See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228953125

Time-optimal interpolation of CNC machines along parametric path with

chord error and tangential acceleration bounds

Article · January 2010

CITATIONS READS

5 211

2 authors:

ChunMing Yuan Xiao-Shan Gao

Chinese Academy of Sciences Chinese Academy of Sciences
66 PUBLICATIONS 832 CITATIONS 220 PUBLICATIONS 4,476 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Xiao-Shan Gao on 27 May 2014.

The user has requested enhancement of the downloaded file.

Mathematics Mechanization Research Preprints
KLMM, Chinese Academy of Sciences
Vol. 29, 165–188, September 2010 165

Time-Optimal Interpolation of CNC Machines

along Parametric Path with Chord Error and
Tangential Acceleration Bounds

Chun-Ming Yuan, Xiao-Shan Gao

KLMM, Institute of Systems Science, Chinese Academy of Sciences
E-mail: (cmyuan,xgao)@mmrc.iss.ac.cn.

Abstract. Interpolation and velocity planning for parametric curves are crucial prob-
lems in CNC machining. In this paper, a time-optimal velocity planning algorithm under
a given chord error bound and a tangential acceleration bound is proposed. The key idea
is to reduce the chord error bound to a centripetal acceleration bound which leads to
a velocity limit curve, called the chord error velocity limit curve (CEVLC). Then, the
velocity planning is to find the time-optimal velocity curve governed by the tangential
acceleration bound “under” the CEVLC. For two types of simple splines, explicit for-
mulas for the optimal velocity curve are given. We implement the methods in these two
cases and use curved segments from real machining parts to show the feasibility of the
methods.

Keywords. Parametric curve interpolation, time-optimal velocity planning, chord error,

velocity limit curve.

1 introduction
In modern CAD systems, the standard representation for free form surfaces is parametric
functions. While the conventional computed numerically controlled (CNC) systems mainly
use micro line segments (G01 codes) to represent the machining path. To convert the para-
metric curves into line segments is not only time consuming but also leads to problems such
as large data storage, speed fluctuation, and poor machining accuracy. In the seminal work
[5, 18, 22], Chou, Yang, and Shpitalni et al proposed to use parametric curves generated in
CAD/CAM systems directly in the CNC systems to overcome these drawbacks.
An interpolation algorithm usually consists of two phases: velocity planning and parame-
ter computation. Let C(u) be the manufacturing path. The phase to determine the feed-rate
v(u) along C(u) is called velocity planning. When the feed-rate v(u) is known, the way to
compute the next interpolation point at ui+1 = ui + 4u during one sampling time period is
called parameter computation.
This paper focuses on velocity planning along a spatial parametric path. For the param-
eter computation phase, please refer to [4, 9, 21] and the literatures in them. Two types of
acceleration modes can be used in the velocity planning: the tangential acceleration and the
multi-axis acceleration where each axis accelerates independently.
Due to its conceptual simplicity, the tangential acceleration is widely used in CNC in-
terpolations. Bedi et al[1] and Yang-Kong[22] used a uniform parametric feed-rate without
166 C.M. Yuan and X.S. Gao

considering the chord error. Yeh and Hsu used a chord error bound to control the feed-rate if
needed and used a constant feedarte in other places [23]. However, the machine acceleration
capabilities were not considered. Narayanaswami and Yong [24] gave a velocity planning
method based on tangential acceleration bounds by computing maximal passing velocities
at sensitive corners. Cheng and Tsai [4] gave an interpolation method based on different
velocity profiles for tangential ACC/DCC feed-rate planning. Velocity planning methods
with tangential acceleration and jerk bounds were considered by several research groups
[7, 11, 13, 14, 12]. In particular, Emami and Arezoo [6], Lai et al[10] proposed velocity
planning methods with confined acceleration, jerk, and chord error constraints by checking
these values at every sampling point and using backtracking to adjust the velocity if any
of the bounds is violated. In all the above methods besides [6, 10, 23], the chord error is
considered only at some special points, and there is no guarantee that the chord error is
satisfied at all points. Furthermore, all these methods are not proved to be time-optimal.
In the multi-axis acceleration case, Borow [3] and Shiller et al [16, 15] presented a time-
optimal velocity planning method for a robot moving along a curved path with acceleration
bounds for each axis. Farouki and Timar [19, 20] proposed a time-optimal velocity planning
algorithm in CNC machining under the same acceleration constraints. Following the method
in [19, 20], Zhang et at gave a simplified time-optimal velocity planning method for quadratic
B-splines and realized real-time manufacturing on industrial CNC machines [26]. Zhang et
al[25] gave a greedy algorithm for velocity planning under multi-axis jerk constraints. These
optimal methods use the “Bang-Bang” control strategy, that is, at least one of the axes
reaches its acceleration bound all the time. But they do not consider the chord error which
is an important factor for CNC-machining.
In this paper, we give a time-optimal velocity planning algorithm along a parametric
curve path C(u) under a given chord error bound and a tangential acceleration bound. The
main contribution of the paper is the introduction of the chord error velocity limit curve
(abbr. CEVLC) and the velocity planning algorithm.
We show that the chord error bound can be approximately reduced to a centripetal
acceleration bound. Furthermore, if the centripetal acceleration reaches its maximal bound,
the velocity can be written as an algebraic function in the parameter u of the curve path
C(u). The graph of this function is called the CEVLC. The CEVLC is significant because the
final velocity curve must be “under” this curve or be part of this curve, which narrows the
range of velocity planning. Also, certain key points on the CEVLC, such as the discontinuous
points, play an important role in the velocity planning.
Similar to the previous work on time-optimal velocity planning, our algorithm also uses
the “Bang-Bang” control strategy, which is a necessary condition for time optimal in our
case. Then the final velocity curve is governed either by the centripetal acceleration bound
or by the tangential acceleration bound, and the later one is called the integration velocity
curve. The main task of the velocity planning is to find the switching points between the
integration velocity curve and the CELVC.
Two main results on the optimal velocity curve are given in this paper. Firstly, as a
theoretical result, we show that the final velocity is the minimum of the CEVLC and all
the integration velocity curves passing through the initial point, the termination point, and
the key points of the CEVLC. We also give a practical algorithm to compute the optimal
 Time-Optimal Control of CNC Machines with Chord Error 167

velocity curve incrementally. For quadratic splines and cubic PH curves, the integration
velocity curve can be given by explicit formulas. We implement our algorithm in these two
cases and conduct experiments using curved pathes from real manufacturing parts.
Moreover, we also give a simple and real-time feed-rate override algorithm based on our
velocity planning curve. For multi-axis velocity planning algorithms, feed-rate override is
complicated and not suitable for real time interpolation.
As a final remark, we want to mention that by combining the CEVLC proposed in this
paper and the method proposed in [19, 20], it is possible to give a time optimal velocity
planning method under a given chord error bound and multi-axis acceleration bounds.
The rest of the paper will be organized as follows. Section 2 gives the time-optimal
velocity planning algorithm for a parametric curve. Section 3 gives the details for computing
the time optimal velocity curves for quadratic B-spline and cubic PH-spline. Section 4 gives
a feed-rate override algorithm for CNC-machining. Section 5 concludes the paper.

2 Time-optimal velocity planning under chord error and tan-

gential acceleration bound
The time-optimal velocity planning algorithm will be presented in this section. We consider a
spatial piecewise parametric curve C(u), u = 0..1 with C 1 continuity, such as splines, Nurbs,
etc. We further assume that each piece of the curve is differentiable to the third order and
has left and right limitations at the endpoints.

2.1 Problem
In order to control the machine tools, we need to know the velocity of the movement at each
point on the curve, which is denoted as a function v(u) in the parameter u and is called the
velocity curve. The procedure to compute the velocity curve v(u) is called velocity planning.
In this subsection, we will show that the chord error bound can be reduced to the centripetal
acceleration bound and formulate the velocity planning problem as an optimization problem
with tangential and centripetal acceleration bounds.
For a parametric curve C(u), we denote its parametric speed to be:

ds
σ(u) = = |C 0 (u)|,
du
where 0 is the derivative w.r.t. u. The curvature and radius of curvature are defined to be:

|C 0 (u) × C 00 (u)| 1
k(u) = 3
, ρ(u) = .
σ(u) k(u)

|PQ|: chord error

Fig. 1. The chord error

168 C.M. Yuan and X.S. Gao

The cutter moves in a line segment between two adjacent points on the curve C(u) and
the distance between the line segment and the curve segment is called the chord error which
is the main source of the manufacturing error (Fig 1). Firstly, we will show that the chord
error bound can be reduced to the centripetal acceleration bound.
Let T be the sampling period, δ the error constraint, and ±A the tangential acceleration
bounds. In practice, the chord error bound is much less than the radius of curvature at each
point on the curve. By the chord error formula [23], at each parametric value u,

8δρ(u) − 4δ 2 8δρ(u) 8δ
q(u) = v 2 (u) ≤ 2
≈ 2
= .
T T k(u)T 2

If we denote by d(u) the interpolating chord error at u with feed-rate v(u), then d(u) =
T 2 v(u)2
8ρ(u) ≤ δ. Let
aN (u) = v(u)2 /ρ(u) = k(u)v(u)2 = k(u)q(u) (1)
be the centripetal acceleration and
8δ
B=. (2)
T2
Then, the chord error bound is transformed to the centripetal acceleration bound:

d(u) ≤ δ ⇐⇒ aN (u) ≤ B.

Since
d ds du d v d
= = , (3)
dt dt ds du σ du
The tangential acceleration is:

dv(u) v(u)v 0 (u) q 0 (u)

aT (u) = = = . (4)
dt σ(u) 2σ(u)

From (3), the time optimal velocity planning problem becomes the following optimization
problem: to find a velocity curve v(u) such that
Z 1
σ(u)
min t = du, (5)
v(u) 0 v(u)

with the tangential and centripetal acceleration constraints

|aT (u)| ≤ A, aN (u) ≤ B, u = 0..1, (6)

where (aT (u), aN (u)) are the tangential and centripetal accelerations and (A, B) are their
bounds respectively. Note that B can be derived from the chord error bound from (2).
 Time-Optimal Control of CNC Machines with Chord Error 169

2.2 The main result

In this section, we present the main result of the paper about the time optimal velocity
planning problem (5). We will give a velocity planning algorithm in section 2.4 and prove
the correctness of the result in section 2.5.
Before giving the main result, we need to introduce several basic concepts. We use
“Bang-Bang” control, that is, at least one of “=” holds among the inequalities in (6). This
will lead to two cases.
B
If the centripetal acceleration reaches its bound B, then from (1) we have q(u) = k(u) ,
which defines a curve
2
qlim (u) = vlim (u) = B/k(u), u ∈ [0, 1] or (7)
p
vlim (u) = B/k(u), u ∈ [0, 1],
in the u-v plane with v as the vertical axis. We call this curve the velocity limit curve with the
chord error constraint, denoted by CEVLC. If C(u) is a piecewise curve with C 1 continuity,
then its CEVLC is the combination of the CEVLCs of its components. From the constraint
aN ≤ B, the velocity curve must be “under” the CEVLC in the v direction.
2u 1−u 2
Example 2.1 If C(u) = ( 1+u 2 , 1+u2 , 0) and B = 100, then vlim (u) = 10, u ∈ [0, 1].
3
If C(u) = (u2 +u+1, u2 +2u−1, 2u+1) and B = 600, then vlim (u) = 10(8u2 +12u+9) 4 , u ∈
[0, 1].
If C(u) = ( 13 u3 − 4u2 − 8u, 23 u3 + 3u2 − 8u, − 32 u3 − 2u2 − 14u) and B = 1000, then
vlim (u) = 30u2 + 40u + 180, u ∈ [0, 1].
If the tangential acceleration reaches its bound, then from (4) the square of the velocity
q(u) = v(u)2 can be obtained by solving the differential equation q 0 (u) = ±2Aσ(u), whose
solution is
q = ±2Aπ(u) + c, (8)
R
where π(u) = σdu is the primitive function of σ(u) and c a constant which can be deter-
mined by an initial point (u∗ , q(u∗ )):
c = q(u∗ ) ∓ 2Aπ(u∗ ).
The curve (8) is called the integration curve or the integration trajectory.
Let vP (u) be the integration trajectory with tangential acceleration A starting from an
initial point P = (uP , vlim (uP )) both for the forward (the +u) and the backward (the
−u) directions. Let v0 (u) be the forward integration trajectory with tangential acceleration
A from the initial point P0 = (0, 0) and v1 (u) the backward integration trajectory with
tangential acceleration A from the initial point P1 = (1, 0). Of curse, all the curves mentioned
above are defined in [0, 1].
Then, the solution to the optimization problem (5) is given below.

Theorem 2.2 Let K be the finite set of key points of the CEVLC to be defined in the next
section. Then
v(u) = min(vlim (u), vP (u), v0 (u), v1 (u)). (9)
P ∈K
Then, v(u) is the solution to the optimal problem (5).
170 C.M. Yuan and X.S. Gao

Obviously, v(u) is a piecewise continuous curve for u ∈ [0, 1]. We will prove the theorem in
Section 2.5.

2.3 Key points of the CEVLC

There exist three types of special points on the CEVLC, which are called switching points
or key points.
The first type switching points are the discontinuous points of the CEVLC. These points
correspond to the curvature discontinuous points of the original curve. They must be the
singular points of the parametric curve or the connection points of two curve segments. If
we denote by C(u) the curve segment, then the singular points of C(u) can be computed
by solving the equations C 0 (u) × C 00 (u) = 0. From (7), the limit velocity vlim is ∞ at these
points and the real velocity curve can not reach it, so we can omit the singular points from
the key points. Thus, only the connection points need to be considered.
At a first type key point, the velocity is not continuous and we call the smaller one of
the left and right side velocities as the minimal velocity or simply the velocity of this point.
The second type switching points are the continuous but non-differentiable points of the
CEVLC (slope discontinuity). These points correspond to the curvature continuous but non-
differentiable points of the original curve. Hence, they must be the connection points of two
curve segments.
For a differentiable segment of the CEVLC divided by the above two types of key points,
we can divide it according to whether the tangential acceleration of the CEVLC is ±A, where
A is from (6). A point on the CEVLC is called a key point of the third type if the tangential
acceleration along the CEVLC at this point is ±A. We can find the third type key points
by solving the following algebraic equation in u

qlim (u) Bσ(u)2

aT (u) = = = ±A.
2σ(u) 2|C 0 (u) × C 00 (u)|

With these switching points, the CEVLC is divided into two types of segments:

1. A curve segment is called feasible if the absolution values of tangential acceleration at

all points are bounded by A. A feasible CEVLC segment can be a part of the final
velocity curve.

2. A curve segment is called unfeasible if the absolution values of tangential acceleration

at all points are larger than A. An unfeasible CEVLC segment cannot be a part of the
final velocity curve. In other words, the final velocity curve must be strictly under it.

If the curve segments on the left and right sides of a second type key point are both
feasible, we delete this key point.

2.4 The time optimal velocity planning algorithm

We use the “Bang-Bang” control strategy. Then the real velocity curve is governed either
by the maximal centripetal acceleration or by the maximal tangential acceleration. Since
the velocity curve must be under or be a part of the CEVLC, we need to find the proper
 Time-Optimal Control of CNC Machines with Chord Error 171

integration trajectory under the CEVLC. We first give the main idea of the velocity planning
algorithm.
Firstly, we compute the CEVLC, find the key points and their speeds. Compute the
forward integration trajectory vs from the starting point (0, 0) with tangential acceleration
A. Find the intersection point (ul , vs (ul )) of vs and the CEVLC. Compute the backward
integration trajectory ve from the ending point (1, 0) with tangential acceleration A. Find
the intersection point (ur , ve (ur )) of ve and the CEVLC.
Secondly, if (ul , vs (ul )) = (ur , ve (ur )), then return the combination of vs and ve as the
final velocity curve. If ul > ur , find the intersection point of vs and ve and return the
combination of vs and ve as the final velocity curve. Otherwise, set Pc = (ul , vs (ul )) to be
the current point and consider the following three cases.
If the next segment of CEVLC from point Pc in the forward direction (the +u direction)
is feasible, then we merge this feasible segment into vs and set the current point (ul , vs (ul ))
to be the end point of this segment.
If the next segment of CEVLC from point Pc in the forward direction is not feasible and
the forward integration trajectory vf with initial point Pc is under the CEVLC, then let
(ui , vf (ui )) be the intersection point of vf and the CEVLC. Merge vf (u), u ∈ [ur , ui ] into vs .
Let ul = ui and Pc = (ul , vf (ul )).
If the next segment of CEVLC from point Pc in the forward direction is not feasible and
the integration trajectory vf with initial point Pc is above the CEVLC, then we find the
next key point Pn on the right hand side of Pc . From Pn , compute the backward integration
trajectory vb , find the intersection point of vb with vs , and merge vb and vs as the new vs .
Set Pn as the new current point.
Finally, with the new current point, we can repeat the procedure from the second step
until the velocity curve is found.
Here by saying a forward curve v1 (u) is under or above another curve v2 (u) from a point
P = v1 (u∗), we mean that v1 (u) is under or above v2 (u) in a neighborhood of u∗ in the
forward direction.
To describe our algorithm precisely, we need the following notations. For a discontinuous
curve f1 (x), we denote by f1+ (x∗ ) and f1− (x∗ ) the limitations of f1 (x) at x∗ from the left
and right hand sides respectively, and define f1 (x∗ ) = min(f1+ (x∗ ), f1− (x∗ )). Let f2 (x) be
a curve with C 0 continuity. If f2 (x∗ ) = f1 (x∗ ) or f2 (x∗ ) is between the left and right
limitations of f1 (x) at x∗ , then define (x∗ , f2 (x∗ )) to be the intersect point of the curves
(x, f2 (x)) and (x, f1 (x)).
We now give the velocity planning algorithm.
Algorithm VP CETA. The input of the algorithm is the curve C(u), u ∈ [0, 1], a chord
error bound d, and a tangential acceleration bound A. The output is the velocity curve
vc (u), u ∈ [0, 1] which is the solution to the optimization problem (5).

1 Compute the CEVLC in (7) and its key points as shown in section 2.3. Denote by Uf the
parametric intervals where the corresponding segments of the CEVLC are feasible.

2 From the starting point (0, 0), compute the forward integration trajectory vs using the
tangential acceleration A. Compute the first intersection point (ul , vs (ul )) of vs and
the CEVLC. If there exists no intersection, denote ul = 1.
172 C.M. Yuan and X.S. Gao

3 From the ending point (1, 0), compute the backward integration trajectory ve using the
tangential acceleration A. Compute the first intersection point (ur , ve (ur )) of ve and
the CEVLC. If there exists no intersection, denote ur = 0.

4 If (ul , vs (ul )) = (ur , ve (ur )), then return the combination of vs and ve as the final velocity
curve. If ul > ur , find the intersection point (ui , vs (ui )) of vs and ve , return v(u),
where (
vs , 0 ≤ u ≤ ui
v(u) = (10)
ve , ui < u ≤ 1.

5 If (ul , vlim (ul )) is a discontinuous point on the CEVLC, there are three possibilities.
+ −
(a1) If vlim (ul ) > vs (ul ) = vlim (ul ), goto step 6.
+ −
(a2) If vlim (ul ) = vs (ul ) < vlim (ul ), goto step 8.
+ −
(a3) If vlim (ul ) ≥ vs (ul ) > vlim (ul ), let un = ul and goto step 9.

6 Now, (ul , vs (ul )) is the starting point of the next segment of CEVLC. There exist three
cases.
(b1) If the next segment of the CEVLC is feasible, goto step 7
(b2) If a−
T (ul ) ≥ A, goto step 8.
(b3) If a−
T (ul ) ≤ −A, then find the next key point (un , vlim (un )) along the +u direction
and goto step 9.

7 Let un > ul be the parameter of the next key point. Then, (ul , un ) ⊂ Uf . Update vs to
be (
vs (u), 0 ≤ u < ul
vs (u) =
vlim (u), ul ≤ u ≤ un .
Let ul = un , goto step 4.

vlim vlim

vlim
vf
vf
vlim u=ul u=u l
(a) Case (a2) (b) Case (b2)
Fig. 2. Two cases of step 8: computation of forward integration curve

8 Starting from (ul , vs (ul )), compute the forward integration trajectory vf with tangential
acceleration A. Find the first intersection point (ui , vf (ui )) of vf and the CEVLC (Fig.
2). If there exists no intersection, denote ul = 1. Update vs to be
(
vs (u), 0 ≤ u < ul
vs (u) =
vf (u), ul ≤ u ≤ ui .
 Time-Optimal Control of CNC Machines with Chord Error 173

vlim

vlim
vs
vb vs
vlim vb
u=un u=un
(a) Case (a3) (b) Case (b3)
Fig. 3. Two cases of step 9: computation of backward integration curve

Let ul = ui , goto step 4.

9 Starting from point (un , vlim (un )), compute the backward integration trajectory vb in the
−u direction with tangential acceleration A. Find the intersection point1) (ui , vb (ui ))
of vb and vs . See Fig. 3 for an illustration. Update vs to be
(
vs (u), 0 ≤ u < ui
vs (u) =
vb (u), ui ≤ u ≤ un .

Let ul = un , goto step 4.

The correctness proof of the algorithm is given in Section 2.5. Further improvements of
the algorithm are given in Section 2.6.
The flow chart of the above algorithm is given in Fig. 5. The details of the algorithm is
omitted in the figure.

110

100
100

90 80

II

80 VI
60
IV

70
V
40 III VII

60

20 I
50

40 0 0.2 0.4 0.6 0.8 1

0 0.2 0.4 0.6 0.8 1
x

(a) CEVLC and key points (b) velocity curve

Fig. 4. An illustrative example of velocity planning

Figure 4 is an illustrative example of the algorithm.

Compute the CEVLC vlim and the key points as shown in Figure 4(a), where ◦ represents
the key points of first type and the corresponding parameters are 0.2, 0.7; ¦ represents the
key point of second type, and the corresponding parameter is 0.4; and ¤ represents the key
points of third type. The gray parts are feasible segments.
1)
We will show that there exists a unique intersection point.
174 C.M. Yuan and X.S. Gao

Starting from point (u, v) = (0, 0), compute the forward integration trajectory vs which
intersects the CEVLC at u = 0.2. Starting from (u, v) = (1, 0), compute the backward
integration trajectory ve which intersects the CEVLC at u = 0.7.
In step 5, the current point Pc is a discontinues point and case (a3) is executed. In
step 9, starting from the first ◦ point, compute the backward integration trajectory vb which
intersects vs at the first point marked by +. Update vs to be the piecewise curve marked by
I, II in Figure 4(b).
From the first point marked by ◦, the CEVLC is feasible. Hence, update vs to be the
piecewise curve marked by I, II and the first gray part of the CEVLC(III).
Let the key point marked by ¦ be the current point. From the current point, the CEVLC
is not feasible and case (b3) is executed. In step 9, we select the next key point which is
the first point marked by ¤. Starting from this point, compute the backward integration
trajectory vb which intersects vs at the second point marked by +. Update vs to be the
piecewise curve marked by I, II, III, IV in the figure.
Starting from the first point marked by ¤, the CEVLC is feasible. Hence, update vs to
be the piecewise curve marked by I, II, III, IV, V in the figure.
Let the second point marked by ¤ to be the current point. Starting from this point, the
CEVLC is not feasible and case (b2) is executed. In step 8, compute the forward integration
trajectory vf which intersects ve at the third point marked by +. The final velocity curve
consists of seven pieces marked by I, II, III, IV, V, VI, and VII in the figure.
Note that the CEVLC can be parts of the final velocity curve quite often, while in [19, 20],
the VLC is rarely a part of the final velocity curve.

Remark 2.3 In the above algorithm, we need to compute the CEVLC and its key points, the
integration trajectory, the intersection points of the integration trajectory and the CEVLC,
and the intersection points of two integration trajectories. In principle, these computations
can be reduced to computing integrations and solving algebraic equations. In Section 3, we
will show how to give explicit formulas for the integration curve for two types of special types
of curves.

2.5 The correctness of the algorithm

In this section, we will show that the velocity curve given by the algorithm in the preceding
subsection is the only solution to the optimization problem (5). The proof is divided into
two parts. We will first show that the velocity curve computed by the algorithm is the curve
defined in (9) and then prove that this velocity curve is the solution to the problem (5). In
order to prove this, we need the following result.

Theorem 2.4 [2][p.24-26] Let y, z be solutions of the following differential equations

y 0 = F (x, y), z 0 = G(x, z),

respectively, where F (x, y) ≤ G(x, y), a ≤ x ≤ b, and F or G satisfies Lipschitz’s condition.

If y(a) = z(a), then y(x) ≤ z(x) for any x ∈ [a, b].
In our case, the above theorem implies the following result.
Time-Optimal Control of CNC Machines with Chord Error 175

Fig. 5. Flow chart of the velocity planning

176 C.M. Yuan and X.S. Gao

Lemma 2.5 Let v1 (u) and v2 (u) be two velocity curves for the curve C(u) and [u1 , u2 ] ⊂
[0, 1]. If v1 (u1 ) ≤ v2 (u1 ) and aT 1 (u) ≤ aT 2 (u) for u ∈ [u1 , u2 ], then v1 (u) ≤ v2 (u) for
u ∈ [u1 , u2 ]. Furthermore, if v1 (u1 ) < v2 (u1 ), then v1 (u) < v2 (u) for u ∈ [u1 , u2 ].
Proof. We may assume that v1 (u1 ) = v2 (u2 ), since if v1 (u1 ) < v2 (u1 ), we may consider
v̄2 (u) = v2 (u) − v2 (u1 ) + v1 (u1 ) which still satisfies the conditions in the lemma. Since C(u)
is differentiable to the order of three, σ(u) and aT 1 (u) must be bounded in [u1 , u2 ]. From
(4), q10 (u) = 2σ(u)aT 2 (u) and q20 (u) = 2σ(u)aT 2 (u). Then, we have q1 (u1 ) = v1 (u1 )2 =
q2 (u1 ) = v2 (u1 )2 , 2σ(u)aT 2 (u) ≤ 2σ(u)aT 2 (u) for u ∈ [u1 , u2 ], and 2σ(u)aT 1 (u) satisfies the
Lipschitz’s condition. Using Theorem 2.4, we have v1 (u) ≤ v2 (u) for u ∈ [u1 , u2 ]. The second
part of the lemma can be proved similarly.
Before giving the correctness proof, we first explain the key steps of the algorithm.
Besides the in initial steps, new velocity trajectories are generated in Steps 7, 8, 9. We
will show that these steps are correct in the sense that they will really generate new velocity
trajectories. Step 7 is obvious, since we will use the next feasible CEVLC segment as the
velocity trajectory. Two cases lead to step 8: cases (a2) and (b2). In case (a2), we have
+ −
vlim (ul ) = vs (ul ) < vlim (ul ), which means there exists a parameter un > ul such that
vlim (u) > vs (ul ) for u ∈ (ul , un ). As a consequence, starting at (ul , vs (ul )), we can use
the tangential acceleration A to generate a piece of integration trajectory without violate
the CEVLC (See Fig. 2(a)). In case (b2), there exists a parameter un > ul such that the
tangential acceleration of the CEVLC must be strictly larger than A for u ∈ (ul , un ). By
Lemma 2.5, starting from point (ul , vs (ul )), we can use the tangential acceleration A without
violate the CEVLC (See Fig. 2(b)). We thus prove the correctness of step 8. The correctness
step 9 can be proved in a similar way. Furthermore, we have

Lemma 2.6 In step 9 of the algorithm, the backward integration trajectory vb intersects vs
only once if there exist no overlap curve segments.
Proof. Two situations lead to the execution of step 9: step (a3) and step (b3). In case
+ −
(a3), the key point at ul is of the first type and vlim (ul ) ≥ vs (ul ) > vlim (ul ). Then vs
and vb must intersect because vb decreasing with the acceleration −A and vs is bounded
by ±A. By Lemma 2.5, vs and vc can intersects only once if there exist no overlap curve
segments. Moreover, if there exists an overlap curve segment of vb and vs , then there exists
no intersection point of vb and vs except the overlap curve segment by Lemma 2.5. Case
(b3) can be proved similarly.
Similarly, we can show that vs and ve in step 4 of the algorithm only intersect once if
there exist no overlap curve segments. And we can use the bisection method to compute the
intersection point.
We now prove the main result of the paper.

Theorem 2.7 The velocity curve computed with Algorithm VP CTEA is the velocity curve
defined in equation (9) and is the only solution to the optimization problem (5).
Proof. Let v(u) be the velocity curve computed with the algorithm. It is clear that v(u)
is below the CEVLC and the tangential acceleration aT (u) of v(u) satisfies |aT (u)| ≤ A.
 Time-Optimal Control of CNC Machines with Chord Error 177

Also, if |aT (u)| 6= A, the corresponding v(u) must be a segment of feasible CEVLC. As a
consequence, v(u) satisfies conditions (6) and is bang-bang.
From the algorithm, it is clear that v(u) consists of pieces of vP (u) for all key points P
of the CEVLC including the start point (0, 0) and the ending point (1, 0). To prove (9), it
suffices to show that for each key point P , if vP (u∗ ) 6= v(u∗ ) for a parametric value u∗ , then
vP (u∗ ) > v(u∗ ). From the algorithm, it is clear that all key points including the starting
and ending points are on or above v(u). Let P = (u0 , v0 ) be a key point. Then v0 ≥ v(u0 ).
In [u0 , 1], the tangential acceleration of vP (u) is A. Then by lemma 2.5, vP (u) ≥ v(u) for
u ∈ [u0 , 1]. In [0, u0 ], if we consider the movement from u0 to 0, then the acceleration of
vP (u) is also A, and hence vP (u) ≥ v(u) for u ∈ [0, u0 ]. As a consequence, vP (u) cannot be
strictly under v(u) at any u. We thus prove that v(s) is the curve in (9).
We now prove that v(s) is an optimal solution. We will prove a stronger result, that is,
the velocity curve q(u) = v 2 (u) obtained by the algorithm is the maximally possible velocity
at each parametric value u. To prove this, we assume that there exists another velocity
curve v∗ (u) satisfying the constraints listed in (6), and there exists a u∗ ∈ [0, 1] such that
v∗ (u∗ ) > v(u∗ ). Let q∗ (u) = v∗2 (u).
The parametric interval [0, 1] is divided into sub-intervals by the key points of CEVLC
on the final velocity curve and intersection points in steps 4, 8, 9 of the algorithm. From the
algorithm, we can see that on each of these intervals, v(u) could be a segment of the CEVLC,
an integration trajectory with tangential acceleration A in the +u direction, which is called
an increasing interval, or an integration trajectory with tangential acceleration −A in the
+u direction, which is called a decreasing interval. Furthermore, if [u1 , u2 ] is an increasing
interval, the starting point (u1 , v(u1 )) must be a key point of the CEVLC; if [u1 , u2 ] is a
decreasing interval, the end point (u2 , v(u2 )) must be a key point of the CEVLC.
According to the definition of the CEVLC, u∗ cannot be on the CEVLC and thus must
be in an increasing or decreasing interval. Firstly, let u∗ be in an increasing interval [u1 , u2 ].
Since (u1 , v(u1 )) is a key point on the CEVLC, we have v∗ (u1 ) ≤ v(u1 ). Since v∗ (u∗ ) > v(u∗ )
and v∗ , v are continuous curves, there exists a u0 ∈ [u1 , u∗ ] such that v∗ (u0 ) = v(u0 ). On
[u0 , u∗ ], using Lemma 2.5, we obtain a contradiction. Secondly, let u∗ ∈ [u1 , u2 ] and [u1 , u2 ]
be a decreasing interval. We can consider the movement from u2 to u1 and the acceleration
of v becomes A and the theorem can be proved similarly to the case of increasing intervals.

2.6 Further improvements of the algorithm

In this section, we present modifications to Algorithm VP CETA to improve its efficiency
by getting rid of some unnecessary computations.
We first prove a lemma.

Lemma 2.8 Let (ul , vlim (ul )) and (un , vlim (un )) be two adjacent key points of the CEVLC.
Then an integration trajectory intersects the curve segment vlim (u), u ∈ (ul , un ) once at most.
Proof. We denote by i(u) an integration trajectory. Without loss of generality, we assume
that the tangential acceleration of i(u) is A; otherwise, consider the −u direction. Since the
two key points are adjacent, there are three cases: (a): aT (u) > A, (b): aT (u) < −A, or (c):
178 C.M. Yuan and X.S. Gao

−A < aT (u) < A2) for u ∈ (ul , un ).

Case (a). Since i(ul ) ≤ vlim (ul ), we have i(u) < vlim (u) for u ∈ (ul , un ) by Lemma 2.5.
That is, if they interest, they must intersect at u = ul .
+
Case (b). There are two cases. If vlim (un ) < i(un ), then they must intersect once since
i(ul ) ≤ vlim (ul ) and the two curve segments are continuous. Due to Lemma 2.5, they cannot
+
intersect more than one time. If vlim (un ) > i(un ), then follow the −u direction, aT (u) > A,
by Lemma 2.5 we have i(u) < vlim (u) for u ∈ (ul , un ), thus they do not intersect.
+
Case (c). There are two cases. If vlim (un ) < i(un ) then vlim (u) and i(u) must intersect.
Assuming that (u∗ , i(u∗ )) is an intersection point. Then, from this point in the +u direction,
we have aT (u) < A, hence, vlim (u) < i(u) for u ∈ (u∗ , un ) by Lemma 2.5. From this point
in the −u direction, we have vlim (u) > i(u) for u ∈ (ul , u∗ ). So i(u) intersects vlim (u), u ∈
+
(ul , un ) only once. If vlim (un ) > i(un ), we have aT (u) > −A in the −u direction. By
Lemma 2.5, we have i(u) < vlim (u) for u ∈ (ul , un ) and they do not intersect.

Remark 2.9 Step 8 of Algorithm VP CETA can be modified as follows. Let ul be the
current parametric value, un the parametric value for the next key point of the CEVLC, and
vf (u) the forward integration trajectory starting from point (ul , vs (ul )). Then the tangential
acceleration of vf (u) is A in the +u direction. We can modify step 8 as follows:

8.1 If vf (un ) < vlim (un ), by Lemma 2.8, vf does not meet the CEVLC in (ul , un ] and we
can repeat this step for the next segment of CEVLC until either un = 1 or vf (un ) ≥
vlim (un ).
+ − + −
8.2 If vlim (un ) ≥ vf (un ) ≥ vlim (un ) or vlim (un ) = vf (un ) ≤ vlim (un ), then vf meets the
CEVLC at (un , vf (un )).
+
8.3 Otherwise, we have vf (un ) > vlim (un ) and vf meets the CEVLC in (ul , un ) at a unique
point P by Lemma 2.8. Furthermore, if the current CEVLC segment is not feasible, we
need not to compute this intersection point. Because, in the next step, we will execute
step 9 by computing the backward integration curve vb from u = un and compute the
intersection point Q of vf and vb . Point P is above vb and will not be a part of the
final velocity curve (Fig. 6(a)).

vlim vlim

P
P Q
Q vf u=un
vb
vf u=ul R vlim
vb vbn
u=u l u=u nn
u=u n
(a) Case 8.3: point P is not needed. (b) Step 9’: vb is not needed.
Fig. 6. Modifications of steps 8 and 9

In step 8.3, we need to compute the intersection point between an integration curve and
a feasible CEVLC or between a backward integration curve and a forward integration curve.
2)
In this case, the curve segment is feasible.
 Time-Optimal Control of CNC Machines with Chord Error 179

We can use numerical method to compute it. A simple but useful method to compute these
points is the bisection method, since the intersection point is unique.
Steps 2 and 3 of the algorithm can be modified similarly as step 8.

Remark 2.10 Step 9 can be simplified as follows. We call a parameter un useless if

+ −
vlim (un ) ≥ vlim (un ), a−
T (un ) ≤ −A, and the next CEVLC segment is not feasible. Note
that the second and third conditions mentioned above are equivalent to the following condi-
tion: aT (u) < −A, u ∈ (un , unn ), where unn is the parametric value for the next key point
after un . If un is useless, then we need not to compute the backward integration trajectory
vb from point (un , vlim (un )). Because the backward integration trajectory vbn starting from
unn will be strictly under vb due to Lemma 2.5 (Fig. 6(b)), and as a consequence vb will
not be a part of the final velocity curve due to Theorem 2.2. So, step 9 can be modified as
follows.
Step 9’. If un is useless, we will choose the next key point as un and repeat this step
until either un = 1 or un is not useless. Use this un to compute the backward integration
trajectory and update vs .

Remark 2.11 Due to Theorem 2.2 and Lemma 2.5, we can give the following simpler and
more efficient algorithm. The input and output of the algorithm are the same as that of
Algorithm VP CETA. After computing the CEVLC and its key points, we compute the
velocity curve as follows.

1 Let P be the set of key points of the CEVLC and the starting and ending points (0, 0)
and (1, 0). Set the velocity curve to be the empty set.

2 Repeat the following steps until P = ∅.

3 Let P = (u, vlim (u)) ∈ P be a point with the smallest velocity vlim (u) and remove P from
P.

4 Let VPf and VPb be the forward and backward integration trajectories starting from point
P respectively, which can be computed with the methods in Remark 2.9. If, starting
from point P , the left (right) CEVLC segment is feasible, VPb (VPf ) is set to be this
segment. Find the intersection points of VPb (VPf ) and the existence velocity curve if
needed. Update the velocity curve using VPf and VPb .

5 Remove the points in P, which are above the curve VPf or VPb . This step is correct due to
Theorem 2.2 and Lemma 2.5.

The main advantage of the above algorithm is that many key points are above these
integration trajectories and we do not need to compute the integration trajectories starting
from these points. Also, all the integration trajectories computed in this new algorithm will
be part of the output velocity curve, because each of them starts from the key point which
are not processed and has the smallest velocity.
180 C.M. Yuan and X.S. Gao

3 Time optimal velocity planning of quadratic B-splines and

cubic PH-splines
R
In Algorithm VP CETA, we assume that σdu is computable. In this section, we will show
that for quadratic B-splines and cubic B-splines, we can have a complete
R and efficient time
optimal velocity planning algorithm by giving explicit formulas for σdu. We implemented
our algorithm in these two cases in Maple and used examples from real manufacturing parts
to show the feasibility of our algorithm.

3.1 Velocity planning for quadratic B-splines

Let C(u), u = 0..1 be a quadratic B-spline. Since C(u) has only C 1 continuity and a quadratic
curve has no singular points, the connection points of the spline are all the key points of first
or second type. To compute the key points of third type, consider a piece of the spline:

C(u) = (x(u), y(u), z(u))

= (a0 + a1 u + a2 u2 , b0 + b1 u + b2 u2 , c0 + c1 u + c2 u2 ),

|C 0 ×C 00 |
where a0 , a1 , a2 , b0 , b1 , b2 , c0 , c1 , c2 are constants. The curvature of C(u) is k(u) = ,
p σ3
where σ = |C 0 | = x02 + y 02 + z 02 . For quadratic curves,
p
|C 0 × C 00 | = (b1 c2 − c1 b2 )2 + (c1 a2 − a1 c2 )2 + (a1 b2 − b1 ∗ a2 )2

is a constant. Hence the CEVLC is

B Bσ 3
q = v2 = = 0 = Dσ 3 ,
|k(u)| |C × C 00 |
3 0
where D is a constant. The tangential acceleration along the CEVLC is aT = (Dσ 2σ
)
=
3 0 3 2 0
2 Dσσ = 4 D(σ ) . Hence, aT (u) is a linear function in the parameter u. Then, the key
points of third type can be computed by solving linear equations aT = ±A. From the
equations, we can see that for each piece of the quadratic B-splines, there are two key points
of third type at most.
Now, we show how to compute the integration trajectory. When the tangential acceler-
ation reaches its bounds ±A, we need to compute the solution of the differential equation:

q 0 = ±2Aσ, (11)
√ √
au2 +bu+c √
where σ = au2 + bu + c. Let i(u) = ±2A[1/4 (2 au+b) a +1/2 ln( 1/2√b+au
a
+ au2 + bu + c)
√
c √1a ]−1/8 ln( 1/2√b+au
a
+ au2 + bu + c)b2 a−3/2 . Then, q(u) = i(u)−i(u∗ )+q(u∗ ) is the solution
of the differential equation (11) with initial value (u∗ , q(u∗ )).
We use an example to illustrate the algorithm. The curve in Figure 7(a) is a quadratic
B-spline consisting of six pieces of quadratic curve segments, which is from the tool path of
the vase in Figure 7(b). Details of the G-codes and splines generated from the G-codes for
the vase can be found in [26]. We set the tangential and centripetal acceleration bounds to
be A = 800 mm/s2 and B = 1000 mm/s2 respectively.
 Time-Optimal Control of CNC Machines with Chord Error 181

–2.2

–2.4

–2.6

–2.8

–3

–3.2

–3.4

–3.6

–3.8

–4
48 49 50 51

(a) A quadratic B-spline (b) The vase

12000

2000

10000

1500
8000

6000
1000

4000

500

2000

0 0.2 0.4 0.6 0.8 1

0 0.2 0.4 0.6 0.8 1
u
u

(c) The CEVLC (d) Velocity curve v 2 (u)

800
0.00012

600

0.0001
400

8e–05 200

6e–05 0 0.2 0.4 0.6 0.8 1

u
–200
4e–05

–400

2e–05
–600

0 –800
0.2 0.4 0.6 0.8 1
u

(e) Chord error δ(u) (f) Tangential acceleration

Fig. 7. Optimal velocity planing for a quadratic B-spline

According to the algorithm, we first compute the CEVLC with the maximal centripetal
acceleration, which is shown in Figure 7(c). Some parts of the 1, 2, 3, 5, 6-th pieces of the
CEVLC are feasible. Secondly, we compute the integration trajectory vs from the starting
point (u0 , v0 ) = (0, 0) by tangential acceleration A, which intersects the feasible part of the
first CEVLC segment. Find the intersection point of vs and the feasible CEVLC segment.
Then, compute the backward integration trajectory ve from (1, 0), similarly as above, it
intersects the sixth CEVLC segment. Then, execute step 7 and merge the feasible part of the
first CEVLC segment into vs . Let the end point of the feasible CEVLC (next key point) be
the current point with parameter u0 . Now, since a− T (u0 ) ≥ A, we execute step 8 by computing
the integration trajectory vf with tangential acceleration A. The segment of vf between the
current point and the first connection point of the spline is under the CEVLC. Merge vf
+
into vs . Let u1 be the parametric value of the first connection point. Since vlim (u1 ) <
− +
vlim (u1 ), we execute step 8 again by computing a new vf starting from (u1 , vlim (u1 )) with
182 C.M. Yuan and X.S. Gao

tangential acceleration A. vf intersects the CEVLC at the second connection point of the
spline. Merge vf into vs . Let u2 be the parametric value of the second connection point.
−
Since vs (u2 ) > vlim (u2 ), according to the algorithm, we execute step 9 of by computing the
−
backward integration trajectory vb from (u2 , vlim (u2 )), which intersects vs only once. Update
vs accordingly. In this way, we can obtain the final velocity curve shown in Figure 7(d),
which consists of eleven pieces, where the 1, 3, 4, 5, 7, 8, 9, 11-th pieces are controlled by the
tangential acceleration, and the 2, 6, 10-th pieces are feasible CEVLC segments. Figure 7(e)
is the chord error of the optimal velocity curve with a sampling period T = 1ms, which
means d = 1.25µm. Figure 7(f) is its tangential acceleration. From these two figures, we
can see that the control is “Bang-Bang.”
Now, we give a more complicated example shown in Figure 8(a), which is a complete tool
path segment C(u) = (x(u), y(u)) of the vase in Figure 7(b) from top to bottom and consists
of 57 quadratic B-spline curve segments and 5 long straight line segments. Figure 8(b) is
its velocity limited curve with centripetal acceleration B = 1000 mm/s2 . Figure 8(c) is
the optimal velocity curve computed with our algorithm with tangential acceleration A =
800 mm/s2 and centripetal acceleration B = 1000 mm/s2 , which consists of 127 curve
segments. Figure 8(d) is the chord error of the optimal velocity curve with a sampling
period T = 1ms, which means d = 1.25µm. Figure 8(e) is its tangential acceleration. Note
that in the connection point of a straight line segment and a quadratic B-spline, the velocity
decreases to zero.

3.2 Velocity planning for cubic PH-splines

Let C(u), u = 0..1 be a cubic PH-spline. Since a cubic PH-spline only has C 1 continuity,
the connection points of the PH-spline are all the key points of first or second type of the
CEVLC. Let r(u) be a piece of cubic PH-curve of C(u). Then, r0 (u) has the following
representation [8]:
r0 (u) =(f (u)2 + g(u)2 − m(u)2 − n(u)2 ,
2(f (u)n(u) + g(u)m(u)), (12)
2(g(u)n(u) − f (u)m(u))).
where f (u), g(u), m(u), n(u) are linear functions in u.
0 00 |
The curvature of r(u) is k(u) = |r σ×r 3 = σE2 , where σ = |r0 | = f (u)2 + g(u)2 + m(u)2 +
n(u)2 and E is a constant. The CEVLC of r(u) is
B
q = v2 = = Gσ 2 ,
|k(u)|
2 0
where G is a constant. The tangential acceleration along the CEVLC is: aT = (Gσ 2σ
)
= Gσ 0 .
Since σ is of degree two, aT (u) is a linear function in the parameter u. The key points of
the third type can be computed by solving linear equations aT = ±A directly.
For a cubic PH-curve, when the velocity is controlled by the tangential acceleration, the
integration trajectory is a polynomial in u with degree three. The integration trajectory is
the solution of the following differential equation:
q 0 = ±2Aσ = ±2A(au2 + bu + c), (13)
 Time-Optimal Control of CNC Machines with Chord Error 183

(a) quadratic B-splines r(u) = (x(u), y(u))

2 (u)
(b) Velocity limited curve qlim (u) = vlim

(c) Velocity curve q = v 2 (u)

(d) Chord error δ(u)

(e) Tangential acceleration aT (u)

Fig. 8. Optimal velocity planing for quadratic B-splines from a vase

where a, b, c are constants. Let i(u) = ±2A(1/3 au3 + 1/2 bu2 + cu). Then, q(u) = i(u) −
184 C.M. Yuan and X.S. Gao

10
25000

20000
6

4 15000

10000

0 5 10 15 20 25

–2 5000

0 0.2 0.4 0.6 0.8 1

(a) Cubic PH-spline (b) The CEVLC

3000

20000
0.00012
2000

0.0001
1000
15000

8e–05
0 0.2 0.4 0.6 0.8 1

10000 u
6e–05
–1000

4e–05
5000 –2000

2e–05
–3000

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(e) Tangential acceleration

u u

(c) Velocity q = v 2 (u) (d) Chord error δ(u)

of velocity curve
Fig. 9. Optimal velocity planning for a cubic PH-spline

i(u∗ ) + q(u∗ ) is the solution of the differential equation (13) with initial value (u∗ , q(u∗ )).
Now, we give an illustrative example for a cubic PH-spline (Fig. 9(a)).
The cubic PH-spline consists of two pieces of PH-curves and has C 1 continuity. Let the
tangential and centripetal acceleration bounds be A = 3000mm/s2 and B = 1000mm/s2
respectively. Here are the details to get the velocity curve. Firstly, we compute the CEVLC
of the curve which consists of two feasible segments. So the key points are the starting
point, the ending point, and the connection point of the spline. Let u∗ be the parametric
value of the connection point of the spline. Secondly, we compute the integration tra-
jectory vs from the starting point (u0 , v0 ) = (0, 0) by tangential acceleration A. Since
vs (u∗ ) > vlim (u∗ ), we see that it must intersect the first CEVLC segment. Find the inter-
section point (u1 , vlim (u1 )) of vs and the first CEVLC segment. Let [0, u1 ] be the defining
interval of vs . Compute the backward integration trajectory ve from the ending point (1, 0).
+ −
Since vlim (u∗ ) < ve (u∗ ) < vlim (u∗ ), ve does not intersect the second CEVLC segment. Let
(u∗ , ve (u∗ )) be the intersection point of ve and the CEVLC and [u∗ , 1] be the defining inter-
val of ve . Starting from (u1 , vlim (u1 )), the CEVLC is feasible, we merge this feasible part
+
of CEVLC into vs . Let (u∗ , vs (u∗ )) be the current point. Since vs (u∗ ) = vlim (u∗ ) < ve (u∗ ),
execute step 8 by computing the integration trajectory vf which intersects ve at (u2 , vf (u2 )).
The final velocity curve has four segments, where the 1, 3, 4-th segments are controlled by
the tangential acceleration and the 2-th piece is a feasible CEVLC segment.
The CEVLC, velocity curve , chord error and tangential acceleration of the velocity curve
are shown in Figure 9, where the sampling period is T = 1ms.
 Time-Optimal Control of CNC Machines with Chord Error 185

4 Feed-rate override in NC-machining

During the NC-machining, there exists another constraint: the maximal feed-rate vmax . In
this situation, all we need to do is to change the velocity curve to q ∗ (u) = min(q(u), vmax ),
where q(u) is the optimal velocity curve obtained in the preceding section. The procedure
of the interpolation is as follows.

Algorithm 4.1 Interpolation algorithm

Input: current parameter ui , the velocity curve q(u), maximum feed-rate vmax , and the
sampling time T
Output: the parameter of the next interpolation point ui+1
p
1. According to the velocity curve and the maximum feed-rate, let vi = min( q(ui ), vmax ).
The step size is ∆L = vi · T .

2. According to the step size ∆L, compute the parameter of the next interpolation point
ui+1 ([4, 21]).

Since for any parametric value u, whenever the value of q ∗ (u) is taken from q(u) or vmax ,
the left and right limitations of the tangential acceleration are satisfied. Hence, q ∗ (u) satisfies
the maximum feed-rate, chord error, and tangential acceleration bounds. Furthermore, in
CNC-machining, the users can change the maximum feed-rate during manufacturing, which
is called feed-rate override. Although the new feed-rate limitation is not required to respond
immediately, the real velocity should decrease to the new lower speed as soon as possible.
The following algorithm solves the feed-rate override problem efficiently.

Algorithm 4.2 Feed-rate override algorithm

Input: the velocity curve q(u), current parameter ui , the current feed-rate v∗ , the modified
0
feed-rate limitation vmax ;
Output: the parametric values ui+1 , ui+2 , . . . of the interpolation points.
0
p
1 If v∗ > vmax , then let vi0 = v∗ − AT , vi00 = max(vi0 , vmax
0 ), vi = min(vi00 , q(ui )). Find
the next interpolation point ui+1 according to Algorithm 4.1, i = i + 1, v∗ = vi , and
repeat step 1;
0
p 0
2 If v∗ ≤ vmax , then let vi = min( q(ui ), vmax , v∗ + AT ). Find the next interpolation
point ui+1 according to Algorithm 4.1. If ui+1 > 1, then let ui+1 = 1 and terminate;
else let i = i + 1, v∗ = vi , and repeat step 2.

Since the final velocity is under the CEVLC, the error bound is satisfied. Step 1 of the
above algorithm is to slow down feed-rate as soon as possible when the current feed-rate is
larger than the modified feed-rate. Step 2 of the above algorithm is exactly Algorithm 4.1,
where the maximum feed-rate is replaced by the modified feed-rate.
One advantage of the algorithm presented in this paper is that feed-rate override can be
carried out easily. In the case of multi-axis acceleration mode, when the maximal feed-rate
is changed to vmax , we cannot simply take q ∗ (u) = min(q(u), vmax ) to be the new velocity
curve.
186 C.M. Yuan and X.S. Gao

5 Conclusion
In this paper, we give the first time-optimal velocity planning method for parametric curves
with confined chord error. We adopt the simplest acceleration mode: the linear acceleration
for tangential accelerations. With the method introduced in this paper, it is not difficult
to give time-optimal velocity planning method with confined chord error and multi-axis
acceleration modes.
The key idea is to reduce the chord error bound to a centripetal acceleration bound.
When the centripetal acceleration reaches its bound, the velocity curve is an algebraic curve
and is called the CEVLC. With the CEVLC, the final velocity curve is the minimum of
all the integration velocity curves starting from the key points of the CEVLC the starting
point, and the ending point. We also give a practical algorithm to compute the time-optimal
velocity curve and implemented the algorithm for two types of curves.

References
[1] D. Bedi, I. Ali, N. Quan. Advanced techniques for CNC machines, Journal of Enginerring
for Industry, 115, 329-336, 1993.

[2] G. Birkhoff, G. Rota. Ordinary differential equations. Newyork, Blaisdell, 1969.

[3] J.E. Bobrow, S. Dubowsky, J.S. Gibson. Time-optimal control of robotic manipulators
along specified paths. Int. J. Robot. Res., 4(3), 3-17, 1985.

[4] C.W. Cheng, M.C. Tsai. Real-time variable feed rate NURBS curve interpolator for CNC
machining. Int. J. Adv. Manuf. Technol., 23, 865-873, 2004.

[5] J.J. Chou and D.C.H. Yang. Command generation for three-axis CNC machining. Journal
of Engineering for Industry, 113, 305-310, 1991.

[6] M.M. Emami, B. Arezoo. A look-ahead command generator with control over trajectory
and chord error for NURBS curve with unknown arc length. Computer-Aided Design,
4(7), 625-632, 2010.

[7] K. Erkorkmaz, Y. Altintas. High speed CNC system design. Part I: jerk limited trajectory
generation and quintic spline interpolation. Int. J. of Mach. Tools and Manu., 41, 1323-
1345, 2001.

[8] R.T. Farouki. Pythagorean-hodograph curves, Springer, Berlin, 2008.

[9] R.T. Farouki, Y.F. Tsai. Exact Taylor series coefficients for variable-feedrate CNC curve
interpolators. Computer-Aided Design, 33(2), 155-165, 2001.

[10] J.Y. Lai, K.Y. Lin, S.J. Tseng, W.D. Ueng. On the development of a parametric inter-
polator with confined chord error, feedrate, acceleration and jerk. Int. J. Adv. Manuf.
Technol. , 37: 104C121, 2008.
 Time-Optimal Control of CNC Machines with Chord Error 187

[11] M.T. Lin, M.S. Tsai, H.T. Yau. Development of a dynamics-based NURBS interpolator
with real-time look-ahead algorithm. Int. J. of Mach. Tools and Manu., 47(15), 2246-
2262, 2007.

[12] S. Macfarlane, E.A. Croft. Jerk-bounded manipulator trajectory planning: design for
real-time applications. IEEE Trans. on Robot. and Automa., 19: 42-52, 2003.

[13] S.H. Nam, M.Y. Yang. A study on a generalized parametric interpolator with real-time
jerk-limited acceleration. Computer-Aided Design 36, 27-36, 2004.

[14] J. Park, S.H. Nam, M.Y. Yang. Development of a real-time trajectory generator for
NURBS interpolation based on the two-stage interpolation method. Int. J. Adv. Manuf.
Technol., 26: 359-365, 2005.

[15] Z. Shiller. On singular time-optimal control along specified paths. IEEE Trans. Robot.
Autom., 10, 561-566, 1994.

[16] Z. Shiller, H.H. Lu. Robust computation of path constrained time optimal motions.
Proc., IEEE Inter. Conf. on Robot. Autom., Cincinnati, OH, 144-149, 1990.

[17] K.G. Shin, N.D. McKay. Minimum-time control of robotic manipulators with geometric
path contraints. IEEE Trans. on Automatic Control, 30(6), 531-541, 1985.

[18] M. Shpitalni, Y. Koren, C.C. Lo. Realtime curve interpolators. Computer-Aided Design
26(1), 832-838, 1994.

[19] S.D. Timar, R.T. Farouki, T.S. Smith, C.L. Boyadjieff. Algorithms for time-optimal
control of CNC machines along curved tool paths. Robotics and Computer-Integrated
Manufacturing, 21(1), 37-53, 2005.

[20] S.D. Timar, R.T. Farouki. Time-optimal traversal of curved paths by Cartesian CNC
machines under both constant and speed-dependent axis acceleration bounds. Robotics
and Computer-Integrated Manufacturing, 23(5), 563-579, 2007.

[21] Z.M. Xu, J.C Chen and Z.J Feng. Performance Evaluation of a Real-Time Interpolation
Algorithm for NURBS Curves. Int. J. Adv. Manuf. Technol., 20: 270-276, 2002.

[22] D.C.H. Yang, T. Kong. Parametric interpolator versus linear interpolator for precision
CNC machining. Computer-Aided Design, 26(3), 225-234, 1994.

[23] S.S. Yeh, P.L. Hsu. Adaptive-feedrate interpolation for parametric curves with a confined
chord error. Computer-Aided Design, 34, 229-237, 2002.

[24] T. Yong, R. Narayanaswami. A parametric interpolator with confined chord errors,

acceleration and deceleration for NC machining. Computer-Aided Design, 35, 1249-1259,
2003.

[25] K. Zhang, X.S. Gao, H. Li, C.M. Yuan. A greedy algorithm for feed-rate planning of
CNC machines along curved tool paths with jerk constraints. MM Research Preprints,
29, 189-205, 2010.
 View publication stats

188 C.M. Yuan and X.S. Gao

[26] M. Zhang, W. Yan, C.M. Yuan, D. Wang, X.S. Gao. Curve fitting and optimal in-
terpolation on CNC machines based on quadratic B-splines(in Chinese). MM Research
Preprints, 29, 71-91, 2010.