Mechanism and Machine Theory 130 (2018) 1–26

Contents lists available at ScienceDirect

Mechanism and Machine Theory

journal homepage: www.elsevier.com/locate/mechmachtheory

Research paper

Design, analysis, and experimental validation of an active

constant-force system based on a low-stiffness mechanism
Zhen Liu∗, Fuliang Niu, Haibo Gao, Haitao Yu, Liang Ding, Nan Li,
Zongquan Deng
State Key Laboratory of Robotics and System, Harbin Institute of Technology, Harbin 150001, China

a r t i c l e i n f o a b s t r a c t

Article history: In low-gravity suspension simulation experiments, the partial gravitational forces of tested
Received 26 April 2018 objects are balanced by the constant vertical forces on cables generated by constant-force
Revised 11 July 2018
systems. To improve system robustness against external payload disturbance, such sys-
Accepted 26 July 2018
tems usually employ low-stiffness mechanisms. The schematic diagram of our proposed
low-stiffness mechanism is derived from an energy approach, which is especially prefer-
Keywords: able when the low-stiffness mechanism comprises two kinds of elastic components. The
Constant-force mechanism mechanism uses a combination of an axially arranged torsion bar and a group of radially
Low-stiffness mechanism arranged springs. While the former exhibits high energy density and generates major out-
Constant-force system put force, the latter offers a negative stiffness to shape the output force curve so that it
Gravity compensation system resembles a constant one. The mechanism has a comparatively smaller overall size, lower
stiffness, and wider adjustable force range. The low-stiffness mechanism is used to form an
active constant-force system. The system, as well as its dynamic model and controller, are
also detailed in this paper. Experimental results demonstrate that the active constant-force
system can be robustly controlled by a proportional-derivative controller with incomplete
derivation to generate a high-accuracy dynamic force.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction

One of the most significant current discussions in astronaut training and spacecraft reliability assessment is the ground
simulation of a low-gravity environment on the planet’s surface [1, 2]. An alternative method for the simulation is using
slings to apply vertical forces on tested objects to compensate for their partial gravity [3, 4]. Such a method usually consists
of a horizontal position tracking system maintaining the sling vertical and a constant-force system keeping the sling force
constant. The performance of a constant-force system is a key factor to ensure high-fidelity simulation. Two requirements
for constant-force systems are high steady-state force accuracy and excellent dynamic response.
Currently, research results regarding to constant-force systems can be classified into passive constant-force systems and
active constant-force systems (ACFSs). Passive constant-force systems are generally classified in terms of compensation com-
ponents into two types: counterweights [5–7] and buffer springs [8–10]. They are simple in structure; however, the counter-
weight inertia and spring deformation significantly affect the force accuracy and dynamic characteristics. Consequently, they
are suitable for low-speed or low-force accuracy applications. ACFSs are divided on the basis of drive source and composition

∗
Corresponding author.
E-mail address: [email protected] (Z. Liu).

https://doi.org/10.1016/j.mechmachtheory.2018.07.019
0094-114X/© 2018 Elsevier Ltd. All rights reserved.
2 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Nomenclature

L Lagrangian operator (J)

VLSM Elastic potential energy of LSM (J)
q Generalised degree of freedom and angle between input and output terminals of low-
stiffness mechanism (LSM) (°)
Q(q) Generalised force (N m)
C Stored elastic energy at initial working point (J)
QC Setting generalised force (N m)
T Torsional stiffness of torsion bar (N m/°)
qst Initial angle of torsion bar (°)
Vtor , Vspr Elastic potential energy of torsion bar and compressed spring (J)
VACFS Elastic potential energy of ACFS (J)
Vrem Elastic potential energy of elastic elements except torsion bar (J)
Vrmax Maximum elastic potential energy of elastic elements except torsion bar (J)
C1 Initial elastic potential energy of elastic elements except torsion bar (J)
qmax(Vrem ) Angle under the maximum elastic potential energy of remaining elastic elements (°)
qH Maximum angle of one side of the initial position of LSM(°)
TEACFS Torque error function of ACFS (N m)
FEACFS Force error function of ACFS (N)
KACFS Equivalent stiffness of LSM and ACFS (N/mm)
d Torsion bar diameter (mm)
Ltor Torsion bar length (mm)
k Stiffness of compressed spring (N/mm)
l0 , lst Rest and installed lengths of compressed spring (mm)
Fpre Preload of compressed springs (N)
l(q) Length function of compressed spring (mm)
e Distance from outer joint of compressed spring to rotation axis of LSM (mm)
r Distance from inner joint of compressed spring to rotation axis of LSM (locking bar with
constant length) (mm)
α (q) Angle function between compressed spring axis and locking bar direction (°)
Fspr Force of compressed spring (N)
R Reel radius (mm)
τ Motor torque (N m)
J1 Equivalent moment of inertia of motor shaft, brake, reducer, and input terminal of LSM
(kg cm2 )
J2 Equivalent moment of inertia of reel and output terminal of LSM (kg cm2 )
Jmotor , Jbrake , Jreducer , Jcoupling Moment of inertia of motor, brake, reducer, and coupling (kg cm2 )
i Transmission ratio
JIn Equivalent moment of inertia of input terminal of LSM (kg cm2 )
μ1 Equivalent viscous friction coefficient of motor shaft, brake, reducer, and input terminal of
LSM (N s/mm)
μ2 Equivalent viscous friction coefficient between input and output terminals of LSM
(N s/mm)
μ3 Equivalent viscous friction coefficient between output terminal of LSM and support
(N s/mm)
θ1 Rotation angle of input terminal of LSM (°) (Positive direction is clockwise viewed from
the motor to the LSM)
θ2 Rotation angle of output terminal of LSM (°) (Direction is the same asθ 1 )
TLSM Actual torque of LSM (N m)
FLSM Actual force of LSM (N)
Fst Required force of LSM (N)
F Sling force (N)
F Sling force error (N)
f Friction force (N)
FC Coulomb friction force (N)
v Relative sliding velocity (m/s)
μ Viscous friction coefficient (N s/mm)
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 3

into pneumatic servo pattern [11, 12], motor-reel pattern [13, 14], and a pattern in which a low-stiffness mechanism (LSM)
providing a low-stiffness force output by using the property of the mechanical structure is used between the motor and
reel [15–17]. The two former patterns are easily achieved. However, the high force accuracy and dynamic performance place
great demands on the controller. Owing to the cushioning and low-stiffness properties of LSMs, the third pattern minimises
the control design effort and exhibits a superior dynamic performance under complex force interactions between payloads
and the environment.
A considerable amount of literature has been published on ACFSs comprising LSMs. The active-response gravity-offload
system developed by NASA innovatively cascades a pneumatic muscle in the sling to improve the dynamic performance. Frey
et al. [15] provided a body weight support system in the form of a parallel connection between a passive spring taking over
the main unloading force and an active electric drive to generate the exact desired force. Yang et al. [16] built an active body
weight support system using an electric cylinder that controlled the cable force by adjusting the spring deformation and a
winch motor controlling the cable displacement to provide a well-controlled supportive force against gravity. Conventional
passive springs, however, have the limitation of large force fluctuations under a large displacement. Xiang et al. [17] devel-
oped an active suspension gravity compensation system composed of a passive constant-force mechanism absorbing high
frequency perturbation and an active controlled electric winch providing the main compensation force to obtain the desired
compensation force. It should be noted that the significant friction caused by the deformation of the braided slings on the
pulleys places limitation on the output force accuracy. A major challenge with ACFSs is to develop an LSM with a simpler
structure, lower output stiffness, larger output force, and less friction.
An LSM is equivalent to a low stiffness spring outputting a large force to balance the partial gravity of the payload with
vertical displacement. When the stiffness is reduced to 0, the LSM is known as a constant-force mechanism. Furthermore,
the design principles of LSMs and constant-force mechanisms are the same. A large volume of published studies has de-
scribed the development of specially designed constant-force mechanisms. Wang et al. [18] conducted a survey of recent
developments on the design and modelling of constant-force mechanisms. These mechanisms were divided according to
working principles into compliant and conventional constant-force mechanisms. The former realises its function by relying
on the deformation of flexible members, while the latter is designed based on a rigid structure and elastic elements. Xu
and co-workers [19–21] and Lan and co-workers [22–24] developed many compliant constant-force mechanisms. The com-
pliant constant-force mechanisms, without friction and backlash, have simple structures and small overall size. They are
therefore widely used in micro-electro-mechanical systems, where the payloads are generally lighter and fixed. The loads
in gravity compensation experiments, however, are much heavier and vary from time to time. It would cost considerable
effort to design compliant LSMs for gravity compensation. In contrast, conventional LSM, with large output ranges and well
adjustability, are suitable for this application, such as the mechanisms developed by Barents et al. [25] and Xiang et al. [17].
According to our experiences, these mechanisms could suffer from backlash and friction caused by deformation of braided
slings on pulleys. This study aims to developed a novel conventional LSM without pulley to achieve a large and exact desired
force with a small overall size, lower stiffness, and wider adjustable force range.
There has been considerable amount of work on design theory of LSMs. Wang et al. [18] pointed out that these mech-
anisms are mainly used in static and quasi-static applications. The theoretical equations are therefore mainly kinetostatic
equations. Dynamic model, which is necessary to develop ACFS controller, hasn’t attracted as much attention. For conven-
tional constant-force mechanisms, theoretical equations are mainly constant force component equation [26], elliptical track
equation [27], zero-free-length spring equation [25], and positive and negative stiffness superposition equation [28, 29]. For
flexible constant-force mechanisms, the theoretical equation can be distributed-compliance model [22] and most of these
mechanisms usually contain heterotypic elastic elements whose shapes can be calculate by the optimization method. The
essence of LSMs is that the work done by the constant force on the output displacement is completely stored as the elastic
potential energy or released from the elastic potential energy. There is a linear relationship between the elastic potential
energy and the displacement, the slope or which is the amplitude of constant force. this paper carries out the LSM design
from the perspective of energy. This approach is not mechanism specific. It’s especially preferable when the LSM consists
of two kinds of elastic components. It illustrates the energy-displacement relation of one when that of the other is known.
The rest of work is then finding a mechanism that produce similar energy-displacement curve, and shaping the curve by
adjusting parameters to resemble the desired curve.
The energy-angle curve of a conventional elastic element is a convex quadratic function. Therefore, a constant torque
output can be achieved by the combination of a conventional elastic element and a specially designed element whose
energy curve is a concave quadratic function. Furthermore, the concave-energy elastic element is a negative-stiffness one
releasing energy when the LSM rotates. Consequently, multiple compressed springs whose maximum preload occurs at the
initial working point are radially arranged between the input and output terminals of LSMs. Although we cannot ignore
the scheme mentioned in reference [22], it should be noted that payload is a key factor contributing to the selection of an
LSM. Considering the large desired force, small overall size, and adjustable characteristics in our low-gravity compensation
application, this study performed a detailed analysis, design, and experimental verification of this scheme.
This paper aims at developing an ACFS based on a novel LSM to realize constant force. In Section 2, the energy analysis
is proposed to analysis the essence of LSM and obtain the schematic of the LSM. The dimensional synthesis and structural
design are carried out. Section 3 establishes the ACFS based on the proposed LSM. A motor is used to rotate the LSM to
achieve a large-range constant force in vertical direction. The dynamic modelling and control system design are performed
respectively in Sections 4 and 5. This is followed by the experimental verification and discussion in Section 6.
4 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

2. Design of LSM used in ACFS

The LSM used in the ACFS is controlled by direct drive motor in closed loop. The proposed LSM differ from static con-
stant force mechanisms in the following ways: (1) the energy density of the proposed LSM should be high so as to ensure
low inertia and better dynamic performance; (2) the equivalent stiffness is low but not 0 in order to meet the control re-
quirements; (3) The required force should be adjustable so that the LSM can produce nearly constant force over a wide
range to cover different load weights.
According to aforementioned requirements, a torsion bar was selected because of its high energy density. It is then
assumed that the torsional bar should work together with a second kind of elastic components to produce nearly constant
force. A constant force mechanism is an energy storage mechanism absorbing or releasing its own elastic potential energy
almost linearly to its generalised movement. The difference between the energy-displacement relation of the torsional bar
and the LSM equals that of the desired second kind of elastic component. The rest of work is then finding a mechanism that
produce similar energy-displacement curve, and shaping the curve by adjusting parameters to resemble the desired curve.

2.1. Energy analysis and schematic design

2.1.1. Energy equation

The design principle of the LSM is that the torque does not change under the relative rotation between the input and
output terminals. The second kind of Lagrangian equation is used to equate the LSM moment balance relationship to the
energy change relationship. The relative rotation angle between the input and output terminals is defined as a generalised
degree of freedom q, while the payload torque is defined as a generalised force Q(q). The general form of the Lagrangian
equation is
 
d ∂L ∂L
− = Q (qi ), i = 1, 2...n, (1)
dt ∂ q˙ i ∂ qi
In the design of the LSM, the kinetic energy of the mechanism should be regarded as 0. The Lagrangian operator L
then satisfies L = −VLSM . In accordance with the quasi-static design principle, the generalised velocity q˙ and generalised
acceleration q̈ should considered to be 0. According to the relationship L = −VLSM and considering that the LSM has only
one rotational degree of freedom, Eq. (1) can be simplified as 1
dVLSM
= Q (q ). (2)
dq
By integrating Eq. (2), the energy-angle relationship of the LSM is

VLSM = Q (q )dq + C. (3)

The constant value C in Eq. (3) is the stored elastic energy at the initial working point to maintain the Q(q) of the LSM.
When Q(q) is a constant value, Eq. (3) changes to

VLSM = QC q + C. (4)
Eq. (4) is the equation needed to design the LSM; that is, the relationship between the energy and rotation angle is a
linear function.
A torsion bar with higher energy density was selected to provide the main torque considering the requirement of large
compensation force and small overall size. The energy-angle relationship of the torsion bar satisfies
1 
Vtor = T (q + qst )2 . (5)
2
It can be concluded from Eq. (5) that the energy-angle relationship of the torsion bar is a convex function. The total
initial energy Cof the LSM can be written as
1  2
C= T qst + C1 . (6)
2
Where C1 is the initial elastic potential energy of the elastic elements except the torsion bar.
The energy of elastic elements except the torsion bar can be obtained by the difference of Eqs. (4) and (5),
Vrem = VLSM − Vtor
. (7)
= − 12 T  q2 + (QC − T  qst )q + C1
It can be seen from Eq. (7) that the energy curve of such mechanisms is a concave quadratic curve. These mechanisms
can be developed from general negative stiffness mechanisms. Given that the mechanism should work together with the
torsion bar, it should be a rotational negative stiffness mechanism. Such elements do not have full-stroke negative stiffness
but exhibit negative stiffness within only a limited stroke after reaching preload.
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 5

Fig. 1. Three positional relationships between the angle interval and the angle at maximum energy.

2.1.2. Energy minimisation for elastic elements except torsion bar

It can be seen from Eq. (7) that the highest energy point of the elastic potential energy of elastic elements except the
torsion bar corresponds to the mechanical rotation angle of
QC
qmax(Vrem ) = − qst . (8)
T
qH is the maximum angle limited on one side of the initial position of LSM, and the angle q satisfies

−qH ≤ q ≤ qH . (9)
Take into account the size reduction of the LSM, the high energy-density torsion bar is placed in axial position. The
energy of other elastic elements needs to be as low as possible, that is, the value of the Eq. (7) should be as small as
possible. The analysis of the minimum energy is performed using a classification method shown in Fig. 1. The green curve
in the figure is the energy curve of the elastic element except the torsion bar, and C1 is the initial energy when q = 0. The
value of C1 can be changed by design, and its change will cause the green curve to translation upwards or downwards in
Fig. 1. The following classification discussion is therefore based on the vertical translation of the curve and the determination
of C1 to derive the maximum energy for each type of situation.
In the following analysis, a supposition is used that the energy of the lowest energy point in the angular range is zero,
that is, the ordinate of the red double dot in Fig. 1 is 0 and the horizontal axis (solid black line) is up-regulated (blue
dashed line). This supposition is an extreme case in the design of the actual LSM, that is, the elastic element except the
torsion bar has zero energy at the lowest point of energy. This goal can be achieved by mechanical design. However, at this
extreme position, the output of the LSM comprise only the torsion bar output, which is not equal to the set value. Once
the mechanism rotates away from the lowest point of energy, the output force of the mechanism is the set force. In other
words, the constant force interval of the LSM under this supposition is (−qH , qH ), not [−qH , qH ].
The energy values at the three angles can be written as:
1
Vrem (−qH ) = − T  qH 2 − (QC − T  qst )qH + C1 , (10)
2

1
Vrem (qH ) = − T  qH 2 + (QC − T  qst )qH + C1 , (11)
2

 2
  QC − T  qst
Vrem qmax(Vrem ) = + C1 . (12)
2T 
For case (a), the following conditions should be met when the energy is minimum:

Vrem (−qH ) = 0
. (13)
qH ≤ qmax(Vrem )

According to Eqs. (8), (10), (11) and (13), the maximum energy Vrmax of the remaining elastic elements can be obtained:

Vrmax = Vrem (qH ) ≥ 2T  q2H . (14)

For case (b), the following conditions should be met when qVmax ≥ 0 and energy are minimum:

Vrem (−qH ) = 0
. (15)
0 ≤ qmax(Vrem ) ≤ qH
6 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Table 1
Constant parameters.

Item Unit Value

QC (N m) 100
qH (°) 15
T (N m/°) 2.5
qst (°) 40
C1 (J) 25
16
π

Fig. 2. Energy curves of LSM.

According to Eqs. (8), (10), (12) and (15), the maximum energy Vrmax of the remaining elastic elements can be obtained:

1  2  
T qH ≤ Vrmax = Vrem qmax(Vrem ) ≤ 2T  q2H . (16)
2
The following conditions should be met when qVmax ≤ 0 and energy are minimum:

Vrem (qH ) = 0
. (17)
−qH ≤ qmax(Vrem ) ≤ 0

According to Eqs. (8), (11), (12) and (17), the maximum energy Vrmax of the remaining elastic elements can be obtained:

1  2  
T qH ≤ Vrmax = Vrem qmax(Vrem ) ≤ 2T  q2H . (18)
2
For case (c), the following conditions should be met when the energy is minimum:

Vrem (qH ) = 0
. (19)
qmax(Vrem ) ≤ −qH

According to Eqs. (8), (10), (11) and (19), the maximum energy Vrmax of the remaining elastic elements can be obtained:

Vrmax = Vrem (−qH ) ≥ 2T  q2H . (20)

Comparing Eqs. (14), (16), (18) and (20), the energy value Vrmax in case b) is the lowest, and the minimum value of
energy Vrmax = 12 T  q2H is obtained at qmax(Vrem ) .
That is, the design condition when the energy of the LSM is the lowest is
 QC
qst = T
1 
. (21)
C1 = 2
T qH 2
The set of parameters selected according to Eq. (21) is shown in Table 1. They are used to draw the energy-angle curves
of the LSM and torsion bar, as shown in Fig. 2. The difference between the two curves is the energy-angle curve of the
elastic elements except the torsion bar.
It should be noted that conditions (21) is satisfied in the most ideal situation, in which the elastic energy of elastic
elements except torsion bar is decreased to 0 at maximum work angle. It implies that these elastic elements are completely
relaxed. The aforementioned analysis has the following guidance on the design of LSM.
First, the minimum energy value of a realistic mechanism cannot reach 0 due to the limitations of specific mechanical
forms. Moreover, the stiffness of the general elastic element in its initial deformation section is smaller than that of the
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 7

Fig. 3. (a) Spring-cam mechanism, (b) double compressed spring mechanism, (c) spring crank-rocker mechanism.

Fig. 4. Design principle diagram of LSM.

working section, so elastic elements are not allowed to fully relax, i.e. potential energy reaches 0 in [−qH , qH ]. In a design
process, the energy curve of the remaining elastic elements should be close to the horizontal axis, rather than touching 0 at
maximum work angle.
Second, in order to make the energy lower in the whole working stroke, the change of Vrem should be smaller, that is,
the stiffness coefficient T of the torsion bar in the concave quadratic function represented by the Eq. (7) should be as small
as the mechanism allow it to be.
Third, the derivative of the energy curve represented by Eq. (7) is a force function, which is linear. Considering that
an actual mechanism’s energy curve does not completely fit the quadratic function of Eq. (7), there is a force error. The
linearization of the force function will contribute to the reduction of the force error.

2.1.3. Schematic design

The key to design an LSM is to design a mechanism that can release energy in both directions, as shown by the remaining
curve in Fig. 2. This type of mechanisms has energy releasing characteristics on both sides of the installation position. Such
mechanisms are easy to be designed, such as: spring-cam mechanism (Fig. 3(a)), spring crank-rocker mechanism (Fig. 3(b)),
double compressed spring mechanism (Fig. 3(c)), etc.
Considering that the LSM is to be used in an ACFS, the mechanism of the LSM should be simple. Moreover, both of its
equivalent stiffness and output torque should be adjustable. The cam in the configuration of Fig. 3(a) guide the spring and
force it to deform. The trajectory of the cam, however, cannot be adjusted after the machining is completed. Neither the
equivalent stiffness nor the output torque can be adjusted in mechanism (a). The mechanism in Fig. 3(b) is complex and the
friction is also large. The mechanism in Fig. 3(c) is relatively simpler. In this mechanism, all the motion pairs are rotation
pairs which introduce much less friction than translational pairs and contact pairs.. After comprehensive consideration, the
mechanism (c) is selected for the LSM in this paper.
Multiple compressed springs whose maximum compression occur at the initial working point are radially arranged be-
tween the input and output ends of the LSM. The compressed springs provide small forces for compensating the force
change of the torsion bar so that the LSM outputs an approximately constant force. The design principle diagram of the
LSM is shown in Fig. 4. The LSM consists of an input shaft, compressed springs, output reel, and torsion bar. Two transitive
force routes exist: one is the force transmitted from the input shaft to the output reel via the torsion bar, and the other is
via the multi-group compressed springs radially arranged between the input shaft and the output reel. When the sling force
is changed abruptly, relative rotation occurs between the input shaft and output reel and the internal elastic elements are
deformed to absorb or release energy. Thus, the external disturbance effect is compensated and the sling force is stable.
8 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Fig. 5. Force analysis of LSM.

2.2. Dimensional synthesis

2.2.1. Static modelling

The initial working point of the LSM is the position where the compressed spring axis passes through the LSM rotation
axis. The constant force working range is ±15° on both sides of the initial working point. When the LSM is in a state of
disturbance, the force analysis is as illustrated in Fig. 5. The parameters in the figure required in the formula derivation
later are defined in the nomenclature table.
According to the geometric relationship, the following equations can be obtained:

l (q ) = r 2 + e2 − 2er cos(q ), (22)

l (q ) e
= . (23)
sin(q ) sin(α (q ))
The energy of the compressed spring assemblies is
1 2
Vspr = k[l0 + e − r − lst − l (q )] . (24)
2
The total elastic potential of the LSM is
1  1 2
VLSM = Vtor + Vspr = T (q + qst )2 + k[l0 + e − r − lst − l (q )] (25)
2 2
The force function of LSM can be obtained by deriving Eq. (25).
TLSM dVLSM 1 T kr
FLSM = = = (q + qst ) − [l0 + e − r − lst − l (q )] sin(α (q )). (26)
R dq R R R
The ideal constant required force Fst satisfies the relationship
Fst R = T  qst . (27)
The energy of the ACFS includes the elastic potential energy VLSM of the LSM and the work performed by the load force
of the LSM, which equals the required force Fst of the ACFS. Therefore, the total energy equation of ACFS is
1  2 1  2 1 2
VACFS = VLSM − Fst Rq = T q + T qst + k[l0 + e − r − lst − l (q )] . (28)
2 2 2
The system energy of ACFS is symmetrical about the initial position (q = 0), so its stable condition is that the system
energy has the smallest value at the initial position, that is, the second derivative of the system energy is greater than 0.
The first derivative of the system energy is the force error function, and the second derivative is the equivalent stiffness
function. The force error function obtained by deriving Eq. (28) is
TEACFS dVACFS 1 T kr
FEACFS = = = q − [l0 + e − r − lst − l (q )] sin(α (q )). (29)
R dq R R R
The equivalent stiffness function obtained by deriving Eq. (29) is
dFEACFS T kre kr e cos(q )l (q ) − er sin(q ) sin(α (q ))
KACFS = = + cos(q ) − (l0 + e − r − lst ) . (30)
dq R R R l 2 (q )
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 9

Fig. 6. Length-stiffness curve of the torsion bar.

2.2.2. Stability constraint on LSM

The first two terms in Eq. (30) are positive values, while the third term is negative. When KACFS = 0, the LSM is a zero
stiffness and critical stabilisation mechanism. For the ACFS, the absolutely stable condition at the initial position (q = 0)
isKACFS > 0. Therefore, the initial installation length of the compressed spring needs to be

T l 2 (q )
lst > (l0 + e − r ) − + e cos(q ) . (31)
kr e cos(q )l (q ) − er sin(q ) sin(α (q ))

The right side of Eq. (31) takes the maximum value at q = 0, so it can be converted into
 
T e−r
lst > (l0 + e − r ) − +e . (32)
kr e

Eq. (32) is the condition that the ACFS can be absolutely stable, that is, the installation length of the compressed spring
is larger than a specific value. When the installation length is longer, the equivalent stiffness is larger, and the stability and
robustness of the system are better.
The relationship between system characteristics and equivalent stiffness is as follows:

(1) The smaller the equivalent stiffness, the smaller force error the LSM produce in responding to the payload impact, and
the better the system’s constant force performance. However, the buffer stroke has to be larger and the buffering ability
to large impact is poorer. This situation is suitable to applications with high force accuracy but slower movement and
weaker payload impact.
(2) The greater the equivalent stiffness, the larger the LSM’s force error under the same displacement, and the worse the
system’s constant force performance. But the buffering ability is stronger. This situation is suitable for applications where
there are strong impacts between load and external environment.

2.2.3. Torsion bar dimensional synthesis

It can be concluded from the Eq. (21) that the smaller the stiffness of the torsion bar, the smaller the energy of the re-
maining elastic elements, and the smaller the size. However, there are still other constraints on the stiffness of the torsional
bar. The derivation is given below. The equation for the torsion bar diameter is

16(T  qH + QC )
d≥ 3
. (33)
π τP
When the torsion bar diameter is the minimum, the length is the smallest, and the following equation is satisfied.

Gπ d 4 Gπ 4 16(T  qH + QC )
Ltor = = 3 (34)
32T  32T  π τP
There is also a correspondence between the minimum length of the torsion bar and the torsion bar stiffness. According
to Eq. (34), the length-stiffness curve of the torsion bar is plotted in Fig. 6 using QC and qH in Table 1.
It can be seen from Fig. 6 that the smaller the torsional stiffness, the longer the torsion bar. The length increases sharply
when the torsion bar stiffness is small than a certain value (5 N m/°). According to the design requirements, the torsion bar
has the available maximum length max (Ltor ).Therefore, the torsion bar stiffness needs to meet the following equation.

T  ≥ T  (max (Ltor ) ) (35)

In addition, the rated load is half of the design load, and the torque of the torsion bar is still positive when the mecha-
nism is rotated to the negative maximum angle, that is, the torque of torsion bar does not reverse during the entire working
10 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

stroke. The torsion bar torque needs to meet the following equation.
QC
T ≤ (36)
2qH
According to comprehensive analysis, the stiffness of the torsion bar still needs to satisfy the Eqs. (35) and (36) when
the value is selected as small as possible.

2.2.4. Compressed spring dimensional synthesis

The compressed spring assemblies of the mechanism are crank-rocker mechanisms whose kinematics relationship is
described by Eqs. (22) and (23). From the point of view of kinematics synthesis, the question is how to determine the size
parameters of the mechanism, namely e and r. The force error function of ACFS is shown in Eq. (29). It can be seen that
Eq. (29) has a nonlinearity, we shall now seek to linearize it. The maximum value of 1-cos q is 0.03407 within ±15°, so
l (q ) ≈ e − rwithin ±15°. Therefore, the error function Eq. (29) is the difference between an approximate sine function and a
linear function. Sin q is close to q in a large range, so the smaller α is, the closer to a linear function the second part in Eq.
(29) is. From Fig. 5 it can be concluded that α equals q plus the angle between the axis of the compressed spring and the
e-axis. The larger e is and the smaller r is, the smaller the angle between the axis of the compressed spring and the e-axis
is and the closer α is to q. Therefore, in terms of the dimensions of the components, e should be as large as possible and r
as small as possible.
While r/e → 0, the error function can be simplified as:
 
T  − kr (l0 − lst ) T  − r Fpre
FEACFS ≈ q= q (37)
R R
From Eq. (37), it can be seen that when the product of the preload of the compressed spring Fpre and r is equal to
the stiffness of the torsion bar, the error is minimal and the LSM is a zero-stiffness mechanism. When the preload of the
compressed spring is reduced, the LSM exhibits a positive stiffness, that is, the Eq. (37) exhibits a monotonically increasing
linear function characteristic regarding to q. This condition is the basis for realising active control of the ACFS. While r/e → 0,
the linearity of the force curve is better, and the linear dynamics modelling performed in subsequent chapters is more
accurate, so the control effect is better.
It is the preload of the compressed spring that determines the magnitude of the compensation force generated by the
compressed spring assemblies. Therefore, the stiffness of the compressed spring can be small, as long as the compression
amount is sufficient. Because low-stiffness compressed springs have less stiffness difference between design and manu-
facturing, LSM tends to choose compressed springs with lower stiffness. However, this will cause the problem of longer
compressed spring lengths and increase LSM radial dimensions. Therefore, the selection of the compressed spring stiffness
needs to be taken into consideration, and it is necessary to refer to the standard component size list and select a relatively
suitable compressed spring size for the actual LSM.

2.2.5. Simulation of the parameter influences on LSM

The LSM is usually designed as a zero-stiffness mechanism, which requires small force error and small energy. After the
design is completed, the LSM is adjusted to a low-stiffness mechanism in the actual ACFS for control needs and stability
requirements. From Eq. (37), it can be concluded that the approximate condition of the force error minimum (FEACFS ≈ 0) is

T  =kr (l0 − lst ). (38)

The energy equation of the compressed spring assembly is

1 T
2
1
Vspr = k(l0 − lst )2 = . (39)
2 2 kr 2
It can be concluded from Eq. (39) that the reduction of the spring energy Vspr can be obtained by reducing the stiffness
T of the torsion bar, increasing the stiffness k of the compressed spring and increasing the distance r between the inner
hinge point of the compressed spring and the axis of the LSM. The characteristics of LSM were calculated and plotted in
Fig. 7 by a single-variable method using different parameters. The effects of these parameters on the LSM energy and force
curves were compared to verify previous analysis.
It can be seen from Fig. 7 that the effect of reducing the stiffness T of the torsion bar on the force error is not obvious,
but the energy of the compressed spring assembly is significantly reduced, which is consistent with the conclusion of Eq.
(39). However, according to Eq. (34), it can be seen that reducing the torsion bar stiffness will significantly increase the
length of the torsion bar, so the torsion bar stiffness should not be too small, and the energy of the compressed spring
assembly cannot be reduced to a very low level. Increasing the stiffness k of the compressed spring will reduce the energy
of the compressed spring, but it will cause fluctuations in the force curve and increase the force error. In addition, the
size of the compressed spring will rise, so the stiffness should not be too large. Increasing r also reduces the energy of the
compressed spring assembly, but it introduces an increase in the force error and an increase in the length of the compressed
spring installation, i.e. an increase in the radial dimension of the LSM. And the simplified process of Eq. (37) requires that
r is as small as possible, which will reduce the force error. For comprehensive analysis, the value of r should take a small
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 11

Fig. 7. Energy and force curves of the LSM under different parameters.

value, and the adjustment of other parameters is used to reduce the energy of the compressed spring. It can be concluded
that the energy of the compressed spring can be reduced by the above method and is consistent with Fig. 2. Due to the
limitations of the design requirements of the mechanism, the energy should be reduced to a reasonable value in the actual
design.

2.2.6. Design example

This section is an example briefly explaining how to determine the key parameters of the LSM according to the required
performance.

(1) Determine the required constant force (10 0 0 N) and the overall dimensions of the LSM: axial dimensions (500 mm) and
radial dimensions (240 mm). The axial dimensions determine the length of the torsion bar (400 mm). The radial dimen-
sion (240 mm) determines the diameter of the reel (200 mm, slightly less than the radial dimension). The constant torque
of the LSM can be determined (100 N m) according to the constant force (10 0 0 N). The radial dimension also determines
the distance e (100 mm) from the outer joint of compressed spring to the axis of LSM.
(2) Determine the rotating working stroke of the LSM (±15°). One half of the design load is taken as the rated load (50 N m).
At the same time, the torsional stiffness (2.52 N m/°) of the torsion bar is determined according to the length of the
torsion bar (400 mm) and Eq. (34).
(3) Determine the torsion bar diameter (9.4 mm) using Eq. (33) according to the torsional stiffness (2.52 N m/°) and the
shear modulus of the material.
12 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Fig. 8. 3D model of LSM.

(4) Determine the number of compressed springs. Considering the axial length and removing the dimensions of the reel and
adjustment assembly, it was determined that the compressed spring assembly could be placed in two groups in the axial
direction, with four compressed springs radially arranged in each group.
(5) Determine the distance r from the inner joint of compressed springs to the axis of rotation. Consider the diameter of the
torsion bar and the diameter of the hollow input shaft to determine r (30 mm). The smaller the value of r, the better the
constant force performance.
(6) Determine the specific parameters of the compressed springs. Fpre can be obtained when taking FEACFS according to Eq.
(37). The stiffness of the compressed spring is selected as small to reduce the diameter. In addition, the working load
of the compressed spring is 20%∼80% of the maximum load to satisfy the strength requirements. According to preload
Fpre and default installation length (e–r) of compressed springs, spring design manual is used to select the standard size.
Finally, all parameters of the compressed spring (k = 18 N/mm,l0 = 100 mm) are determined.

Considering the strength and rigidity of each structural member of the mechanism, design the required force adjustment
mechanism, the equivalent stiffness adjustment mechanism and other structures.

2.3. Structural design of LSM

2.3.1. Overall structure

The 3D model of the LSM designed according to the constant force principle is illustrated in Fig. 8. The LSM consists of
eight groups of compressed springs with both ends hinged, a torsion bar, a reel, a required force adjusting mechanism of
worm and gear, a mechanical limit mechanism, and two rotary encoders.
The LSM has the size of 500 mm × 240 mm × 240 mm, which is smaller compared to existing mechanisms [16,17] in the
application of low gravity suspension equipment. In the design of the LSM, the following works were carried out to reduce
the overall size. First, a torsion bar with high energy density was used to provide main output torque and it was placed on
the axis of a hollow input shaft to decrease the overall size. Second, radial springs are arranged in 2 groups to reduce the
individual radial length. Third, a torque motor is used to achieve a large range of constant-force output, avoiding the problem
of increasing the overall size caused by a large-range constant-force output design. The LSM still can be miniaturised in the
future research and the following methods can be used. First, profiled springs with negative stiffness can be designed to
replace compressed springs to reduce the radial dimension. Second, the torsion bar with non-linear torsional stiffness can
be designed, or a combination of multiple torsion bars can be used to achieve segmented torsional stiffness [30]. It has a
small torsional stiffness near the initial position to reduce the compressed spring load and the radial dimension while a
large torsional stiffness in the part far from the initial position to shorten the axial length of the torsion bar and the LSM.

2.3.2. Required force adjustment mechanism

From Eq. (26), the sling force expression at the initial working point, which is also the LSM required force, can be ob-
tained as

Fst = T  qst /R. (40)

Eq. (40) indicates that the required force is related to the torsional stiffness and initial torsion bar angle, as well as the
reel radius. Therefore, the LSM required force can be conveniently adjusted by changing the initial torsion bar angle. In the
LSM, a worm and gear mechanism is arranged on one side of the torsion bar. The worm is driven by a gear to rotate one
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 13

Fig. 9. Required force adjustment mechanism.

Fig. 10. Force curves under different initial torsion bar angle.

end of the torsion bar relative to the other end fixed to the output frame, so that the initial torsion bar angle, as well as the
required force of the LSM, can be adjusted, as shown in Fig. 9. Adjustment of the required force is performed prior to the
closed-loop experiments; therefore, Fst is a constant value during the experimental process. A change in the required force
will add an upward and downward translation to the total force curve; however, it has no influence on its fluctuation error
and does not change the constant force accuracy and the equivalent LSM stiffness. This is shown in Fig. 10, which shows
the five groups of force curves for different torsion bar initial installation angles.
This adjustment method is a stepless adjustment method, which does not lose the precision of constant force. Xiang et
al. [17] adjusted the required force by changing the preload of the spring. This adjustment would affect the constant force
accuracy. The force error is small near the design working force but large when the force is far from the rated value. Chen
et al. [23] adjusted the force by adjusting the preload position of the profiled spring. The adjustment of this mechanism
has little effect on the force accuracy in the constant-force section. Therefore, it is a better adjustment method applied
in the linear force adjustment mechanisms. From the perspective of constant-force accuracy, the adjustment process of
required force should not have influence on the constant force error. It should just shift the constant-force curve upwards
and downwards like the adjustment scheme in this article, as shown in Fig. 10. When the adjustment process has a coupling
relationship with the shape of the curve, the constant force performance at different required force is distinct. So there will
be an optimal working force range, which will affect the load adaptability of the mechanism.

2.3.3. Equivalent stiffness adjustment mechanism

The equivalent stiffness is the slope of the LSM force curve. Eq. (30) gives the expression of the equivalent stiffness
KACFS . The equivalent stiffness KACFS can be adjusted by adjusting the installation lengths lst of compressed springs. When
lst becomes larger, the equivalent stiffness becomes larger and vice versa. An adjustment example of the installation length
of the compressed spring and the equivalent stiffness adjustment mechanism are shown in Fig. 11. The upper turning block
can be screwed to realize the up and down movement. Thus, the LSM can demonstrate positive equivalent stiffness, negative
equivalent stiffness, and even zero stiffness (the optimal condition) at the initial working point. The LSM force curves with
different installed lengths of the compressed springs are illustrated in Fig. 12. The equivalent stiffness KACFS of the LSM is
the slope of the force curve indicated in Fig. 12 in the vicinity of q=0.
From the perspective of the ACFS, the mechanism should be in a stable equilibrium. That is, it should exhibit a tendency
to restore the initial equilibrium state after the disturbance disappears. The equivalent stiffness is 0, corresponding to the
14 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Fig. 11. Equivalent stiffness adjustment mechanism.

Fig. 12. LSM force curves with different preloads of compressed springs.

condition whereby the input and output terminals of the LSM are in a fully dynamic isolated state. Therefore, the system
cannot be restored to the initial working point after being disturbed. A lower equivalent stiffness results in a higher output
force accuracy. Therefore, the curve should be adjusted to exhibit a small positive equivalent stiffness (KACFS > 0) indicated
by the green curve in Fig. 12 to design the control system. This conclusion is also provided in Section 3.2 to prove this
theory. The blue curve (KACFS < 0) is used in the LSM design to satisfy the minimum force error throughout the entire
stroke.
There is a minimum equivalent stiffness for this LSM. A smaller equivalent stiffness result in higher constant force accu-
racy, but it will also result in a larger buffer stroke when dealing with external disturbances. Due to the limited cushion-
ing stroke, an unreasonable small equivalent stiffness will reduce the disturbance rejection of the LSM. There is therefore
a suitable minimum equivalent stiffness. Considering the application of low-gravity simulation, the rotation angle of the
LSM should be less than 5° within a sling force variation of 10 N, so the minimum equivalent torsional stiffness should be
0.2 N m/°.

2.3.4. Effects of prototype manufacturing on LSM used in ACFS

There is a deviation between the actual LSM and designed LSM. The deviation results from the following three aspects:
first, material properties: the different between the design stiffness and the actual stiffness of compressed springs and
torsion bar due to the material composition, manufacturing tolerances and the technology for heating processing; second,
manufacturing tolerances: backlash and friction caused by manufacturing tolerances in each kinematic pair; third, assembly
tolerances: The rotary damping of revolute pairs and the linear damping of guide rods caused by the positional error and
coaxial error during the assembly process.
In response to these three aspects, we have done the following work on the design of LSM and ACFS to reduce their
influences. First, the stiffness error of the torsion bar is allowed fluctuate within a limited range. The rated load of LSM is
500 N while the design load is 1000 N, so the changes of stiffness and load capacity of the torsion bar have little effect on the
rated load. There is a correspondence between the stiffness of the torsion bar and the compressed springs. By adjusting the
installed lengths of compressed springs, the torsion bar with different stiffness can be adapted to ensure that the constant-
force characteristics of the LSM still meet the requirements. Second, the effects of friction and backlash can be suppressed by
the control algorithm using a non-model-based PID controller. Third, many positioning structures are set during the design
process to ensure the position accuracy. The main rotary shaft holes are machined after assembly. Each axis of rotation is
provided with a bearing and the compressed spring guide bar is provided with a sliding bearing pair to reduce the effect of
friction.
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 15

Fig. 13. Prototype of ACFS.

Fig. 14. Overall scheme of electrical system.

3. Establishment of ACFS

3.1. Prototype overall design

The ACFS constructed based on the LSM is illustrated in Fig. 13. It mainly comprises the shafting line, mechanical frame,
and electrical system. The shafting line consists of the motor, brake, reducer, coupling, and LSM. The motor provides the
power for the entire system, while the brake locks the shaft, and the payload is maintained in the current position once
the experiment is completed. The LSM reel is intertwined with the sling on which the force sensor is placed. The frame is
constructed from rods and ball hinges. The vertical payload movement varies between 0–1.5 m, and the maximum acceptable
mass is 50 kg. The electrical system provides the motor power supply circuit and the hardware platform of the real-time
control system.

3.2. Real-time electrical platform

The real-time ACFS platform is based on the host and target computers. The target computer provides a real-time ex-
ecutable environment for the program designed by the host computer, and the platform is implemented by the MATLAB
Simulink Real-Time module. The overall scheme of the electrical system is illustrated in Fig. 14. MATLAB is installed on the
host computer, and the target computer has NI’s PCI-6229 data acquisition card and Softing’s CAN card installed. The PCI-
6229 is used to receive feedback signals from the encoders and brake, and the CAN card is used to receive the position,
speed, current motor information, and the force value of the force sensor.

4. Dynamic modelling and analysis

4.1. Dynamic modelling

A schematic diagram of the ACFS is shown in Fig. 15. From right to left are the following: motor, brake, reducer, cou-
pling, LSM, and sling. The LSM is divided into two parts: the input and output terminals. The input terminal is directly
16 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Fig. 15. Schematic of ACFS.

Fig. 16. LSM force curve and its fitting curve.

connected to the reducer output shaft, while the output terminal is connected to the reel. The variables used in Fig. 15 and
the following equations are defined in the nomenclature table. The equivalent stiffness of the LSM is positive (KACFS > 0).
(1) Establishment of dynamic differential equation
According to the design parameters, the theoretical force curve of LSM under an equivalent stiffness is plotted as shown
in Fig. 16. The curve in this figure is fitted and equivalent for later dynamic modelling.
To simplify this model, MATLAB is used to fit FLSM . The equation is obtained by fitting the curve with the cubic
polynomial
FLSM = 0.01q3 − 1.6 × 10−15 q2 + 0.45q + 500. (41)
In Eq. (41), the third term and quadratic coefficient are small, and FS can be simplified as a function near the initial
working point (q = 0) as
FLSM = 0.45q + 500. (42)
Eq. (42) is changed into the form of a general symbolic variable as
FLSM = KACFS q + Fst = TLSM /R. (43)
The simplification of Eq. (43) as a linear function works only for ±5°. The working stroke of ±5° is sufficient. Under the
control of the closed loop, the rotation angle between the input and output terminals of the LSM is very small.
The dynamic model in a differential equation form is established based on Newton’s second law for ACFS.
The expression of J1 is
J1 = (Jmotor + Jbrake )i2 + Jreducer + Jcoupling + JIn . (44)
The friction in this ACFS is mainly Coulomb-viscous friction, and the expression is
f = μv + Fc sgn(v ), (45)

+1 v > 0
sgn(v )= 0 v = 0 . (46)
−1 v < 0
In Eqs. (45) and (46), the Coulomb friction in the system is a typical nonlinear link which can be suppressed by the
closed-loop control system. The Coulomb friction effect is neglected when the equations are deduced, and f is simplified to
viscous friction only.
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 17

The system differential equations can be obtained as follows.

Motor drive part:

τ = J1 θ̈1 + μ1 θ˙ 1 + TLSM +μ2 (θ˙ 1 − θ˙ 2 ). (47)

LSM part:

TLSM +μ2 (θ˙ 1 − θ˙ 2 )=J2 θ̈2 + μ3 θ˙ 2 + F R. (48)

The differential equation of the ACFS is obtained by combining Eqs. (47) and (48):

τ = J1 θ̈1 + J2 θ̈2 + μ1 θ˙ 1 + μ3 θ˙ 2 + F R. (49)

It can be observed from Eq. (49) that μ2 is the internal friction and is not reflected in the entire equation, as well as
TLSM , which belongs to the system internal force.

(2) Derivation of system transfer function

The transfer function is deduced according to the system differential equation established above and the LSM force equa-
tion.
The rotation angle θ 2 of the LSM output terminal reflects the payload displacement change in the vertical direction. For
the low-gravity planet rover simulation, θ 2 reflects the up and down information of the ground. Treating the motor rotation
angle θ 1 and mechanism output angle θ 2 as the system input while the system output is the sling force, the derivation of
its transfer function is as follows.
Substitute Eqs. (43) into Eq. (48):

KACFS R(θ1 − θ2 ) + Fst R+μ2 (θ˙ 1 − θ˙ 2 )=J2 θ̈2 + μ3 θ˙ 2 + F R. (50)

The F in Eq. (50) is

F = Fst + F . (51)
The motor is operated in speed mode. The relationship between the speed command sent to the motor and the motor
output angle is
θ1 ( s ) Km
Gm ( s ) = = . (52)
θ˙ 1a (s ) Tm2 s2 + 2ξm Tm s + 1
In Eq. (52), Km , Tm , and ξ m are motor parameters.
From Eqs. (50) to (52), F (s ) can be obtained as

Km (μ2 s + KR ) J2 s2 + (μ2 + μ3 )s + KACFS R

F (s ) =   θ˙ 1a (s ) − θ2 ( s ) . (53)
R Tm2 s2 + 2ξm Tm s + 1 R

Eq. (53) is the ACFS transfer function. The sling force error arises from two parts: the motor speed command θ˙ 1a (s ) and
payload displacement θ 2 . The former transfer function is recorded as G1 (s) and the latter as G2 (s).

4.2. Analysis of dynamic characteristics

The system dynamic model was established as discussed in Section 4.1. In this section, the dynamic characteristics ex-
pressed by the transfer functions G1 (s) and G2 (s) are analysed.

4.2.1. Dynamic characteristics between motor speed and sling force

The transfer function between the motor speed command and force error is G1 (s). The equation of Gm (s) indicates the
characteristic of the integral link at a low frequency and the addition of an inertia lag in the high-frequency region. The
μ2 value is small; therefore, G1 (s) demonstrates the integral characteristic at a low frequency.

4.2.2. Dynamic characteristics between output angle and sling force

When the ACFS is subjected to high frequency disturbances, the motor response lags; therefore, the influence of the mo-
tor angle on the sling force becomes hysteretic. Considering an extreme situation where the motor is completely unrespon-
sive, the speed θ˙ 1 =0 and the angle θ1 =θ1−const . For the sling force, θ 1 is a fixed disturbance; therefore, we mainly analyse
the influence of the output angle θ 2 (s) on the sling force. Separating G2 (s) from Eq. (53), the deformation is expressed as
R
θ2 ( s ) = − F (s ). (54)
J2 s2 + (μ2 + μ3 )s + KACFS R
(1) Self-recovery characteristic
18 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

The LSM sling is excited with a unit pulse force signal to deduce the response under the time domain. The Laplace
transform of the unit impulse force signal is
F (s ) = L[δ (t )] = 1. (55)
Substituting Eq. (55) into Eq. (54), the result is expressed in a standard second-order vibration form as
KACFS R
1
θ2 ( s ) = − J2
( μ2 + μ3 )
, (56)
KACFS s2 + s + KACFS R
J2 J2

μ +μ
where ωn = the natural frequency and ξ = √ 2 3
KACFS R
J2 is 2 KACFS RJ2
is the system damping ratio. Thus, the constant force curve
must have a positive stiffness. That is, the green curve in Fig. 7. The ACFS is an under-damped system because the damping
ratio is less than 1. The solution to Eq. (56) is
ω 
θ2 (t ) = − n e−ξ ωn t sin ωn 1 − ξ 2 t, (57)
KACFS 1−ξ 2

where θ 2 has a value of 0 at the initial working point. When the system is disturbed, the amplitude of the second-order
oscillating system is attenuated according to Eq. (57). Therefore, it satisfies
lim θ2 (t ) = 0 (58)
t→∞

This indicates that the LSM has the ability to automatically recover to the initial working point, while the force is also
restored to the required force.
(1) Oscillation characteristic
Eq. (56) demonstrates that the natural frequency and damping ratio of the system are related to the equivalent stiffness
K of the LSM, the system moment of inertia, and the viscous friction coefficient. Once the design and machining of the
LSM are completed, the moment of inertia and viscous friction coefficient are determined. The equivalent stiffness of the
constant force curve is the only adjustable parameter, and the required natural frequency and damping ratio can be obtained
by adjusting the equivalent stiffnessKACFS .  
The natural frequency is proportional to KACFS , while the damping ratio is inversely proportional to KACFS . The system
is stable at a value θ2−stable with a step excitation signal input.
1
θ2−stable = − (59)
KACFS
A reduction in the equivalent stiffness KACFS will reduce ωn , increase ξ , increase the rise time tr , increase the peak
time tp , stabilise the adjustment time ts , decrease the overshoot MP , and decrease the oscillation timeN. From a control
engineering point of view, a reduction in the equivalent stiffness of the LSM will lead to a slower response of the second-
order oscillation system, improvement in the oscillation performance, and a larger range after stabilisation. The system
response speed and oscillation characteristics can only be optimised by selecting appropriate ξ and ωn values and then
determining the appropriate equivalent stiffness value. As the constant force range of the LSM is only ±15°, the KACFS value
selected should not be extremely small.
From the above analysis, it can be concluded that in the case of positive stiffness, the LSM can automatically recover the
equilibrium state and exhibit second-order vibration characteristics. Adjusting the equivalent stiffness KACFS can alter the
system performance. The equivalent stiffness KACFS cannot be taken as 0 but should be a reasonable value larger than 0.

5. Control system

5.1. Control strategy

The control target of the ACFS is to stabilise the sling force within the allowable error range. The direct force control
mode consists of using the sling force measured by the force sensor to construct the closed loop. Only the transfer function
with low-frequency characteristics was established in the ACFS dynamic modelling; therefore, the controller uses a model-
independent proportional-derivative (PD) controller with a correction link to achieve a constant force control.
The motor driver includes three modes of torque, speed, and position. From a force control point of view, the motor
rotation angle and speed in the torque loop are not controlled. There is no guarantee that the angle difference between the
LSM input and output is always approximately 0, which leads to exceeding the best LSM working area following a distur-
bance; thus, the torque loop should not be used. The speed loop is faster than the position loop. Following motor tuning,
the friction between the motor shaft and LSM input terminal can be managed by the speed loop. After a comprehensive
analysis, the speed loop is selected as the inner loop of the closed loop.
From Eq. (36), the force error F (s ) is affected by the motor velocity command θ˙ 1a (s ) and the displacement of the
suspended payload θ 2 (s) (also the LSM output angle). In this study, θ 2 (s) is treated as an interference, and the influence of
the motor angle is mainly considered. The control block diagram of the direct force control is shown in Fig. 17.
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 19

Fig. 17. Closed-loop control block diagram based on direct force control.

Fig. 18. Bode plots of experiment and fitting function.

5.2. Transfer function identification of controlled object

Using an identification to obtain the controlled object transfer function forms the basis of the system bandwidth design.
The identified transfer function is the forward channel G1 (s) of the closed loop. The ACFS G1 (s) contains the characteristics of
the motor and LSM. The input signal is the control speed command θ˙ 1a (s ), while the output signal is the sling force errorF ,
and the LSM required force is 500 N. The identification process makes use of the sweep method, and a special sweep signal
is designed for the ACFS. The ACFS sweep signal frequency range is 0.1–15 Hz, and the sinusoidal signal at each frequency
point is scanned for a full 10 cycles. Following idling for 10 s, the next frequency is scanned. The swept sinusoidal signal has
a 90° phase difference, and the excitation signal amplitude increases correspondingly at the frequency points after 1 Hz.
The force curves are analysed by means of a Fourier transform, and the amplitude and phase of the linear components
are separated. A Bode graph is drawn using MATLAB after obtaining a series of amplitude and phase values of the linear
component. MATLAB’s invfreps function is used to fit the transfer function and draw the Bode diagram of the fitting transfer
function. The experimental Bode plot and that of the fitting function are plotted into the graph shown in Fig. 18.
The fitting transfer function is

4.031 × 105 s + 2.87 × 105

G (s ) = . (60)
s4 + 29.65s3 + 6233s2 + 4.372 × 104 s + 1912
It can be clearly observed that the fitting effect is improved and the deviation is small. Furthermore, it can be observed
from Eq. (60) that using one zero and four poles can achieve the most appropriate effect in fitting the actual system. How-
ever, the transfer function in Eq. (53) is in a structure with one zero and two poles. The reason for the difference of the two
poles is analysed in the following simulation.
There is a certain amount of error in the dynamic model established in Section 4.1. The ACFS model error mainly includes
two parts: the unmodelled model error and the model simplified error. The former is Coulomb friction, while the latter
includes the linearisation error of the LSM output force and the moment of inertia simplified error of the compressed
spring assemblies. The influence of the model error on the system characteristics can be obtained by the ACFS dynamic
simulation analysis using Adams.
Adams simulates the Coulomb friction by incorporating the friction caused by the reaction force. The entire LSM model
is placed into Adams so that the output force nonlinearity and the moment of inertia of the compressed spring assemblies
are included. The torsion bar is replaced by a torsion spring, and the viscous friction in the torsion and compressed springs
is set under the damping option.
20 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Fig. 19. Excitation velocity and sling force curves at different frequencies.

Fig. 20. Closed-loop block diagram.

The force response under the motor velocity excitation is analysed by fixing the LSM output terminal. The sinusoidal
excitation velocity signals at frequencies of 1 Hz, 10 Hz, and 30 Hz are added to the LSM input shaft. The excitation velocity
and sling force curves are illustrated in Fig. 19.
It can be observed from Fig. 19 that the phase lag of the force curve is 90° behind the velocity curve at 1 Hz and 10 Hz.
The low frequency is considered as an integral characteristic, as demonstrated in the analysis in Section 4.2.1. The nonlinear
effect caused by the Coulomb friction is more obvious at 10 Hz. The force curve is no longer in a sine form, but a sine with
clutters. Commutation of the Coulomb friction results in a sudden force change.
The hysteresis phase angle of the force curve reaches approximately 270° at 30 Hz. By changing the frequency of the
excitation speed signal, a significant phase lag occurs from approximately 90° to 270° at 16 Hz. If the frequency increases
sequentially, the mechanism will always lag by approximately 270°. The blocks at the end of the compressed springs cause
a phase lag after a certain frequency. After 16 Hz, the denominator of G1 (s) in Eq. (53) adds two poles to produce a phase
lag from 90° to 270°, which is the high-frequency modelling error. Within the system bandwidth of 3 Hz, the impact of
the high-frequency unmodelled characteristics is weak and can be ignored. The previous dynamic modelling and the fitting
transfer function of the actual system including four poles are reasonable.

5.3. Bandwidth design

The ACFS bandwidth reacts to the response speed and stability of the system. A higher bandwidth results in the system
being able to withstand a higher frequency disturbance. The bandwidth design in this study is based on the open-loop Bode
graph, and the bandwidth value ωBW is expressed by the shear frequency ωc of the graph.
The transfer function of the controlled object was obtained using Eq. (60) in Section 5.2. The forward path of the control
system shown in Fig. 20 consists of a PD controller, corrective link, and Eq. (60). The selection of the gain and corner
frequency of the inertia link imposes a certain constraint: both excessively large and small values are unacceptable. A closed-
loop block diagram is constructed and the Bode diagram for the open-loop transfer function is drawn using Simulink. The
desired amplitude–frequency characteristic curve is obtained by adjusting the PD and correction link parameters. The closed-
loop block diagram is shown in Fig. 20, and the Bode diagram for the open-loop transfer function following the bandwidth
design is illustrated in Fig. 21.
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 21

Fig. 21. Bode diagram for open-loop transfer function following the bandwidth design.

Fig. 22. Principle diagram of static constant force characteristic experiment.

The PD parameters following the design are

N 550
P D (s ) = P + D N
= 4.5 + 0.006 550
. (61)
1+ s
1+ s

The correction link is

0.01s + 1
Gc ( s ) = . (62)
0.8s + 1
Fig. 21 illustrates that the ACFS bandwidth is 3 Hz following the bandwidth design, while the amplitude margin is 11 dB
and the phase margin is 30°.

6. Experimental analysis

6.1. Static constant force characteristic experiment

The main purpose of the static constant force experiment is to test the LSM constant force performance and verify
the theoretical design accuracy. The principle diagram for the experiment is illustrated in Fig. 22. The brake, reducer, and
coupling between the motor and input terminal are not shown. The end of the sling is fixed first, and then the motor is
used to rotate the input terminal without closed-loop control. The change in the sling force is simultaneously recorded by
the force sensor. The experimental force curve and an enlarged view of a part of the curve are shown in Fig. 23.
It can be observed from Fig. 23 that the constant force characteristic of the LSM is effective. The sling is constantly
tightened, and the force increases sharply in the force loading area where the LSM has not yet entered the work space.
Fig. 23 shows that the force is 500 ± 3 N when the ACFS enters the work area. Compared to the theoretical LSM, the real
LSM exhibits a smaller force error for the following reasons. First, the stiffness of the compressed springs and torsion bar
22 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Fig. 23. Force curves of static constant force characteristic experiment.

Fig. 24. Experimental force curve under direct force control method.

is inconsistent with the theoretical design, and the initial deformation stiffness is small. Therefore, the real LSM force curve
is obtained by adjusting the initial angle of the torsion bar and the preload of the compressed springs. The tendencies of
the two curves are the same, but the error sizes differ. Second, the force curve fluctuation is caused by the compressed
spring assemblies. The friction direction is opposite to the output force direction of the compressed spring assemblies. The
frictional effect weakens the output force of the compressed spring assemblies, resulting in a smaller force error.

6.2. Dynamic constant force characteristic experiments

The dynamic constant force experiments are the key to testing the steady-state force accuracy and anti-disturbance per-
formance of the ACFS under closed-loop control. The results of different excitation signals can be obtained by an excitation
source for which the output motion can be changed randomly. The ACFS required force can be adjusted within the range of
0–550 N. As the rated design force is 500 N, the experimental force is also selected as 500 N.

6.2.1. Steady-state force accuracy experiment

To test the ACFS steady-state force error and verify the accuracy of the control algorithm design, the end of the sling is
fixed as shown in Fig. 16, and the ACFS with closed-loop control enters the steady state. The steady-state force error is 0.5 N
as shown in the force curve of Fig. 24.

6.2.2. Sinusoidal excitation experiment

The sinusoidal excitation experiment is the key to testing the ACFS constant force performance in response to an alter-
nating load. The principle diagram is illustrated in Fig. 25(a). The force response curves can be obtained using the input of
the sinusoidal excitation signals at different amplitudes and frequencies produced by the adjustable excitation source motor.
The prototype is illustrated in Fig. 25(b). The motion excitation for the sling with a large force requires a motor with a large
torque output capability. Therefore, a Kollmorgen motor CH053A with a 20 N m torque output capability is selected as the
excitation source motor. The output shaft diameter is 30 mm and the sling diameter is 3 mm; therefore, the maximum force
that the excitation source can carry is 1212 N, which meets the experimental requirements.
Sinusoidal velocity excitation signals at frequencies of 0.5–5 Hz and amplitudes of 3.5–17.3 mm/s are applied to the ACFS.
The force response curve for the sinusoidal velocity excitation signal at a frequency of 0.5 Hz and amplitude of 17.3 mm/s is
shown in Fig. 26.
The dynamic response of the system is obtained by analysing the force response curves. The force error data for each
frequency and amplitude are extracted to draw the curves illustrated in Fig. 27. Following the frequency analysis of the
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 23

Fig. 25. Principle and prototype diagrams of sinusoidal excitation experiment.

Fig. 26. Sinusoidal velocity excitation signal and resulting force response curve (0.5 Hz, 17.3 mm/s).

Fig. 27. Force error curves under sinusoidal excitation.

excitation speeds and response forces, the phase angle data for each frequency and amplitude are obtained, and the curves
are plotted as shown in Fig. 28. The following conclusions can be drawn by analysing the curves in Figs. 27 and 28.
(1) Amplitude frequency characteristics
It can be observed from Fig. 27 that the force error initially increases and then decreases with an increase in frequency
when the excitation remains at the same amplitude. The transfer function between the excitation velocity signal and output
force error is the sensitivity function. The ACFS can withstand a sinusoidal excitation signal with a frequency of 2 Hz and
amplitude of 17 mm/s under a maximum dynamic force error of 20 N.
(2) Phase frequency characteristics
24 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

Fig. 28. Phase hysteretic angle curves under sinusoidal excitation.

Fig. 29. Force and angle difference curves in sinusoidal excitations of different amplitudes and frequencies.

It can be observed from Fig. 28 that the system enters the damped zone after 3 Hz and the phase angle lags and becomes
greater than 180° under a small amplitude excitation. With a large amplitude excitation, the system enters the attenuation
zone from 2 Hz, after which the motor hysteresis becomes severe and the disturbance suppression capability deteriorates.

(3) Comprehensive comparison

A greater excitation amplitude at the same frequency will result in a greater force error in the low-frequency band. After
entering the high-frequency area, the motor tends to be unresponsive at a greater amplitude excitation, and the LSM plays
the role of suppressing the disturbance. A higher frequency results in a smaller excitation displacement; thus, the force error
is relatively reduced.
The ACFS force error mainly arises from the adjustment effect of the LSM when the motor is in a hysteretic state. As
shown in Fig. 29, the force error is positively correlated with the angle difference between the LSM input and output termi-
nals. Once again, it can be observed that the LSM plays the role of suppressing the disturbance.

6.3. Discussion

The performance of the constant-force systems can be evaluated in two aspects: disturbance rejection and constant force
accuracy. Figs. 27 and 28 illustrate the performance of the proposed ACFS. Detailed description is as follows.

(1) Disturbance rejection: The disturbance rejection is evaluated by the bandwidth of the ACFS. The wider the bandwidth,
the wider range of disturbances frequency the system can suppress, and the smaller the system’s dynamic force error.
The designed bandwidth of the proposed ACFS is 3 Hz, and it can still have excellent anti-disturbance ability when the
disturbance frequency reaches 5 Hz. In comparison, the stable bandwidth achieved in [17], which is also a research of
our laboratory, is 1 Hz. The motion frequency of planet rovers and patients in rehabilitation training is generally around
1 Hz, so it is inferred that the systems in [15,16] can perform stable gravity compensation within 1 Hz.
(2) Constant force accuracy: The force error of the passive counterweight system results from the inertial force of the
weight, so the error is large. The force accuracy of a passive elastic system depends on the stiffness and length of the
spring, and its force error is also large. The comparison experiments of these two systems were carried out in [15],
and the force errors are respectively about 2 times and 10 times of that of an active constant force system. For slow
moving objects, such as planet rovers [4], it is relatively easy to achieve high force accuracy (<1 N). The proposed ACFS
can achieve a force error of 0.5 N. Fast moving objects includes patients in rehabilitation training and astronauts in low
 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26 25

gravity simulation, the maximum force error of gravity compensation for patients in [15] is about 20 N, and the average
error is 10 N. The device in [10] can achieve high-precision gravity compensation under different required forces at slow
speeds, but its dynamic force accuracy might be limited by inner friction. In our previous research, we have performed a
1/6 G walking simulation on an existing constant force system [17]. The maximum force error in a walking state reaches
50 N (Walking frequency 1 Hz). The proposed ACFS has a force error less than 10 N at 1 Hz. Under 5 Hz excitation, it
still maintains a force error less than 20 N, while its load can reach 50 0N∼10 0 0 N. Overall, the performance of the ACFS
is satisfactory compared with current active constant force systems, and much better than the passive constant force
system.

The maximum force error occurs at a sinusoidal excitation with an excitation speed of 17 mm/s and a frequency of
2 Hz. The sling must complete two periods of reciprocating motion within 1 s. This condition is worse than that of a usual
actual application conditions, where the motion frequency of astronauts or spacecraft is usually less than 1 Hz. So it shall
be considered acceptable that the force error is 20 N in this harsh condition.
The factors contributing to dynamic force error are as follows:

(1) Amplitude and frequency of the excitation signal: The larger the amplitude of the excitation signal, the higher the
frequency, and the more complex the excitation situation, the more difficult it is for the ACFS to suppress disturbance,
and therefore the greater the error.
(2) Internal friction of the mechanism: The main components of the internal friction are Coulomb friction and viscous
friction. The Coulomb friction is mainly expressed as a constant force whose direction is opposite to the moving direction
of the payload.
(3) Controller performance: The servo period of the control algorithm determines the execution frequency of the closed-
loop control. The performance of the hardware determines the upper limit of the closed-loop frequency, thus restricting
the ability to suppress disturbances.

The recommended solution to the aforementioned factors are shown below:

(1) A better force sensor measures disturbance more accurately, thereby improving the accuracy of the feedback loop and
achieving higher control accuracy.
(2) The effects of friction can be suppressed by the control algorithm using a non-model-based PID controller. In addition,
many positioning structures are set during the design process to ensure the position accuracy. The main rotary shaft
holes are machined after assembly. Each axis of rotation is provided with a bearing and the compressed spring guide bar
is provided with a sliding bearing pair to reduce the effect of friction.
(3) A faster hardware platform can be used to achieve higher closed-loop frequency.

7. Conclusion

In this study, an approach of designing LSM from an energy perspective was proposed. Then a novel ACFS was devised
based on an LSM capable of outputting a large force with low and adjustable stiffness. A theoretical analysis proved that the
LSM could buffer disturbances and restore the initial state automatically once the disturbances were removed. The mecha-
nism without pulley offers the advantages of a small overall size, low stiffness, and large adjustable force range (0∼10 0 0 N).
A proportional-derivative controller with incomplete derivation was applied to the ACFS to achieve a wide constant-force
output range. Finally, the ACFS was experimentally demonstrated to achieve a high static and dynamic force accuracy and
excellent robustness. The compensative force in the experiments was 500 N, while the steady-state force error was 0.5 N.
When the sinusoidal motivation speed varied from 1 Hz to 5 Hz, the ACFS was stable and maintained a small force error
until the excitation amplitude reached 17 mm/s and the force error reached 20 N.
Future work will focus on a constant-force control algorithm for floating loads to simulate a zero-gravity environment.
The long-term goal is to establish an ACFS to simulate both low-gravity and zero-gravity environment for astronauts and
spacecraft. The ACFS can also be applied to rehabilitation training, cargo balance handling, and vibration isolation.

Acknowledgements

This study was supported by the National Natural Science Foundation of China (Grant No. 51405109), the Self-Planned
Task (No. SKLRS2014MS09) of the State Key Laboratory of Robotics and System (HIT), and the ‘111’ Project (B07018).

Supplementary materials

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.mechmachtheory.
2018.07.019.

References

[1] A. Kemurdjian, U.A. Khakhanov, Development of simulation means for a gravity forces, Robotics 20 0 0 (20 0 0) 220–225.
26 Z. Liu et al. / Mechanism and Machine Theory 130 (2018) 1–26

[2] H.A. Fujii, K. Uchiyama, H. Yoneoka, et al., Ground-based simulation of space manipulators using test bed with suspension system, J. Guid. Control
Dyn. 19 (5) (1996) 985–991.
[3] M.G. Kim, S. Cho, T.Q. Tran, et al., Scaled jump in gravity-reduced virtual environments, IEEE Trans. Vis. Comput. Graph. 23 (4) (2017) 1360–1368.
[4] Z. Liu, H.B. Gao, Z.Q. Deng, Design and implementation of a large-scale gravity compensation system for lunar rover, applied mechanics and materials,
Trans. Tech. Publ. 385 (2013) 759–767.
[5] H.B. Brown, J.M. Dolan, A novel gravity compensation system for space robots, in: Proceedings of the ASCE Specialty Conference On Robotics for
Challenging Environments, ASCE, Albuquerque, USA, 1994, pp. 250–258.
[6] G.C. White, Y. Xu, An active vertical-direction gravity compensation system, IEEE Trans. Instrum. Meas. 43 (6) (1994) 786–792.
[7] Y. Shibata, S. Imai, T. Nobutomo, et al., Development of body weight support gait training system using antagonistic bi-articular muscle model, in:
Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2010, pp. 4468–4471.
[8] J. Yoon, B. Novandy, C.H. Yoon, et al., A 6-DOF gait rehabilitation robot with upper and lower limb connections that allows walking velocity updates
on various terrains, IEEE/ASME Trans. Mechatron. 15 (2) (2010) 201–215.
[9] T.M. Griffin, N.A. Tolani, R. Kram, Walking in simulated reduced gravity: mechanical energy fluctuations and exchange, J. Appl. Physiol. 86 (1) (1999)
383–390.
[10] W. Xiu, K. Ruble, O. Ma, A reduced-gravity simulator for physically simulating human walking in microgravity or reduced-gravity environment, in:
Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2014, pp. 4837–4843.
[11] D.A. Kienholz, Simulation of the Zero-Gravity Environment For Dynamic Testing of Structures, Csa Engineering Inc, Palo Alto, CA, 1996.
[12] F. Gazzani, A. Fadda, M. Torre, et al., WARD: a pneumatic system for body weight relief in gait rehabilitation, IEEE Trans. Rehabil. Eng. 8 (4) (20 0 0)
506–513.
[13] M. Glauser, Z. Lin, P.E. Allaire, Modelling and control of a partial body weight support system: an output regulation approach, IEEE Trans. Control Syst.
Technol. 18 (2) (2010) 480–490.
[14] K. Gordon, B. Svendesen, S. Harkema, et al., Closed-loop force controlled body weight support system, U.S. Patent 7 381 163, Jun. 3, 2008.
[15] M. Frey, G. Colombo, M. Vaglio, et al., A novel mechatronic body weight support system, IEEE Trans. Neural Syst. Rehabil. Eng. 14 (3) (2006) 311–321.
[16] Z. Yang, Y. Sun, Y. Hao, et al., Active mass-offloading with cable-driven SEA for tailored support to lower limb rehabilitation, in: Proceedings of the
International Conference on Advanced Robotics and Mechatronics (ICARM), IEEE, 2016, pp. 202–206.
[17] S. Xiang, H. Gao, Z. Liu, et al., A novel active suspension gravity compensation system for physically simulating human walking in microgravity, in:
Proceedings of the IEEE International Conference on Robotics and Biomimetics (ROBIO), IEEE, 2016, pp. 1052–1057.
[18] P. Wang, Q. Xu, Design and modelling of constant-force mechanisms: a survey, Mech. Mach. Theory 119 (2018) 1–21.
[19] Y. Liu, Y. Zhang, Q. Xu, Design and control of a novel compliant constant-force gripper based on buckled fixed-guided beams, IEEE/ASME Trans.
Mechatron. 22 (1) (2017) 476–486.
[20] Y. Liu, Q. Xu, Design of a 3D-printed polymeric compliant constant-force buffering gripping mechanism, in: Proceedings of the IEEE International
Conference on Robotics and Automation (ICRA), IEEE, 2017, pp. 6706–6711.
[21] P. Wang, Q. Xu, Design of a flexure-based constant-force XY precision positioning stage, Mech. Mach. Theory 108 (2017) 1–13.
[22] C.W. Hou, C.C. Lan, Functional joint mechanisms with constant-torque outputs, Mech. Mach. Theory 62 (2013) 166–181.
[23] Y.H. Chen, C.C. Lan, An adjustable constant-force mechanism for adaptive end-effector operations, J. Mech. Des. 134 (3) (2012) 031005.
[24] Z.W. Yang, C.C. Lan, An adjustable gravity-balancing mechanism using planar extension and compression springs, Mech. Mach. Theory 92 (2015)
314–329.
[25] R. Barents, M. Schenk, W.D. van Dorsser, et al., Spring-to-spring balancing as energy-free adjustment method in gravity equilibrators, J. Mech. Des. 133
(6) (2011) 061010.
[26] Y. Liu, D.-J. Li, D.-P. Yu, et al., Design of a curved surface constant force mechanism, Mech. Based Des. Struct. Mach. 45 (2) (2017) 160–172.
[27] P. Lambert, J.L. Herder, An adjustable constant force mechanism using pin joints and springs, New Trends in Mechanism and Machine Science, Springer,
2017, pp. 453–461.
[28] D. Xu, Y. Zhang, J. Zhou, et al., On the analytical and experimental assessment of performance of a quasi-zero-stiffness isolator, J. Vib. Control 20 (15)
(2014) 2314–2325.
[29] G. Chen, S. Zhang, Fully-compliant statically-balanced mechanisms without prestressing assembly: concepts and case studies, Mech. Sci. 2 (2) (2011)
169–174.
[30] G. Radaelli, R. Buskermolen, R. Barents, et al., Static balancing of an inverted pendulum with prestressed torsion bars, Mech. Mach. Theory 108 (2017)
14–26.