7 = 0 following the
suggestion of ref. [13]. As depicted in Fig. 3, errors in TOOL
where Jgi on the right-hand side was previously calculated can be modeled with three positions d6 , a6 , a7 and three
following Eq. (9). orientations θ6 , α6 , θ7 , of which only a7 , θ7 need to
be identified.
2.3. The identifiable parameters
Links 1, 4, 5, and 6 do not belong to the closed-loop chain,
and hence their error parameters are safe from redundancy. 3. Simulation
The remaining redundancy is due to the design of the A simulation study was conducted to validate the proposed
parallelogram structure whose opposite links are nominally calibration model. To begin with, simulated dimensional and
parallel. Thanks to the simple analytic form of the Jacobian angular errors in the range of ±4 × 10−3 m and ±4 × 10−3
matrix, the redundant error parameters of the loop can be rad were added to the nominal robot parameters. Next, the
simply determined: Since links 4 and 2 are designed to be passive joint angles θ3 ,θ4 were computed from the simulated
parallel and x4 = −x2 , it can be seen from Eq. (19) that θ2 , θ2 and link lengths a2 , a2 , a3 , a4 following the position
Ja4 = −Ja2 , thus, a4 and a2 are dependent (a3 and equations of a four-bar linkage provided in Appendix A1.
a2 are not though). Likewise, it is also possible to prove As a result, the actual tool positions differ with the ones
that θ3 and θ2 are dependent, while (α4 , β4 ) are computed by the nominal kinematic model by about 11 mm.
grouped into (α3 , θ4 ). The redundance must be omitted If the proposed error model, denoted as model 1, succeeds,
from the error vector g and their corresponding columns identified error parameters will be identical with the
must be discarded from the Jacobian matrix J . Resulting simulated ones and tool pose errors will be significantly
identifiable parameters are given in Table II. Notice the reduced. The calibration result of model 1 will be compared
existence of θ2 , a2 even though link 2 is not part of the with result of another model, namely model 2, implemented
chain (12 . . . 6). based on the method given in ref. [8]. In this paper, the
Eq. (A1) was differentiated directly to obtain an error model
2.4. Base and tool transformations of the parallelogram structure, which is similar to Eq. (18).
In case the base and tools transformations require calibration, The identification algorithm for both models is described in
the kinematic model (1) is written as: Fig. 4. The condition number of the two regression matrices
is 92.04 < 100 indicating that both of these identification
T = BASE · T60 · TOOL, (20)
systems are well-conditioned.5 The least-square solutions of
the two models are solved by the Gauss–Newton method.
where BASE and TOOL are transformations defining the
The simulation result is shown in Fig. 5. Model 1
robot base frame w.r.t. the metrology frame and the probe
converged quickly after five iterations and the average
frame w.r.t. the tool frame, respectively. Of the 12 parameters
residual error is 3.4 × 10−5 mm. Model 2 converged much
(three orientations and three positions for each) required to
slower, taking more than 400 iterations to obtain the
model errors in BASE and TOOL, only six are identifiable,4,6
comparable result. This can be explained by that fact that
and therefore, these two transformations require proper
the linearization error of model 1 is mild since it is only due
coordinate arrangements. The BASE transformation can be
to sine and cosine functions while that of model 2 is much
set up as:
more severe. In addition to the speed of convergence, Model 1
offers the advantage of simplicity whereas the derivation of
BASE = Tran(x, a0 )Tran(y, b0 )Rot(x, α0 )Rot(y, β0 )
Model 2 is a tedious work, even with the aid of symbolic
Rot(z, θ0 )Tran(z, d0 ). (21) Matlab programming. Apart from these, the error parameters

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252


Calibration of serial robots having closed-chain mechanisms 967

l4
Start

Initialize link parameters g a3


(Table 1, 2) link 4
link 2 a2
P4

Compute differential vector Δ x τ2


P2
l2 θ2
Update parameters: A lw aw
g = g + Δg x1
Compute Jacobian Matrix J τ 2'
(except θ 3', θ 4' use [Eqs. (9), (19)] θ 2'
Eqs. A1.1, A1.2) Pw
Solve linear least-square for Δg
Fig. 6. Free body diagram of forces in the x1 y1 plane of frame 1
based on which the joint compliance was modeled.
N
Converge ?
and the counterweight and l2 , a2 , l4 , a3 , lw , aw (m) are offsets
Y between joint axes and gravity centers. Equation (22) can be
Stop written as:
θ2C = G1 cθ2 ,
Fig. 4. Flow chart for kinematic identification of manipulators (23)
having the parallelogram mechanism.
θ2 C = G2 cθ2 + G3 sθ2 ,
where the dimensionless constants G1 , G2 , and G3 :
identified by the two models are almost identical, as depicted G1 = C2 (P2 l2 + P4 a2 ),
in Table III. It can be seen that parameters of the loop (links G2 = Pw lw − P4 l4 , (24)
2 3 4 3) were identified with accuracy better than 10−5 m and G3 = Pw aw − P4 a3 ,
10−5 rad while those of the open chain were identified with
higher accuracy. are identified by merging Eq. (23) with Eq. (8). The Jacobians
of G1 , G2 , and G3 are given as:
4. Compensation for Deflections due to Link Weights JG1 = Jθ2 cθ2 ,
The proposed model can be expanded with the inclusion of JG2 = Jθ2 cθ2 , (25)
nongeometric errors, for example, joint elasticity induced by JG3 = Jθ2 sθ2 ,
the robot’s heavy structure. Deflections mostly occur at joints
2 and 2 due to the weights of links 2, 4 and the counterweight where Jθ2 and Jθ2 on the right-hand side were calculated as
at the end of link 2 (Fig. 6). These joints can be modeled as shown in Eq. (19).
linear springs with constant stiffnesses,15 thus, the deflections
θ2C , θ2 C due to gravitational loading are formulated as:

θ2C = C2 τ2 = C2 (P2 l2 + P4 a2 )cθ2 , 5. Experiment Results and Discussions


θ2 C = C2 τ2 = C2 ((Pw lw − P4 l4 )cθ2 (22) In Section 5, we describe calibration experiments with
+ (Pw aw − P4 a3 )sθ2 ), the Smart H4 manipulator. The robot repeatability quoted
by Comau is ±0.3 mm, serving as the upper bound for
In Eq. (22), C2 , C2 (rad/Nm) are the compliance coefficients calibration results. Robot tool poses were observed using
of joints 2 and 2 , τ2 , τ2 (Nm) are the joint torques generated a Leica laser tracker equipped with a reflector mounted
as the result of P2 , P4, Pw (N): the weights of links 2 and 4 at the robot flange surface via an L-shape aluminum

Fig. 5. (Colour online) Simulation results.

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252


968 Calibration of serial robots having closed-chain mechanisms
Table III. Identified parameters.

Model 1 (after 7 iterations) Model 2 (after 450 iterations)


Link i  Input Identified Error Identified Error Note about identified values

1 θ1 −1 −0.9999 5.72e−6 −0.9999 3.90e−6


d1 −4 −4.0000 −2.67e−5 −4.0000 1.64e−5
a1 −1 −1.0000 −1.04e−5 −0.9999 1.58e−5
α1 4 4.0000 −2.33e−6 4.0000 9.26e−6
2 θ2 3 2.9905 −9.54e−3 2.9905 −9.54e−3
a2 −1 −1.0000 −1.61e−5 −1.0001 −8.50e−4
2 θ2 −2 −1.9992 8.38e−4 −1.9992 8.43e−4 θ2 + θ3
d2 −3 Dependent parameter, identified via d3
a2 −2 2.0089 8.91e−3 2.0089 8.91−3 a2 − a4
−3
α2 3 2.9962 −3.77e 2.9962 −3.77e−3
β2 0 −4.6e−6 0 −1.61e−6 0
3 θ3 Position-dependent error
d3 −2 Dependent parameter, identified via d3
a3 −3 −3.0089 −8.88e−3 −3.0089 −8.98e−3
−5
α3 2 2.0000 −2.05e 2.0000 −1.96e−5
β3 0 −2.16e−5 0 −1.48e−5 0
4 θ4 Position-dependent error
d4 −1 Dependent parameter, identified via d3
a4 −4 Dependent parameter, identified via a2
α4 1 Dependent parameter, identified via α 3 and θ 4
β4 0 Dependent parameter, identified via α 3 and θ 4
3 θ3 0 Dependent parameter, identified via θ 2
d3 4 −2.0024 −0.24e−3 −2.0024 −0.24e−3 d2 + d3 + d4 + d3
−3
a3 1 0.9949 −5.05e 0.9949 −5.00e−3
α3 −1 −0.8785 −4.17e−4 −0.8785 −4.13e−4 α3 + α4 cos(7◦ )
−3 −3
4 θ4 3 3.9888 −3.73e 3.9888 −3.74e θ4 + α4 sin(7◦ )
−4 −4
d4 3 3.0004 3.66e 3.0004 3.52e
a4 2 2.0000 1.31e−5 2.0000 4.67e−6
−6
α4 −2 −2.0000 −1.01e −2.0000 −9.05e−6
−6
5 θ5 2 2.0000 2.57e 2.0000 −8.22e−7
−6
d5 2 2.0000 −3.06e 2.0000 1.29e−6
−6
a5 3 3.0000 −9.90e 3.0000 −9.04e−6
α5 −3 −3.0000 1.10e−6 −3.0000 7.94e−6
6 θ6 1 1.0000 −9.67e−7 1.0000 1.10e−6
d6 1 1.0000 4.53e−6 0.9999 −5.00e−6
−6
a6 4 4.0000 −2.31e 4.0000 −8.88e−6
−6
α6 −4 −4.0000 −1.63e −4.0000 3.15e−6
(Units: length: 10−3 m, angles: 10−3 rad)

bracket (Fig. 7). The reflector, namely the Leica TMAC geometric parameters described in Sections 2.3 and 2.4. The
probe, provides full pose [6 d.o.f. (or degree of freedom)] third and fourth models are extended from the first and second
measurements with the position and orientation accuracy ones with the compliance parameters G1 , G2 , G3 presented
better than 5 × 10−2 mm and 10−2 deg. in its near field. The in Section 4. Parameters of models 1, 2, 3, and 4 were
robot internal parameters along with those in the BASE (the solved by the Gauss–Newton method. Accuracy measures
robot base w.r.t. laser tracker frames) and the TOOL (the between the models are the residual errors in positions
TMAC (Tracker - Machine control sensor) w.r.t. TCP (Tool and orientations (Roll–Pitch–Yaw angles) between the laser
centre point) frames) transformations will be calibrated from tracker’s measurements and the four models’ estimates.
a set of 90 robot configurations in its workspace. It can be seen from the calibration results given in Table IV
To demonstrate how the proposed method improves robot that all the models were able to predict the orientation parts
accuracy, we performed parameter identification with several accurately. This can be explained as parameters of the last
calibration models. In the first model, the robot was regarded three links that control the TCP orientations are identical in
as a standard serial-link manipulator (by computing Eq. (1) the four models. On the other hand, differences between the
with Eq. (3), thus, neglecting the parallelogram structure). models: the parallelogram and joints 2, and 2 compliance
This model uses 6 × 4 + 6 = 30 “complete and minimal” parameters only, affect the position accuracy of the TCP.
parameters.6 The second is the proposed model using 34 From Table IV, the mean value of residual position errors

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252


Calibration of serial robots having closed-chain mechanisms 969
Table IV. Residual position and orientation errors of the models.

Position errors Orientation errors


Models Mean p Std. dev. σp Mean θX Std. dev. σθX Mean θY Std. dev. σθY Mean θZ Std. dev. σθZ

Nominal 5.03 0.66 2.380 1.283 0.787 0.571 3.530 0.917


Model 1 1.14 0.60 0.015 0.013 0.026 0.020 0.021 0.016
Model 2 1.08 0.38 0.015 0.011 0.020 0.014 0.020 0.015
Model 3 0.78 0.35 0.015 0.010 0.020 0.010 0.020 0.010
Model 4 0.38 0.18 0.008 0.007 0.008 0.005 0.010 0.007
(Units: length: 10−3 m, angles: degree)

outperformed 3 by 100% even though they utilized the same


deflection error model. Since joint deflections θ2C , θ2 C
that contribute to the accuracy improvements of these models
are functions of joint angles θ2 and θ2 [see Eq. (23)], it can
be inferred that θ2 and θ2 were better estimated by Model 4,
thanks to the error model of the parallelogram structure.
Therefore, the proposed modeling does improve calibration
accuracy. Estimated geometrical errors of the parallelogram
mechanism and their standard deviations are given in Table V.

6. Conclusion
The paper addresses modeling issues for serial robots
Fig. 7. (Colour online) Calibration experiment with the Comau having closed-loop chains. We have presented a method
Smart H4 robot. for modeling errors in both the main open-loop and
initially was 5.03 mm (before calibration) then reduced to the subkinematic chain of the manipulators based on
1.14, 1.08, 0.78, and 0.38 mm with the four models. There linearization of the constraint equations. The advantage
is only a slight improvement in the result of Model 2 over of the proposed model is its simplicity compared with
Model 1 because they both deteriorated due to deflection error other competitive models and the possibility to apply
[(Fig. 8(a)]. When this nongeometric error was eliminated, existing calibration techniques of standard serial-link robots
the difference became much clearer [Fig. 8(b)]. Model 4 to manipulators of this type, as has been demonstrated

Fig. 8. (Colour online) Calibration results of the Smart H4 robot obtained from the four models: serial link model (Models 1 and 3) and
proposed model (Models 2 and 4).

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252


970 Calibration of serial robots having closed-chain mechanisms
Table V. Parameters of the parallelogram structure identified by Model 4.

Link i θ (σθ ) d(σd ) a(σa ) α(σα ) β(σβ )

2 −0.05 (0.02) n/a −0.06 (0.02) 0.02 (0.01) 0.01 (0.006)


2 −0.56 (0.02) n/a −1.52 (0.37) n/a n/a
3 n/a n/a −1.12 (0.37) −0.01 (0.004) 0.01 (0.003)
4 n/a n/a n/a n/a n/a
3 n/a −0.98 (0.25) 0.34 (0.17) −0.02 (0.007) n/a
(Units: length: 10−3 m, angles: degree)

throughout this work by the use of analytic formations 3. W. Khalil, Modelling, Identification and Control of Robot
of the Jabobian matrix to analyze parameter identifiablity (Taylor & Francis, New York, 2002) pp. 257–290.
(Section 2.3), joint deflection error model (Section 4), 4. W. Khalil, S. Bernard and P. Lemoine, “Comparison study of
the geometric parameter calibration methods,” Int. J. Robot.
and linear least square method for parameter identification Autom. 15(2), 56–67 (2000).
(Sections 3 and 4). Calibration experiments were performed 5. J. Hollerbach, W. Khalil and M. Gautier, Handbook of Robotics
with a large and heavy Comau Smart H4 robot containing (Springer, London, UK, 2007) pp. 321–344.
a parallelogram mechanism. When both geometric and 6. K. Schröer, S. Albright and M. Grethlein, “Complete, minimal
deflection errors were identified, the proposed model and model-continuous kinematic models for robot calibration,”
Robot. Comput. Integr. Manuf. 13(1), 73–85 (1997).
improved calibration accuracy by 100%, compared with the 7. K. Schröer, S. Albright and A. Lisoukin, “Modeling closed-
conventional model that regards the robot as a simple open- loop mechanisms in robots for purposes of calibration,” IEEE
loop structure. J. Robot. Autom. 13(2), 218–229 (1997).
8. S. Marie and P. Maurine, “Elasto-Geometrical Modelling
of Closed-Loop Industrial Robots Used For Machining
Appendix Applications” IEEE International Conference on Robotics and
Automation, Pasadena, CA, USA (May 19–23, 2008) pp. 1294–
A1. Angular positions of a four bar linkage 1300.
9. G. Alici and B. Shirinzadeh, “A systematic technique to
Based on ref. [16], θ3 andθ4 of a four-bar linkage (Fig. 2) estimate positioning errors for robot accuracy improvement
are computed as follows. Defining: using laser interferometry based sensing,” Mech. Mach. Theory
40, 879–906 (2005).
x = a2 cθ2 − a2 cθ2 , 10. S. P. Ananthananyanan, C. Szymczyk and A. Goldenberg,
y = a2 sθ2 − a2 sθ2 ,  “Identification of Kinematic Parameters of Multiple Closed
 Chain Robotic Manipulators Working in Coordination,” IEEE
ϕ2 = atan2(y,
 x) + acos a4 − x − y − a3 /
2 2 2 2
International Conference on Robotics and Automation, Nice,
2a3 x 2 + y 2 , France (1992) pp. 358–363.
ϕ1 = atan2((y + a3 sϕ2 )/a4 , (x + a3 cϕ2 )/a4 ), 11. B. Siciliano, L. Sciavicco and L. Villani, Robotics Modelling,
Planning and Control (Springer, Berlin, 2009) pp. 70–72.
12. D. J. Bennett and J. Hollerbach, “Autonomous calibration of
then: single-loop closed kinematic chains formed by manipulators
with passive endpoint constraints,” IEEE J. Robot. Autom.
θ3 = ϕ2 − θ2 , 11(5), 597–606 (1995).
(A1)
θ4 = π + ϕ1 − ϕ2 . 13. W. K. Veitsschegger and C. H. Wu, “Robot calibration and
compensation,” IEEE J. Robot. Autom. 4(6), 643–656 (1988).
14. M. R. Driels and U. S. Pathre, “Simulation experiments on
A2. Sub-Jacobian matrix for a0 , b0 , α0 , β0 parameter identification for robot calibration,” J. Adv. Manag.
The Jacobians for a0 , b0 , α0 , β0 can be derived with Tech. 5, 13–33 (1990).
the aid of symbolic Matlab programming as: 15. R. P. Judd and A. B. Knasinski, “Technique to calibrate
⎡ ⎤ industrial robots with experimental verification,” IEEE J.
0 0 1 0 Robot. Autom. 6(1), 20–30 (1990).
⎢0 0 cα0 ⎥
16. Integration Engineering Lab. at University of California,
⎢ 0 ⎥ “Four Bar Linkage” http://iel.ucdavis.edu/chhtml/toolkit/
⎢ ⎥
⎢0 0 0 sα0 ⎥ mechanism/fourbar/fourbarpos.html, visited on 12/10/2011.

JBASE = ⎢ ⎥. (A2) 17. B. W. Mooring and G. R. Tang, “An improved method for
⎢1 0 0 b0 sα0 ⎥
⎥ identifying the kinematic parameters in a six axis robot,” Proc.
⎢ ⎥
⎣0 1 0 −a0 sα0 ⎦ Int. Comput. Eng. Conf. Exhibit., 1, 79–84 (1984).
18. S. Hayati and M. Mirmirani, “Improving the absolute
0 0 −b0 a0 sα0 positioning accuracy of robot manipulators,” J. Robot. Syst.,
2(4), 397–413 (1985).
19. T. W. Hsu and J. L. Everett, “Identification of the Kinematic
Parameters of a Robot Manipulator for Positional Accuracy
Improvement,” In: Proceedings of the International Computers
References in Engineering Conference and Exhibition, Boston, MA, USA
1. R. Bernhardt and S. L. Albright, Robot Calibration (Chapman (Aug. 4–8, 1985) pp. 263–267.
& Hall, London, UK, 1993). 20. W. Veitschegger and C. Wu, “A Method for Calibrating
2. B. Karan and M. Vukobratovic, “Calibration and accuracy and Compensating Robot Kinematic Errors,” In: IEEE
of manipulation robot models - An overview,” Mech. Mach. International Conference on Robotics and Automation,
Theory 29(3), 479–500 (1994). Raleigh, NC, USA, (1987) pp. 39–44.

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252


Calibration of serial robots having closed-chain mechanisms 971

21. M. R. Driels and U. S. Pathre, “Generalized joint model for 23. H. Zhuang, K. Wang and Z. S. Roth, “Error-model-based
robot manipulator calibration and compensation,” J. Robot. robot calibration using a modified CPC model,” Int. J. Robot.
Syst. 4(1), 77–114 (1987). Comput. Integr. Manuf. 10(4), 287–299 (1993).
22. H. W. Stone and A. C. Sanderson, “A Prototype Arm Signature 24. M. A. Meggiolaro, “An Analytical Method to Eliminate the
Identification System,” In: Proceedings of IEEE International Redundant Parameters in Robot Calibration,” In: Proceedings
Conference on Robotics and Automation, (1987) pp. 175– of ICRA, San Francisco, CA, USA (Apr. 2000) pp. 3609–
182. 3615.

http://journals.cambridge.org Downloaded: 11 Apr 2015 IP address: 169.230.243.252

You might also like