Computer-Aided Design 127 (2020) 102868

Contents lists available at ScienceDirect

Computer-Aided Design
journal homepage: www.elsevier.com/locate/cad

Reparameterization of Ruled Surfaces: Toward Generating Smooth

Jerk-minimized Toolpaths for Multi-axis Flank CNC Milling
∗
Ali Hashemian a , , Pengbo Bo b , Michael Bartoň a,c
a
BCAM – Basque Center for Applied Mathematics, Alameda de Mazarredo 14, 48009 Bilbao, Basque Country, Spain
b
School of Computer Science and Technology, Harbin Institute of Technology, West Wenhua Street 2, 264209 Weihai, China
c
Ikerbasque – Basque Foundation for Sciences, Maria Diaz de Haro 3, 48013 Bilbao, Basque Country, Spain

article info a b s t r a c t

Article history: This paper presents a novel jerk minimization algorithm in the context of multi-axis flank CNC ma-
Received 3 April 2020 chining. The toolpath of the milling axis in a flank milling process, a ruled surface, is reparameterized
Received in revised form 18 April 2020 by a B-spline function, whose control points and knot vector are unknowns in an optimization-based
Accepted 25 April 2020
framework. The total jerk of the tool’s motion is minimized, implying the tool is moving as smooth
Keywords: as possible, without changing the geometry of the given toolpath. Our initialization stage stems from
Jerk-minimized toolpath measuring the ruling distance metric (RDM) of the ruled surface. We show on several examples that
B-spline reparameterization this initialization reliably finds close initial guesses of jerk-minimizers and is also computationally
Ruled surface efficient. The applicability of the presented approach is illustrated by some practical case studies.
Ruling distance metric (RDM) © 2020 Elsevier Ltd. All rights reserved.
Multi-axis flank CNC milling

1. Introduction lowers deviation of the machined surface from its (ideal) designed
state. Moreover, high values of jerk (as the time derivative of ac-
Multi-axis computer numerically controlled (CNC) machining celeration) are closely related to sudden changes of drive torques
is the leading subtractive manufacturing technology, especially which induce undesirable vibrations [3]. Last but not least, small
for hard materials such as titanium or steel–nickel alloys. The jerk reduces the dynamic cutting forces that, in a consequence,
machining process has typically three stages: (i) roughing, that leads to lower energy consumption of the whole milling process.
quickly removes the waste material, e.g. using a large ripper cut-
ter, and generates a rough shape, (ii) semi-finishing, that is mostly 1.1. Previous work
realized by end milling with a smaller cutter than for roughing,
and (iii) finishing, that aims at finalizing the shape of the object, Most of research on jerk minimization relates to contour ma-
reaching fine machining tolerances (e.g. for engine/aeronautical chining [4–7]. Zhang et al. [4] use a G4 -continuous B-spline curve
components a few micrometers for objects of the size of tens of for the toolpath and employ a jerk-smooth feedrate scheduling
centimeters). Our research focuses on this final (finishing) stage of process in order to obtain a smooth jerk profile along the tool-
machining. In particular, we consider 5-axis flank milling where path. By using an optimal control problem formulation, Bosetti
the milling cutter (aka tool) and the to-be-manufactured object and Bertolazzi [5] present an algorithm for CNC kernels that
touch along a whole curve (i.e., grazing curve) [1], in contrast to aims at solving the axes interpolation problem. They assume a
(i) and (ii) where the tool has (theoretically) a single-point contact predefined path tracking tolerance and apply constraints on the
with the material block. kinematic specifications of the machine axes to calculate the
There are many factors that influence quality of a machined trajectory that satisfies the given constraints.
product: from geometric accuracy of the path-planning algo- Jahanpour and Alizadeh [6] present an adaptive acceleration-
rithms, over analysis of cutting forces and calibration of the actual jerk-limited NURBS interpolation method based on an optimized
machine, up to the error associated to the tool wear. In this work, S-shaped C 2 quintic feedrate planning scheme. They modify the
we discuss the kinematic part of the path planning process that feedrate profile for each sharp corner to minimize the total execu-
deals with the physics (velocities, acceleration, and jerks) of the tion time. Using a piecewise constant approximation of the third
tool [2]. Having a smooth and jerk-minimized profile in multi-axis derivative of the path, Zhang et al. [7] perform a time-optimal tool
machining is fundamental as it affects tracking performance and trajectory generation for CNC manufacturing systems. In their
approach, the axis jerk constraints are also introduced into the
∗ Corresponding author. problem where the desired smoothness of the path could be
E-mail address: [email protected] (A. Hashemian). accomplished by adjusting the values of constraints.

https://doi.org/10.1016/j.cad.2020.102868
0010-4485/© 2020 Elsevier Ltd. All rights reserved.
2 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868

Beudaert et al. [8] use 5-axis toolpath smoothing techniques

in order to maximize the feedrate and to reduce the machining
time in both end and flank milling processes. Based on the ve-
locity, acceleration, and jerk limits of each drive, they compute
the maximum reachable feedrate which is then used to localize
the areas where the toolpath is smoothed. In a similar research,
Sencer et al. [2] present a feed optimization process to minimize
the machining time based on the drive constraints. They use a
B-spline model for the feedrate profile and find its unknown
control points as design variables of an optimization process.
An optimization method for obtaining a minimum time fee-
drate profile that takes into account the kinematic characteris-
tics of a particular machine is proposed by Beudaert et al. [9].
They consider both tangential and axis jerk and, in an itera-
tive algorithm, compute the minimum time feedrate profile by Fig. 1. Toolpaths of 5-axis flank milling with conical tools, represented by ruled
surfaces (yellow). The envelopes of the tool (transparent) accurately approximate
simultaneously satisfying the drive constraints.
a benchmark freeform workpiece (the path-planning results are courtesy of Bo
A feedrate scheduling method for the 5-axis machining with et al. [21]). On the left path, observe distribution of the milling tools that is
geometry, process, and drive constraints is presented by Sun uniform in time, yet quite non-uniform in space, and would require higher
et al. [10]. The initial feedrate profile is constructed using a chord acceleration and/or jerk at the upper end of the toolpath.
error, and then iteratively adjusted by a curve evolution strategy.
Hu and Tang [11] consider dynamic behavior of the milling
tool in the context of high-speed machining. A tool path is gen- 1.2. Contributions
erated such that the maximal angular accelerations of the rotary
axes are reduced. By defining smoothing blocks on the toolpath of
Our main contributions are as follows:
CNC contour machining, Essid et al. [12] perform a local toolpath
smoothing. In their kinematic model, they took into account the
• We propose utilizing the reparameterization technique upon
drive parameter axes defined by the manufacturer of the CNC
which the relationship between the motion path parame-
machine, however, in contrast to our work, the milling path is
ter of the flank machining toolpath and the system time
being modified. In a closely related research, Song and Ma [13]
variable is expressed by means of an optimal transfer func-
perform an interval partition-based feedrate scheduling with ax-
ial drive constraints for the 5-axis flank machining. Ma et al. [14] tion such that the total jerk of the tool’s motion is min-
use a dual NURBS interpolator to detect feedrate-sensitive and imized. To this end, we use a B-spline transfer function
nonsensitive regions and design constant and smooth transition whose control points and the knot vector are variables in
speeds in these regions. an optimization-based framework.
Huang et al. [15] present a real-time feedrate planning method • We introduce a new initialization strategy using a proper
for 5-axis machining to synchronize the linear and angular trajec- ruling distance metric (RDM) to adequately estimate dis-
tories, in order to reach smooth linear and angular motions con- tances between neighboring positions of the milling tool.
sidering axial kinematic constraints. Recently, Weng and Kuo [16] This initialization serves as a good guess for global opti-
have proposed a jerk decision making process that synchronize mization in a discretization of the space of transfer functions
the path planning stage with the dynamic response for a 5-axis (reparameterizations).
machine tool. Such a synchronization results in better surface • Given a milling toolpath for multi-axis flank milling, rep-
finish and also shortens the machining time. resented as a B-spline ruled surface, our proposed research
Another relevant family of research deals with optimization of preserves the geometric accuracy and optimizes the motion
tool orientations. Farouki and Li [17] consider a tool-orientation of the tool. Our results can therefore be easily combined
problem in the context of 5-axis ball-end milling and look for with any toolpath design strategies.
rotation-minimizing tool motions that preserve the angle be- • We compare our results with state of the art path plan-
tween the tool axis and the surface normal. Sun et al. [18] also
ning techniques and show on several case studies that the
propose a tool orientation adjustment method for 5-axis ball-
reparameterized paths reduce the total jerk considerably.
end machining to obtain an optimized toolpath considering the
kinematic constraints. For the given toolpath and feed profile
The remainder of this paper is organized as follows.
of the tool-end (both expressed as B-splines), they find analytic
Section 2 discusses ruled surfaces as toolpaths in flank machining
relations of angular feed, acceleration, and jerk with respect to
and introduces basic notions of the kinematics of the tool and
the geometric and tangential feed parameters of the tool-end
reparameterization. In Section 3, the toolpath jerk minimization
trajectory. Afterwards, conditional inequalities of the kinematic
constraints are applied to adjust the tool orientation. Given a problem is formulated and a ruling distance metric is introduced,
contact point, Sharma et al. [19] consider a multi-point tool followed by various case studies and numerical examples in
positioning and propose a method to adapt the tool to have a Section 4. Finally, Section 5 draws the conclusions of this work.
multiple contact with the reference geometry.
Finally, and probably in the closest research to ours, focusing
2. Ruled surfaces as toolpaths in flank CNC milling
on freeform curves expressed by B-spline and/or NURBS repre-
sentation, Hashemian et al. [20] present a NURBS reparameter-
ization technique to minimize the jerk value in trajectories of The tool motion in a flank CNC milling process can be repre-
robot manipulators. They reparameterize a single contact path sented by the position of the tool-end and the direction of the
(i.e., curve) and use a quintic B-spline function with uniform tool. Therefore, the toolpath is defined as the locus of the milling
knot vector for relating the path parameter to the execution time tool axis in the Euclidean 3-space, which is a ruled surface as
variable. shown in Fig. 1.
 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868 3

2.1. Ruled surfaces as toolpaths of rigid axes

For the sake of completeness, we begin with the definition of a

B-spline curve that can be thought of as the locus of the tool-end
positions in space. A B-spline space curve c (u) : u ∈ [0, 1] → R3
of degree n with N + 1 control points p0 , p1 , . . . , pN is defined
as:
N
∑
c (u) = Bin (u) pi , (1)
i=0

where the univariate parameter space u ∈ [0, 1] is characterized

by the knot vector u,

u = [0, 0, . . . , 0, un+1 , un+2 , . . . , uN , 1, 1, . . . , 1] , (2)

Fig. 2. A ruled surface s (u, v) with boundary curves c1 (u) and c2 (u) and straight
     
n+1 n+1
lines (i.e., rulings) of equal length (herein and in the following, u and v are
and the B-spline basis functions Bin (u) are expressed by the Cox– referred to as motion and ruling parameters, respectively).
deBoor recursion formula as follows [22]:
ui ≤ u < ui+1
{
1
Bi0 (u) =
0 otherwise
(3)
u − ui ui+n+1 − u
Bin (u) = Bin−1 (u) + Bin+−11 (u) .
ui+n − ui ui+n+1 − ui+1
Consider two B-spline curves c1 (u) and c2 (u) (referred
∑N n to ask
boundary curves) defined by Eq. (1), that is, ck (u) = i=0 Bi (u) pi
(k = 1, 2). We assume that c1 (u) and c2 (u) are of the same degree
n and spanned over the same knot vector u. A ruled surface
s (u, v) is then defined as
s (u, v) = (1 − v ) c1 (u) + v c2 (u), [u, v] ∈ [0, 1] × [0, 1] , (4)
where u is the parameter controlling the tool motion (i.e., pseudo-
time parameter) and v controls the position of the point on the
axis. The boundary curves are linearly interpolated and, therefore,
we have a straight line segment (referred to as a ruling) in the
v direction connecting the respective points on the curves c1 (u)
and c2 (u). If the boundary curves have different knot vectors, the
standard knot refinement algorithm is used to bring them to a
common u (see, e.g., [22]).
Fig. 3. Simulation of flank milling in a table-tilting 5-axis CNC machine. The
motions spanned by the X and Y directions are provided by translational
Remark 1. Alternatively, one can define a ruled surface by movements of the base (i.e., the table), while the Z motion is provided by vertical
one boundary curve and the motion of the ruling vector. This is translation of the spindle. The rotational movements B and C are controlled by
a more typical description of a ruled surface in scenarios such rotations of the table around Y and Z directions, respectively.
as ball-end milling where the contact paths are given and the
orientations are used, e.g., for collision avoidance. However, these
two representations of ruled surfaces are equivalent and one can Fig. 3). Considering that the toolpath is given by a B-spline ruled
easily convert one to another. surface, regardless of the machine type and how the motion law
In our context, the ruled surface is considered as a toolpath is governed by the G-codes, the kinematic specification of the tool
of the milling axis in flank machining, that is, the rulings repre- motion can be represented by means of velocity (V), acceleration
senting the tool axis have to be of equal length (see Fig. 2). This (A), and jerk (J) profiles of boundary curves. These entities are
constraint is formulated as computed as time derivatives of ck (u) (k = 1, 2) using a chain
rule as follows
∥ c1 (u) − c2 (u) ∥ = const. , for all u ∈ [0, 1] , (5) ∂ ck ∂ ck ∂ u
Vk (t) = = , (7)
and its differentiation with respect to u gives ∂t ∂u ∂t
)2
c1 ′ (u), c1 (u) − c2 (u) = c2 ′ (u), c1 (u) − c2 (u) , ∂ 2 ck ∂ 2 ck∂u ∂ ck ∂ 2 u
⟨ ⟩ ⟨ ⟩ (
(6)
Ak (t) = = + , (8)
where ⟨· , ·⟩ is the Euclidean scalar product. Eq. (6) is the first- ∂t 2 ∂ u2 ∂t ∂u ∂t2
order length-preserving constraint, also known as the projection ∂ 3 ck
( )3
∂ 3 ck ∂ u ∂ 2 ck ∂ 2 u ∂ u ∂ ck ∂ 3 u
rule (see, e.g., [23,24]). Jk (t) = = + 3 + , (9)
∂t3 ∂ u3 ∂ t ∂ u2 ∂ t 2 ∂ t ∂u ∂t3
2.2. Kinematics of the tool motion where t ∈ [0, τ ] is the system time variable.
In the above equations, the rth derivative of B-splines curves
The tool motion in the multi-axis CNC machining centers is is obtained as
represented in terms of G-codes and governed by the combi- N
nation of movements provided by the drives of multiple axes ∂ r ck ∑ ∂r
= B n (u) pik , (10)
(e.g., X, Y, Z, B, and C in the table-tilting 5-axis milling, see ∂u r ∂ ur i
i=0
4 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868

Fig. 4. Reparameterization. Left: A boundary curve in the typical and (arbitrarily) reparameterized forms. Red circles show the typical linear transformation between
u and t upon which equal time steps lead to equal steps on the parameter space, while green squares illustrate the parameters created by reparameterization upon
which the parameter step-lengths are governed by the transfer function u = f (t). Right: The respective linear (red) and general (green) transfer functions. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

where optimal smooth toolpath (noting that, herein and in the following,
∂ nr
n ∂ r −1 the optimality criterion is to reach a jerk-minimized toolpath).
B (u) = B n−1 (u) In the proposed method, along with the B-spline representa-
∂ ur i ui+n − ui ∂ ur −1 i
tion of the ruled surface, the reparameterization function u = f (t)
n ∂ r −1 n−1 is also introduced as a B-spline curve in terms of time variable.
− B (u) . (11)
ui+n+1 − ui+1 ∂ ur −1 i+1 B-spline formulation offers sufficient flexibility (e.g., arbitrary
The kinematic specifications of boundary curves provided by order of continuity in the function and its derivatives) for our jerk
Eqs. (7)–(9) will then be used to measure and optimize the minimization purpose. Hence, we recall Eq. (1) to define the TF
smoothness of the tool’s motion. as an m-th degree B-spline function with M + 1 control points
(i.e., coefficients) q0 , q1 , . . . , qM spanned over the knot vector t
2.3. B-spline reparameterization of the toolpath as follows:
M
∑
When the flank milling toolpath is parametrically given by a u = f (t) := Bim (t) qi , (12)
B-spline ruled surface, the G-code command lines are typically i=0
generated from a dense set of points sampled along the bound-
ary curves (or alternatively one boundary curve and the ruling t = [0, 0, . . . , 0, tm+1 , tm+2 , . . . , tM , τ , τ , . . . , τ ] . (13)
direction at each point). The procedure of developing command      
lines in CNC programmers (e.g., Siemens NX) generally considers m+1 m+1

equal time steps for tool motion between sampled points. An It should be noted that since the curve parameter u is sup-
important point to mention is that these sampling points are also posed to be in [0,1], the first and last control points of f (t) are
obtained by uniform parameter increments in most occasions. As q0 = 0 and qM = 1, respectively. All other qi s must lie within
a result, the execution system time domain t ∈ [0, τ ] is linearly the interval [0,1] and be monotone (i.e., qi > qi−1 ). The control
transferred to the motion parameter space u ∈ [0, 1] by means points and knot sequence of TF play crucial roles in smoothing
of a constant scaling coefficient (i.e., u = t /τ ). In this way, all the kinematic specifications of the toolpath. In the next section,
time derivatives of the trajectory are scaled by the powers of we will present an optimization-based framework that looks for
1/τ . However, a constant scaling cannot always guarantee a kine- the control points qi s and the knots tj s of the TF to minimize the
matically smooth motion especially at the start and end of the jerk of the toolpath.
path (see, e.g., [25]). Therefore, finding an optimal transformation
between t and u, which results in a smooth motion, is essential. 3. Toolpath jerk minimization
In this research, we propose utilizing the reparameterization
technique upon which the motion path parameter u on both 3.1. Optimization problem
boundary curves ck (u) is related to the time variable t by means
of a strictly increasing transfer function u = f (t) where f (0) = 0
As stated in the Introduction, the jerk minimization is taken
and f (τ ) = 1. Fig. 4 shows an original parameterization and its
into consideration in this research as a significant optimality cri-
reparameterized variant of a 3D curve. In the context of flank
terion in smoothing toolpaths of multi-axis flank CNC machining.
CNC machining, one needs to consider two boundary curves
Hence, in order to find the jerk-minimized toolpath, we need to
reparameterized by single transfer function (TF). It is interesting
determine the optimal reparameterized form of the respective
to note that, in contrast to linear transformation between u and
ruled surface, i.e., s̃ (t , v). Since the optimized boundary curves
t upon which equal time steps lead to equal steps on the pa-
c̃k (t) need to share the same parameterization to preserve the
rameter space, applying this reparameterization means that the
projection rule of Eq. (6), we seek a transfer function u = f (t)
parameter step-lengths (and their respective mappings on bound-
to be applied to both curves. For this purpose, we perform an
ary curves) are governed by the transfer function f (t). We also
optimization process in the context of functional analysis, i.e., we
recall a well-known fact that reparameterization does not change
minimize the jerk magnitudes on both boundary curves consider-
the shape of boundary curves (and consequently the ruled sur-
ing the reparameterization function f (t), represented by control
face), but affects solely their derivatives. More precisely, for each
points qi s and knot values tj s in Eqs. (12) and (13), as the un-
boundary curve, the original form ck (u) and reparameterized one
knowns. The objective function is defined as a weighted sum of
c̃k (t) := ck (f (t)) are geometrically identical but parametrically dif-
ferent (see, e.g., [20,22,26]). A similar scenario can be expressed L2 -norms of jerks of the respective boundary curves as
τ τ
for the ruled surface, i.e., s (u, v) ≡ s̃ (t , v). Consequently, by find-
∫ ∫
ing an appropriate expression for f (t), we are able to reach an F := w1 ∥J1 (t)∥2 dt + w2 ∥J2 (t)∥2 dt , (14)
0 0
 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868 5
2 ⟨
where Jk (t) = Jk (t), Jk (t) is the square of Euclidean norm of
 ⟩
the jerk. The weights w1 and w2 in the above integral might
be set based on different considerations, e.g., due to conditions
imposed on the boundary curves or due to special type of the
tool or milling machine. The optimization problem can be further
written as:

i = 1, 2, . . . , M − 1
{ }
qi
Find which minimize F
tjj = m + 1, m + 2, . . . , M
⎧
1
⎨ qi − qi−1 ≥
⎪
subject to α (M − m + 1)
τ
⎩ tj − tj−1 ≥
⎪
β (M − m + 1)
(15)

where α, β ≥ 1/(M − m + 1) are parameters to control the prox-

imity of adjacent control points and knots, respectively. Large Fig. 5. A 3D curve example reproduced from [20] for accuracy assessment of
values of α and β push the consecutive knots and control points the function evaluation procedure of the proposed toolpath jerk minimization
to be closer, and to eventually lower the continuity in the limit method.
(β → ∞). On the other hand, smaller values result in wider gaps
between the knots and/or the control points (for instance, for
knots, this constraint pushes the knot vector toward the uni- with the exact analytical one. We consider a random TF (i.e., non-
form case). If not stated differently, we set the default values to optimized) as its shape is not relevant to this exactness test. The
α = β = 10 in our implementation. The optimization constraints degree and number of control points of the TF are m = 5 and
express our wish of increasing sequence of knots and control M + 1 = 15, respectively. For performing the integration, various
points such that f (t) is well-defined and increasing. Finally, we numbers of Gauss points (GPs) are employed to evaluate the
also require the boundary constraints as q0 = 0, qM = 1, tm = 0, objective function.
and tM +1 = τ assuming the total execution time is τ . However, In the second part of the test, Table 2 shows the absolute
these boundary knots and values are set directly and therefore and relative errors of local (or so-called element-wise) Gaus-
are not included as variables in Eq. (15). sian integration of the objective function F for the same curve
where the integration is performed locally over the knot spans
Remark 2. Since obtaining the jerk profiles entails computing [tj , tj+1 ] of the transfer function. Again, we assume the degree
the third derivative of ck (u), employing quartic boundary curves of the TF is m = 5, but the number of control points varies as
for the toolpath would guarantee only a C 0 -continuous jerk. In M + 1 = 7, 11, and 15, so that we can observe the accuracy of
order to have a smooth jerk profile (at least C 1 -continuous), the element-wise numerical integration for different numbers of
ck (u) is required to be of degree n = 5 or higher. In addition, knot spans. For lower number of knot spans (i.e., lower M), we
referring to Eq. (9), and for the same reason, the transfer function need more GPs to reach an accurate integration since, in this
needs be at least of degree m = 5 to reach a smooth jerk–time case, the knot spans are wider and the integrand may oscillate
representation. more. On the other hand, by increasing the number of knot spans
of the TF, lower number of GPs is required for the element-
3.2. Evaluation of the objective function wise integration, guaranteeing fast and accurate evaluation of
the objective function for the iterative optimization process. The
In order to perform the jerk minimization, we need to cal- results of these tables indicate that local integration with 7–
culate the objective function F defined by Eq. (14). Depending 8 GPs can provide sufficient accuracy if a suitable number of
on the degrees of ck (u) and u = f (t), and based on the chain knot spans is considered for the TF. Another option to reduce
derivative rule of Eq. (9), the degree of the objective function can the computational cost of integration with such integrand of
grow drastically. For instance, if the TF and boundary curves are the polynomial spline type is to apply the spline Gaussian rules
all expressed in terms of quintic B-spline bases (i.e., n = m = 5), (see, e.g., [28,29]) that require even fewer quadrature points, yet
the degree of F would be 44. Since the analytical integration of guaranteeing the exactness of the integration.
functions of such high degrees is computationally expensive, and
needs to be computed repeatedly in an iterative optimization 3.3. Ruling distance metric
procedure, we prefer Gaussian quadrature rules as numerical
integration schemes, because they are known to be computa- In this research, we perform a gradient-based optimization
tionally cheap (see, e.g., [27]). To test the numerical exactness of by employing the sequential quadratic programming (SQP) as an
the integration, we experimented with various Gaussian rules as iterative method for constrained nonlinear optimizations, widely
follows. used on spline-based problems with a twice continuously differ-
Consider an example boundary curve that is a quintic B-spline entiable objective function (see, e.g., [30,31]). The SQP method
with N + 1 = 11 control points spanned over a uniformly-spaced solves a sequence of subproblems, each of which optimizes a
knot vector (see Fig. 5). It is assumed that this path, which is quadratic model of the objective function, guaranteeing a su-
a reproduction of a 3D example in [20], is to be traversed in perlinear convergence by using a quasi-Newton updating proce-
τ = 1 min. dure (for more details regarding the mathematical aspects of this
For the first part of the test, Table 1 shows the absolute and method, the readers are referred to, e.g., [32,33]). As the objective
relative errors of global Gaussian integration of the objective func- function in Eq. (14) is highly nonlinear, one needs to have a good
tion F for the 3D curve shown in Fig. 5 (herein, the term global initial guess of the unknown design variables (i.e., qi s and tj s) to
denotes the integration over the entire time domain [0, τ ]). The avoid the optimization process to be terminated at an undesirable
error values are obtained by comparing the numerical integral local minimum. Even by performing thousands of optimization
6 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868

Table 1
Accuracy assessment of global numerical integration of the objective function for the 3D curve shown in Fig. 5 with an
arbitrary (non-optimal) quintic TF with 15 control points (10 knot spans). Absolute and relative errors are obtained by
various numbers of Gauss points (# GPs) and the path is assumed to be traversed in τ = 1 min.
Analytical integration Global Gaussian quadrature Error
F (m2 /min6 ) # GPs F (m2 /min6 ) Absolute Relative
5 18239 17230 48.58%
10 39298 3828.4 10.79%
20 36528 643.0 2.99%
35470
30 34987 482.5 1.36%
40 35616 146.11 0.41%
50 35495 25.0 0.07%

Table 2
Accuracy assessment of local numerical integration of the objective function for the 3D curve example of Fig. 5 with arbitrary (non-optimal) quintic TFs with 7, 11,
and 15 control points (i.e., 2, 6, and 10 knot spans). Absolute and relative errors obtained by different numbers of Gauss points (# GPs) and the path is assumed to
be traversed in τ = 1 min.
#GPs 2 knot spans 6 knot spans 10 knot spans
F (m2 /min6 ) Error F (m2 /min6 ) Error F (m2 /min6 ) Error
Analytical Numerical Abs. Rel. Analytical Numerical Abs. Rel. Analytical Numerical Abs. Rel.
5 141830 84358 147% 37839 1552 4.28% 35797 327.72 0.92%
6 119950 62482 109% 35257 1030 2.84% 35321 148.91 0.42%
57472 36287 35470
7 87841 30370 52.8% 36618 331.7 0.91% 35507 37.35 0.11%
8 51604 5868 10.2% 36587 300.1 0.83% 35477 6.99 0.02%

runs with different random initial guesses for unknown variables,

there may be still the risk that the best optimal TF cannot be
captured. Furthermore, considering the computational cost of the
iterative optimization process, this cluster of runs is completely
far from the desired fast toolpath smoothing strategy that is the
main goal of this research.
In order to overcome the above-mentioned obstacles, an inno-
vative initialization technique for the jerk minimization process is
presented by means of an adequate ruling distance metric (RDM).
This metric is defined on a space of finite lines (i.e., rulings) in
3D-space and reliably captures distance between the positions of
the milling tool. Using this metric, we show that almost-optimal
jerk minimized path can be found with a single round of the Fig. 6. A bilinear patch (black) that approximates the ruled surface patch (blue)
optimization process. of s (u, v) between rulings ℓi and ℓi+1 . (For interpretation of the references to
Defining line ℓi := (ai , bi ) as a ruling that connects points color in this figure legend, the reader is referred to the web version of this
article.)
ai := c1 (ûi ) and bi := c2 (ûi ) on the boundary curves (where û is
an arbitrary parameter in [0,1]), for any set of two consecutive
rulings ℓi and ℓi+1 obtained by uniformly-spaced parameters ûi
and ûi+1 , and their uniformly-spaced time domain counterparts t̂i jerk values. Taking into account that most of CNC programmers
and t̂i+1 , we construct a bilinear patch on the ruled surface s (u, v) consider linear transformation between u and t, the generated
as shown in Fig. 6. We follow the idea of the metric of finite toolpath may reveal undesirable kinematic specifications if the
respective ruled surface is inappropriately parameterized. The
rulings [34] and define the RDM metric d(ℓi , ℓi+1 ) as a measure
main advantage of defining this metric is that we can reorient
for the distance traversed by the milling tool between rulings
the time domain based on the RDM magnitude to have larger
ℓi and ℓi+1 in the time step [t̂i , t̂i+1 ]. This metric is introduced
time steps for larger ruling deviations, leading to a good ini-
as the L2 -norm of the Euclidean distance function of rulings as
tialization for jerk minimization. Since the ruling distances are
follows (multiplied by 3 for convenience):
governed by the original parameterization of boundary curves,
1
this reparameterization implies that the tool goes on a shorter
∫
d(ℓi , ℓi+1 ) := 3 ∥(1 − v ) (ai − ai+1 ) + v (bi − bi+1 )∥2 dv time step when the rulings are close and on a larger one when
0
the rulings are further one from another. We herein refer to time
= ∥ai − ai+1 ∥2 + ∥bi − bi+1 ∥2 reorientation as an alternative definition of reparameterization
+ ⟨(ai − ai+1 ) , (bi − bi+1 )⟩ . (16) in which we keep the parameters equally-spaced while the time
steps are different (see Fig. 8). This definition will help to create
From the kinematic point of view, the RDM metric d is a the initial TF for the optimization process based on the RDM
measure that indicates how fast the tool is traveling between metric in an effective manner. Consider now Nd + 1 uniformly-
rulings in uniform time steps. Equally importantly, and unlike spaced parameters û0 , û1 , . . . , ûNd ∈ [0, 1] in Fig. 8, then in this
the Plücker’s metric of infinite lines (see, e.g., [35]), this metric discrete setup, the reoriented time values t̃i corresponding to each
captures sudden changes of one endpoint of the ruling, while the ûi are defined as follows:
other end stays (almost) fixed, see Fig. 7.
It is obvious that for larger RDM values, under uniform time ∑i
j=0 dj
steps, the tool should move faster to reach the next ruling (i.e., the t̃i := ∑N τ, i = 1, 2, . . . , Nd , (17)
next parameter û) that will result in higher acceleration and d
j=0 dj
 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868 7

Fig. 9. Left: An original path of a cylindrical milling tool (red) uniformly

sampled in time (we used 10 equal time steps for visualization of the tool). No
reparameterization is applied, i.e., u = t, on the boundary curves c1 (u) and c2 (u).
Fig. 7. RDM metric. Two milling paths with a conical tool (transparent) are Right: The toolpath after the jerk-minimized reparameterization is applied, c.f.
shown. While the Plücker’s metric gives almost the same value for the neigh- Figs. 10–13. The time-space t is sampled uniformly, while the parameter space is
boring positions of the milling axis (yellow), RDM assigns larger values to the obtained via u = f (t). Observe a more uniform distribution of the green cylinders
motion with the more intensive deviation of the axis (below). Consequently, low after reparameterization (with the same 10 equal time steps). The toolpaths
values of RDM capture better motions preferable for milling. correspond to 5-axis flank milling of a blade of the blisk model (framed) [23].
(For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

It will be shown in numerical examples that the kinematic spec-

ification (i.e., velocity, acceleration, and jerk) generally become
lower than the original ones after reparameterization.

4.1. The applicability of the ruling distance metric

The main goal of the first case study is to validate the correct-
ness of our proposed ruling distance metric (RDM) described in
Section 3.3 for the jerk minimization problem. For this purpose,
we explore the space of admissible transfer functions that should
be monotone and meet the conditions of u(0) = 0 and u(τ ) = 1.
∑M
Fig. 8. An example of the initial transfer function u = f (t) = m
i=0 Bi (t) qi for
We then sample a set of 5000 random initializations of u = f (t)
jerk-optimization process obtained by ruling distance metric. The function is for the optimization process and compare the results with that
fitted to equally-spaced parameters ûi corresponding to reoriented time values
obtained by the RDM metric. Fig. 9 demonstrates a toolpath of a
t̃i .
flank milling process on the turbine blisk benchmark workpiece.
The ruled surface has quasi-parallel rulings (i.e., with slight ruling
rotations) and is constructed by two quintic boundary curves with
where t̃0 := 0, d0 := 0, and dj := d(ℓj−1 , ℓj ). To get a smooth ini-
∑M m N + 1 = 15 control points spanned over a uniformly-spaced knot
tial TF : u = f (t) = i=0 Bi (t) qi for the jerk-optimization pro- vector. The toolpath of this example, which is assumed to be
cess, we apply B-spline fitting on the (t̃i , ûi ) values. In particular, traversed in τ = 5 s, is taken from [23].
we follow the scheme proposed in [22] and initialize the (non- Starting with the case of 5000 random initializations, Fig. 10
uniform) knot vector t by applying the De Boor’s algorithm and depicts the optimized TFs obtained by the optimization process,
initialize the control points qi by minimizing the fitting error in where equal weights are considered for the objective function F
the least-squares sense. (i.e., the total L2 -norms of jerks on both endpoints of the tool’s
axis). In this figure, the color coding reflects the quality with
4. Case studies respect to the quality measure, i.e., the value of the objective
function Eq. (14). The red color indicates large values of F while
In this section, the applicability of the proposed method of dark blue signals vanishing F . Observe that not all initial guesses
this paper is demonstrated on several benchmark examples. We converge to the true minimizers and get stuck at some local min-
again emphasize that the objective of this work is to kinematically imum, including the linear transformation depicted as a dashed
smooth a toolpath obtained a priori by some toolpath design black line. Most of the initial guesses converge to a roughly
strategies. We also assume that the given toolpath already meets similar TFs highlighted as a blue cluster at the middle of the
the axes constraints of a particular milling machine. It is impor- graph. In this example, we considered a quintic TF : u = f (t) (see
tant for the smoothing procedure not to violate these constraints. Remark 2) with M + 1 = 15 control coefficients qi where the
8 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868

Fig. 10. Optimal transfer functions u = f (t) for the first case study obtained
Fig. 11. Optimal transfer functions obtained by the proposed RDM metric with
by 5000 samples. The colormap shows how good the optimal TFs are with
different numbers of control points (# CPs) compared with that obtained by the
respect to the objective function value F (i.e., the total L2 -norms of jerks). The
random initializations. The respective values of the objective function (F ) and
black dashed diagonal line represents the linear parameter–time transformation
maximum jerk magnitudes are reported in Table 3.
(i.e., no reparameterization). (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

integrations are performed by element-wise Gaussian quadrature

with 7 GPs.
As a comparison with the random sampling of the initial
transition function, we employ the RDM metric to construct a
single initialization based on the geometry of the ruled surface.
Considering Nd = 200 in Eq. (17), the initial and optimal TFs with
different numbers of control points are obtained by the proposed
metric and visualized in Fig. 11. These TFs are also compared
with the best case of the 5000 random initializations. The figure
indicates that optimal TFs computed by both RDM metric and
random sampling are (almost) identical. More interestingly, the
RDM-based TF with 15 control points is almost the same as
the one obtained by random sampling with the same number
of control points (see the cyan curve in Fig. 11). The respec-
tive objective function values and maximum jerk magnitudes on
boundary curves are reported in Table 3. Again, it can be seen
that both random and RDM-based initializations lead to almost
similar optimal values, considering that using the RDM metric, Fig. 12. Velocity (top), acceleration (middle) and jerk (bottom) profiles for
the results are computed by a single optimization process and boundary curves c1 and c2 of the example shown in Fig. 9. The values are
computed by RDM-based optimal TF : u = f (t) and compared to the linear
therefore with a significantly lower computational time. Partic- TF : u = t /τ . The zoomed-in boxes show in detail the comparison of the four TFs.
ularly in this comparison, the computational time is governed When applying the reparameterization, observe a minor worsening in the middle
by the number of iterations, which depends on how far the of the domain in contrast to a major improvement close to the boundaries.
initial guesses are from the optimized values. Using the proposed
RDM metric, we find an initialization for the TF that is closer
to the final optimal one and, consequently, the computational In order to have a better interpretation of the reparameter-
time is decreased. Regarding the number of control coefficients, ization technique and see how it may change the kinematic
it should be also noted that by increasing the number of qi s, the specifications along the toolpath, Fig. 13 depicts the velocity,
TF is more flexible and a (slightly) better jerk profile may be acceleration, and jerk vectors obtained by linear and optimal
found. On the other hand, it leads to more variables and there- transformations at some sampled points on the toolpath. Under
fore a more expensive optimization. As a result, a compromise
either of parameterizations we set (linear or optimal), the milling
between computational cost and required accuracy should be
tool reaches the shown locations at the same parameter u, but
considered.
at different times t. The reason is that, for the same parameter
The velocity, acceleration, and jerk plots of the original tool-
values, reparameterization reorients the time domain and the
path with linear parameter–time relationship (i.e., without
reparameterization) and the optimized toolpath obtained by em- velocity magnitude is changed accordingly. However, referring to
ploying an optimal TF with 15 control points are shown in Fig. 12. Eqs. (7)–(9), the velocity direction remains unchanged and tan-
It can be seen that the jerk profile is substantially optimized and gent to the path (for the acceleration and jerk, both direction and
smoothed compared to the original input toolpath (see Table 3). magnitude are changed). Observe that the acceleration and jerk
We again emphasize that the geometric shape of the toolpath values, which are excessive at both ends of the original path (see
remains unchanged through the jerk minimization procedure so Fig. 12) are well decreased by the presented reparameterization
that any consideration in the path planning stage is kept. algorithm.
 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868 9

4.2. The impeller benchmark workpiece

In order to show the applicability of the presented method-

ology, in this section, we study two toolpaths from the impeller
benchmark workpiece where a higher variation of ruling di-
rections occurs. Fig. 14 shows two toolpaths on the impeller
workpiece obtained by a toolpath generation algorithm (see [1]).
Both corresponding ruled surfaces have quintic boundary curves,
where the first toolpath (Figs. 14a and b) has uniform knots
and 15 control points, while the second one (Figs. 14c and d)
is spanned over a non-uniform knot sequence with 17 control
points.
The optimal (quintic) transfer function for each toolpath is
computed by the presented reparameterization technique where
the initialization for the optimization process is obtained by the
proposed RDM metric. The respective optimal TFs and the plots
of velocity, acceleration and jerk are depicted in Figs. 15 and 16,
respectively. We assume that both paths are to be traversed in
τ = 10 s and the kinematic specifications do not violate the
machine constraints. Both results consider equally important jerk
minimization on either ends, i.e., w1 = w2 in Eq. (14).

4.3. Discussions and limitations

Various optimization weights. In all our case studies we

Fig. 13. (a) Velocity, (b) acceleration, and (c) jerk vectors at some sampled points
considered equal weights in Eq. (14) because our research is
on the ruled surface of the example shown in Fig. 9 (the view is changed
to better visualize the vectors). Red vectors correspond to (original) linear motivated by 5-axis flank milling and the jerk of the tool needs to
transformation between u and t, while the green ones correspond to the TF be minimized throughout the whole grazing curve. Since the jerk
obtained by the RDM metric. The time values of reaching each point in different of the intermediate axis points is just a linear combination of the
configurations are also reported. The acceleration and jerk vectors of the original jerks on the axis’ ends, we require the jerk to be minimized on
toolpath are scaled at both ends to be fitted in the figure.
both ends equally. One can, however, consider jerk minimization

Fig. 14. Two milling paths, (a,b) and (c,d), respectively, on a blade (gray) of an impeller (e). It is a simulation of 5-axis flank milling, data courtesy of [1], using
two distinct conical milling tools. The red positions of the tool are related the original (linear) parameterization, while the green positions show the motion after
jerk-minimized reparameterization. All positions are uniformly sampled in the time space. Similarly to Fig. 9, the reparameterized motion better captures distances
between particular positions of the tools, resulting in smoother motion (in terms of jerk). (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)
10 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868

Table 3
Maximum jerk magnitudes on boundary curves and the objective function value (F ) for the first case study. Optimal TFs are obtained by the random initializations
and the proposed RDM metric, respectively. The effect of employing different numbers of control points (# CPs) for the transfer function is shown and the advantage
of implementing the RDM metric (in terms of computational cost) is highlighted.
Linear TF Optimal TF by 5K random initializations Optimal TF by proposed RDM metric
(non-optimal) (# CPs = 15) # CPs = 10 # CPs = 15 # CPs = 20
max(||J1 (t)||) (mm/s3 ) 31.2402 2.4308 2.5456 2.4303 2.4207
max(||J2 (t)||) (mm/s3 ) 78.0993 2.6108 2.3940 2.6133 2.6037
Objective function value F (mm2 /s6 ) 853.7908 12.1626 12.6436 12.1628 12.1195
Computational time (s) – 61.7a (average time per case) 14.8b 45.5 111.4
# Iterations – 73 (average # iterations per case) 31 39 44
a
The reported time for the case of 5000 random initializations, is the average of parallel runs on 50 CPUs (100 cases per run). The total time is 6171 s.
b
The reported time for the proposed method is obtained by a single run on one CPU.

Fig. 15. Jerk minimization of the 1st toolpath of the impeller workpiece (Figs. 14a and b). (a) Optimal vs. linear TFs compared to the RDM-based initial TF. (b)
Velocity, acceleration, and jerk profiles over time obtained by linear and optimal TFs on boundary curves c1 and
 c2 .(c), (d), and (e) Contours of velocity, acceleration,
and jerk on the ruled surface obtained by linear and optimal TFs where the values of ∥Vm (t)∥, ∥Am (t)∥, and Jm (t) are the means of respective values on boundary
curves in (b). Note we changed our view to be in accordance with global xyz coordinate system.

in the context of ball-end milling and there it makes sense to freedom. We experimented with a higher number of control
minimize the jerk of the whole tool by prioritizing the weight on coefficients of the transfer function to increase the number of
the side closer to the contact point. This can be easily controlled degrees of freedom which increased the computational cost and
by changing the weights in Eq. (14). improved the jerk only negligibly (as reported in Sections 3.2 and
Considerations on the transfer functions. We initialize the 4.1). After these experiments, we set the degree to 5 and the
transfer function using a quintic B-spline to meet the mini- number of the control points to 15 in our implementation.
mum requirement of having a C 1 -continuous jerk–time profile G-code input. In the case when the toolpaths are given in
(see Remark 2). One can eventually experiment with higher de- terms of discrete data points or in terms of a G-code, our algo-
grees in situations where the initial data of the transfer function rithm can be applied as well. We just need to know the directions
(obtained by the RDM metric) would indicate more degrees of and length of the rulings that can be easily extracted from motion
 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868 11

Fig. 16. Jerk minimization of the 2nd toolpath of the impeller workpiece (Figs. 14c and d). (a) Optimal vs. linear TFs compared to the RDM-based initial TF. (b)
Velocity, acceleration, and jerk profiles over time obtained by linear and optimal TFs on boundary curves c1 and
 c2 .(c), (d), and (e) Contours of velocity, acceleration,
and jerk on the ruled surface obtained by linear and optimal TFs where the values of ∥Vm (t)∥, ∥Am (t)∥, and Jm (t) are the means of respective values on boundary
curves in (b). Note we changed our view to be in accordance with global xyz coordinate system.

path goto files or from G-codes. Our experiments were applied to 5. Conclusions
a smooth (B-spline) input, see e.g. Fig. 9, however, our follow-up
project that deals with physical validations will deal directly with We have introduced a new jerk-minimizing algorithm for
G-codes. multi-axis machining. The algorithm looks for a reparameteri-
Kinematic limits and smooth motions on machine drives. zation of the milling motion such that the jerk is globally min-
In this work, we focus on the jerk values of the tool motion that imized. In our initialization–optimization based framework, we
are derived from the geometry of the toolpath. Hence, the opti- use the ruling distance metric that measures the distance be-
mization objective is evaluated with reference to the workpiece tween two finite lines in 3D-space. We have shown that this
coordinate system (WCS). From the practical point of view, it is metric can be used for the initialization of the motion reparam-
more relevant to optimize the jerk values on machine drives, eterization as it reliably captures changes of the milling tool’s
i.e., X, Y, Z, B and C in 5-axis machining, which are referred to axis. We have validated our algorithm on several case studies
as the machine coordinate system (MCS). Note that the currently and have shown that the proposed method quickly decreases the
smoothed toolpath in WCS may result in a non-smooth motion jerk, resulting in smooth milling motions. Our follow-up research
in MCS. Such an approach, requires to consider a particular ma- aims at physical validations of the proposed methodology, and
chine (and the limits on its drives) as a part of the optimization
incorporating other physical entities such as kinematic limits
algorithm. However, optimization in MCS would require a differ-
on machine drives, tool vibration and/or cutting forces into the
ent metric since the current RDM measures distances between
optimization loop.
the rulings and does not correspond to an equal distribution of
the motion in MCS, considering each drive may have different
kinematic limits. Then, with such a metric at hand, fixing the Declaration of competing interest
workpiece setup and going through the inverse kinematic trans-
formation (IKT) to convert the optimization objective from WCS The authors declare that they have no known competing finan-
to its counterpart in MCS is an interesting topic for a follow-up cial interests or personal relationships that could have appeared
project. to influence the work reported in this paper.
12 A. Hashemian, P. Bo and M. Bartoň / Computer-Aided Design 127 (2020) 102868

Acknowledgments [15] Huang J, Lu Y, Zhu L-M. Real-time feedrate scheduling for five-axis machin-
ing by simultaneously planning linear and angular trajectories. Int J Mach
The first author has been partially funded by the BCAM ‘‘Severo Tools Manuf 2018;135:78–96. http://dx.doi.org/10.1016/j.ijmachtools.2018.
08.006.
Ochoa’’ accreditation of excellence, Spain (SEV-2017-0718), and
[16] Weng W-H, Kuo C-FJ. Jerk decision for free-form surface effects in
the Basque Government, Spain through the BERC 2018–2021 multi-axis synchronization manufacturing. Int J Adv Manuf Technol
program. The second author has been partially supported by the 2019;105:799–812. http://dx.doi.org/10.1007/s00170-019-04164-5.
National Natural Science Foundation of China (61672187) and the [17] Farouki RT, Li S. Optimal tool orientation control for 5-axis CNC milling
Shandong Provincial Key R & D Program, China (2018GGX103038). with ball-end cutters. Comput Aided Geom Design 2013;30:226–39. http:
The third author has been partially supported by Spanish Ministry //dx.doi.org/10.1016/j.cagd.2012.11.003.
[18] Sun Y, Bao Y, Kang K, Guo D. A cutter orientation modification method for
of Science, Innovation and Universities: Ramón y Cajal with
five-axis ball-end machining with kinematic constraints. Int J Adv Manuf
reference RYC-2017-22649 and the European Union’s Horizon Technol 2013;67:2863–74. http://dx.doi.org/10.1007/s00170-012-4699-6.
2020 research and innovation programme under agreement No. [19] Sharma SK, Duvedi RK, Bedi S, Mann S. A method for generating multiple
862025. solutions for multipoint five-axis tool positioning. Int J Adv Manuf Technol
2019;100:2509–20. http://dx.doi.org/10.1007/s00170-018-2871-3.
References [20] Hashemian A, Hosseini SF, Nabavi SN. Kinematically smoothing trajecto-
ries by NURBS reparameterization – An innovative approach. Adv Robot
2017;31:1296–312. http://dx.doi.org/10.1080/01691864.2017.1396923.
[1] Calleja A, Bo P, González H, Bartoň M, López de Lacalle LN. Highly accurate
[21] Bo P, Bartoň M, Pottmann H. Automatic fitting of conical envelopes to free-
5-axis flank CNC machining with conical tools. Int J Adv Manuf Technol
form surfaces for flank CNC machining. Comput Aided Des 2017;91:84–94.
2018;97:1605–15. http://dx.doi.org/10.1007/s00170-018-2033-7.
[2] Sencer B, Altintas Y, Croft E. Feed optimization for five-axis CNC machine http://dx.doi.org/10.1016/j.cad.2017.06.006.
tools with drive constraints. Int J Mach Tools Manuf 2008;48:733–45. [22] Piegl L, Tiller W. The NURBS Book. 2nd ed. New York, NY: Springer-Verlag;
http://dx.doi.org/10.1016/j.ijmachtools.2008.01.002. 1997, p. 646.
[3] Constantinescu D, Croft EA. Smooth and time-optimal trajectory [23] Bo P, Bartoň M, Plakhotnik D, Pottmann H. Towards efficient 5-axis flank
planning for industrial manipulators along specified paths. J Robot CNC machining of free-form surfaces via fitting envelopes of surfaces of
Syst 2000;17:233–49. http://dx.doi.org/10.1002/(SICI)1097-4563(200005) revolution. Comput Aided Des 2016;79:1–11. http://dx.doi.org/10.1016/j.
17:5<233::AID-ROB1>3.0.CO;2-Y. cad.2016.04.004.
[4] Zhang Y, Zhao M, Ye P, Zhang H. A G4 continuous B-spline transition [24] Bo P, Bartoň M. On initialization of milling paths for 5-axis flank CNC
algorithm for CNC machining with jerk-smooth feedrate scheduling along machining of free-form surfaces with general milling tools. Comput Aided
linear segments. Comput Aided Des 2019;115:231–43. http://dx.doi.org/10. Geom Design 2019;71:30–42. http://dx.doi.org/10.1016/j.cagd.2019.04.012.
1016/j.cad.2019.04.004. [25] Biagiotti L, Melchiorri C. Trajectory planning for automatic machines and
[5] Bosetti P, Bertolazzi E. Feed-rate and trajectory optimization for CNC robots. Berlin, Heidelberg: Springer-Verlag; 2008.
machine tools. Robot Comput-Integr Manuf 2014;30:667–77. http://dx.doi. [26] Hosseini SF, Hashemian A, Reali A. On the application of curve reparame-
org/10.1016/j.rcim.2014.03.009. terization in isogeometric vibration analysis of free-from curved beams.
[6] Jahanpour J, Alizadeh MR. A novel acc-jerk-limited NURBS interpolation Comput Struct 2018;209:117–29. http://dx.doi.org/10.1016/j.compstruc.
enhanced with an optimized S-shaped quintic feedrate scheduling scheme. 2018.08.009.
Int J Adv Manuf Technol 2015;77:1889–905. http://dx.doi.org/10.1007/ [27] Gautschi W. On generating Gaussian quadrature rules. In: Hämmer-
s00170-014-6575-z. lin G, editor. Numerische integration: Tagung im mathematischen
[7] Zhang Q, Li S, Guo J. Smooth time-optimal tool trajectory generation for
forschungsinstitut oberwolfach. Basel: Birkhäuser Basel; 1979, p. 147–54.
CNC manufacturing systems. J Manuf Syst 2012;31:280–7. http://dx.doi.
[28] Bartoň M, Calo VM. Gauss–Galerkin quadrature rules for quadratic and
org/10.1016/j.jmsy.2012.06.001.
cubic spline spaces and their application to isogeometric analysis. Comput
[8] Beudaert X, Pechard P-Y, Tournier C. 5-axis tool path smoothing based on
Aided Des 2017;82:57–67. http://dx.doi.org/10.1016/j.cad.2016.07.003.
drive constraints. Int J Mach Tools Manuf 2011;51:958–65. http://dx.doi.
[29] Bartoň M, Calo VM. Optimal quadrature rules for odd-degree spline
org/10.1016/j.ijmachtools.2011.08.014.
[9] Beudaert X, Lavernhe S, Tournier C. Feedrate interpolation with axis jerk spaces and their application to tensor-product-based isogeometric analysis.
constraints on 5-axis NURBS and G1 tool path. Int J Mach Tools Manuf Comput Methods Appl Mech Engrg 2016;305:217–40. http://dx.doi.org/10.
2012;57:73–82. http://dx.doi.org/10.1016/j.ijmachtools.2012.02.005. 1016/j.cma.2016.02.034.
[10] Sun Y, Zhao Y, Xu J, Guo D. The feedrate scheduling of parametric [30] Fußeder D, Simeon B. Algorithmic aspects of isogeometric shape optimiza-
interpolator with geometry, process and drive constraints for multi-axis tion. In: Jüttler B, Simeon B, editors. Isogeometric analysis and applications
CNC machine tools. Int J Mach Tools Manuf 2014;85:49–57. http://dx.doi. 2014. Cham: Springer International Publishing; 2015, p. 183–207.
org/10.1016/j.ijmachtools.2014.05.001. [31] Hosseini SF, Moetakef-Imani B, Hadidi-moud S, Hassani B. Pre-bent
[11] Hu P, Tang K. Improving the dynamics of five-axis machining through shape design of full free-form curved beams using isogeometric
optimization of workpiece setup and tool orientations. Comput Aided Des method and semi-analytical sensitivity analysis. Struct Multidiscip Optim
2011;43:1693–706. http://dx.doi.org/10.1016/j.cad.2011.09.005. 2018;58:2621–33. http://dx.doi.org/10.1007/s00158-018-2041-0.
[12] Essid M, Gassara B, Baili M, Hbaieb M, Dessein G, Bouzid Saï W. Analytical [32] Bonnans J-F, Gilbert J, Lemarechal C, Sagastizábal C. Numerical opti-
modeling of the CNC machine axis motion in high-speed milling with local mization, theoretical and practical aspects. 2nd ed. Berlin, Heidelberg:
smoothing. Int J Adv Manuf Technol 2019;105:457–70. http://dx.doi.org/10. Springer-Verlag; 2006.
1007/s00170-019-04157-4. [33] Nocedal J, Wright SJ. Numerical optimization. New York, NY: Springer-
[13] Song D-N, Ma J-W. Interval partition-based feedrate scheduling with axial Verlag; 2006.
drive constraints for five-axis spline toolpaths. Int J Adv Manuf Technol [34] Chen H-Y, Pottmann H. Approximation by ruled surfaces. J Comput Appl
2019;105:4701–14. http://dx.doi.org/10.1007/s00170-019-04433-3. Math 1999;102:143–56. http://dx.doi.org/10.1016/S0377-0427(98)00212-
[14] Ma J-W, Jia Z-Y, Qin F-Z, Song D-N, Jiang W-W, Chen S-Y. A five-axis X.
dual NURBS interpolator with constant speed at feedrate-sensitive regions
[35] Pottmann H, Wallner J. Computational line geometry. Berlin, Heidelberg:
under axial drive constraints. J Manuf Sci Eng 2019;141. http://dx.doi.org/
Springer-Verlag; 2001.
10.1115/1.4043256.