2014 IEEE International Conference on Robotics & Automation (ICRA)

Hong Kong Convention and Exhibition Center

May 31 - June 7, 2014. Hong Kong, China

An Admittance Control Scheme for Haptic Interfaces Based on

Cable-Driven Parallel Mechanisms
Alexis Fortin-Côté, Philippe Cardou and Clément Gosselin

Abstract— This paper presents a cable-driven parallel mech-

anism as a haptic interface and its underlying control method.
This human-sized, three-degree-of-freedom mechanism has a
tetrahedral architecture, four cables and evolves in three-
dimensional space. A brief review of the kinematics of the
mechanism is presented. Also, an admittance control law
coupled with a closed-loop velocity controller is proposed. The
control method is then refined by introducing adaptations for
smooth surfaces and sharp edges. This control method is then
validated by experimental results. Furthermore, the geometry
of the mechanism is identified by a method that does not require
any other sensor than the motor encoders.

I. INTRODUCTION Handle
With virtual reality becoming more and more prominent
with CAVE-style environments and wearable commercial
devices like the occulus rift [1], arises a need to physically
interact with these virtual worlds. Some devices providing
force feedback and haptic rendering are already available to F/T
consumers, e.g., the PHANToM [2] and Falcon (by Novint sensor
inc). Both of these devices use rigid-link robots to transmit a Fig. 1: Photograph of the CDPM used in this work.
force to the user. Another kind of device is the Maglev 200
(by Butterfly Haptics, LLC), which, as its name implies, uses
magnetic fields to transmit haptic rendering to the user. All
these solutions work for their respective domains but suffer friction. More work, [7], [8], followed up on those initials
from a common drawback: their limited workspace. research, but did not concentrate either on the control law.
This is where a cable-driven parallel mechanism (CDPM) Nevertheless, we could infer that they relied on an impedance
as that illustrated in Fig. 1 can provide an advantage relative control scheme, meaning that the authors had to mitigate
to conventional mechanisms. CDPMs are known to possess the effects of friction to obtain a good rendering. Williams
a large workspace and a low inertia, two traits of a good [10] worked on a planar device and also used an impedance
haptic feedback mechanism. CDPMs are currently used in control law without a force sensor. Morizono et al. [11]
commercial applications like the Skycam [3] and as a control used an impedance scheme with force feedback on each
module for a radio telescope [4], for example. string to control the wrench applied by a sport simulating
The advantages mentioned above have also led researchers device. In this paper, we rely on an admittance control
to use them as haptic interfaces. Several research initiatives scheme, as it seems better suited to render contacts with hard
addressed the development of haptic CDPMs, beginning surfaces. Such an approach to haptic rendering is explained
as early as 1989 [5]. Most reports concentrate on the by Carignan and Cleary [12], but not for a CDPM. Gosselin
architecture and components of the CDPMs. Among the et al. [13] worked on an admittance control scheme for a
pioneering designs, we find the SPIDAR [6], a spatial haptic planar CDPM, but used closed-loop position feedback as
device that generates forces on the user’s finger, but little is opposed to velocity feedback, which is used here.
shown of the control laws that must have been used. Also, Furthermore, it is desired to take advantage of the re-
Gallina [9] presented the Feriba-3: a planar three-degree-of- configuration capabilities of CDPMs in order to adapt their
freedom haptic interface sliding on compressed air to limit workspace to different virtual environments. This requires
that the mechanism be able to determine its geometry effi-
*This work was supported by the Natural Sciences and Engineering ciently, i.e., that it calibrates itself. Miermeister and Pott [14]
Research Council of Canada (NSERC) and by the Canada Research Chair
program. expose a method of for the calibration of these devices with
The authors are with the Laboratoire de robotique, Département de génie force sensors. Borgstrom et al. [15] use a self-calibration
mécanique, Université Laval, Québec, Canada, G1V 0A6 method without force sensor to determine the initial lengths
[email protected] and
[email protected] and of the cables but not the position of the anchor points. This
[email protected] paper presents a method of calibration without force sensor

978-1-4799-3685-4/14/$31.00 ©2014 IEEE 819

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on November 12,2020 at 05:17:08 UTC from IEEE Xplore. Restrictions apply.
 and able to determine the initial lengths of the cables and
position of the anchor points. This self-calibration method, Am
which was briefly presented in [16], is further detailed in
Section III. Experimental results are presented to demon-
strate the effectiveness of this method. Rm ··
·
The mechanism studied, shown in Fig. 1, is a three- Ri
degree-of-freedom mechanism used for positioning a point
in 3D space. It consists of a rigid platform, for the user to P ri
ρi
manipulate, held by a set of four cables whose lengths are
controlled by reels that are affixed to a rigid frame. The Ai
platform shape is that of a regular tetrahedron, which allows R1
each cable to be attached at a different vertex. A face of the p
ai
tetrahedron is directed towards the user so as to increase ···
access to the handle and to limit interferences between
user and cables. The reels are also positioned to form a A1 O
regular tetrahedron having a similar shape and orientation
as the platform. This four-cable configuration allows point
positioning in three-dimensional space. The orientation of Fig. 2: Geometric model of a generic cable driven parallel
the moving platform is not restrained but remains fairly mechanism.
constant; the end-effector being much smaller than the fixed
frame. The platform itself consists of a handle mounted on
a force sensor, which is attached to the platform frame. The low at 1.5 g/m. Thus, the assumption of rigid and massless
handle is positioned at the centroid of the tetrahedron, which cables is considered reasonable.
prevents the forces applied by the user on the platform from Geometric constraints prescribe the length ρi of the ith
generating large moments. This centroid is the point where cables as
forces applied by the cables intersect when the platform is
ρ2i = (p − ai + ri )T (p − ai + ri ). (1)
centred in its workspace. The forces applied by the user are
measured using a force sensor and are used for control. A By differentiation and rearrangement of (1), the correspond-
measure of these forces is crucial, as it is the input to the ing velocity equation is written as
admittance control scheme. This method of control has been
ρi ρ̇i = (p − ai + ri )T ṗ. (2)
chosen over an impedance control because it is less sensitive
to friction and other perturbations. This leads, for m cables, to the following first-order kine-
This paper is structured as follows: the kinematics of a matic relationship
generic CDPM are presented in Section II; Its calibration ρ̇ = Jṗ (3)
is explained in Section III; the haptics control scheme is
detailed in Section IV; and finally, experimental results are where
(p − a1 + r1 )T
 
presented in sec. V.  
ρ̇1   ρ1
II. K INEMATICS  .. 
 
..
ρ̇ =  .  and J =  (4)
 
 .
Consider a rigid platform, which is to be manipulated by  
 (p − am + rm )T 
ρ̇m
the user, moving in three-dimensional space and constrained
by m cables attached by reels to the fixed frame, as shown ρm [m×3]
schematically in Fig. 2. The points of attachment to the is the Jacobian matrix, which will be used in the control
reels are defined by vectors ai with respect to reference scheme of Section IV.
point O, while the corresponding attachment points on the
platform are defined by vectors ri expressed relative to the III. C ALIBRATION
position of the platform reference point P . The position of The identification of the model’s parameters is crucial
the latter point is given by vector p. The orientation of the to its accuracy and, in turn, to the quality of the haptic
rigid platform is considered constant. Further experiments rendering. The parameters that need identification are the
proved that this simplification is acceptable for the proposed position of each fixed cable anchor point ai and the initial
application, as a change of 15◦ in orientation induces an error length of each cable ρi,0 due to the use of incremental
in the position p of less than 10 mm which is acceptable in a encoders. The position ri of the moving attachment points
context of haptic rendering where user cannot easily perceive are known accurately, since the end effector was built using
this error. It is important to note that cable elasticity and fast prototyping. Hence, these parameters are treated as
mass are considered negligible and are not accounted for constant in the calibration procedure.
in this model. Under the maximum tension, the cable axial Since a cable-driven parallel mechanism possesses a large
deformation is less than 0.5%, and the cable mass is very workspace, standard measurement methods are cumbersome

820

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on November 12,2020 at 05:17:08 UTC from IEEE Xplore. Restrictions apply.
 or give inaccurate results. The ideal calibration method over ai , ρi,0 , i = 1, . . . , m,
does not require any other sensor than those used for the
where
haptic control of the CDPM, namely, the motor encoders
and the force sensor. The calibration technique used for ui,j =k pj − ai + ri k22 −(∆ρi,j + ρi,0 )2 . (9)
this mechanism uses only the encoder measurements and is
The solution to this nonlinear least-square problem is ob-
performed in two steps. The first step is an approximation
tained via Matlab’s implementation of the interior-reflective
of the positions of Ai , i = 1, . . . , m, by reading the value
Newton method. This method is invoked by the lsqnonlin
of the encoders when the platform is located at each of the
function of the Matlab optimization toolbox.
anchoring points Ai . This method gives two estimates of the
An algebraic Jacobian H, not to be confused with the
distances between pairs of anchoring points, which can then
general Jacobian matrix of (3), can be calculated by differ-
be averaged to yield an initial estimate of their positions.
entiation of
These initial estimates are refined in step two.
u = [uT1 · · · uTn ]T , (10)
The second step of the calibration method consist in
exploiting the measurement redundancy of the mechanism. with
An over-determined system of equations can be generated uj = [u1,j · · · um,j ]T , (11)
if the assumption of rigid and massless cables is valid. To
and
this end, the platform is moved across the workspace by the T
v = aT1 · · · aTm pT1 · · · pTn

user with cables under constant tension, thus generating n ρ0,1 · · · ρ0,m , (12)
cable-length measurements for each of the m cables. For the m·n m (4m+3n)
where u ∈ R , uj ∈ R , and v ∈ R . The
j th calibration position, it is possible to rearrange (1) and
expression of this Jacobian matrix is then
replace ρi,j by
∂u1,1 T ∂u1,1 T
 
ρi,j = ∆ρi,j + ρi,0 , (5) ∂u1,1
 ∂a1 ··· ···
∂ρ0,1 ∂pn 
where ∆ρi,j is the measured length and ρi,0 is the initial ∂u 
.. .. .. .. ..

=H= ,
 
length of the cable at initialization of the encoder. It is then ∂v  . . . . . 
T T
possible to generate a system of nm equations, each of which  ∂un,m ∂un,m ∂un,m
can be written as ··· ···
∂a1 ∂ρ0,1 ∂pn
(13)
0 = (pj − ai + ri )T (pj − ai + ri ) − (∆ρi,j + ρi,0 )2 , (6) where the components can be written as
with i = 1, . . . , m, and j = 1, . . . , n, where m is the
(
∂ui,j −2pj + 2ai − 2ri if i = k,
number of cables and n is the number of positions used for =
∂ak 0 otherwise.
the calibration. Note that vectors ri are considered constant, (3
since they are small compared to the length of the cables and ∂ui,j −2(∆ρi,j + ρi,0 ) if i = k,
= (14)
since the orientation of the platform remains approximately ∂ρ0,k 0 otherwise.
( 3
constant. In order to avoid having to deal with an under ∂ui,j 2pj − 2ai + 2ri if j = k,
determined system of equations, the number of equations =
∂pk 03 otherwise.
should be greater than or equal to the number of unknowns.
Here, m equations are obtained for each position of the This Jacobian H is given as an argument to the Matlab
moving platform, for a total of mn. The unknowns appearing function to increase the calculation speed of the algorithm.
in the equations are ai , ρi,0 , i = 1, . . . , m, and pj , j = A conditioning of approximately 103 is obtained for n > 15,
1, . . . , n, for a total of 4m + 3n unknowns. Moreover, it is which is deemed acceptable. Notice that the end-effector
important to note, that six components of the ai vectors need positions used for the calibration were distributed approx-
to be prescribed, so as to constrain the mechanism in space. imately evenly across the workspace. Using positions that
This yields a revised total of 4m + 3n − 6 unknowns and the are not sufficiently far apart would lead to a higher condition
condition to avoid an under determined system becomes number of the Jacobian.
4m − 6 By implementing this method, the mechanism is self-
n> . (7) calibrated and resulting identified parameters are the position
m−3
of each fixed cable anchor point ai shown on Table I and
Since, in our case, m = 4, we must measure a minimum of
initial length of each cable ρi,0 shown on Table II.
10 distinct moving platform positions.
A solution can be obtained by formulating the nonlinear TABLE I: Position of each TABLE II: Initial length
least-squares problem and using as an initial guess the cable anchor point Ai of each cable ρi,0
outcome of the first calibration step. Mathematically, this
problem takes the following form x (m) y (m) z (m) x (m)
A1 0 1.827 0 ρ1,0 1.241
minimize n
m X A2 0 0 0 ρ2,0 0.977
X 2
(ui,j ) (8) A3 1.886 0.936 1.713 ρ3,0 1.528
A4 0 0.866 1.994 ρ4,0 1.242
i=1 j=1

821

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on November 12,2020 at 05:17:08 UTC from IEEE Xplore. Restrictions apply.
 As an indication of the accuracy of the calibration, a plot a displacement as input and outputs a force. The second
of the value of a model parameter as a function of the number class of control schemes are called admittance controllers.
of points used in the calibration is shown in Fig. 3. They use the opposite principle, taking forces as inputs, and
A minimum of n = 10 points is necessary to perform supplying displacements as outputs. Admittance control was
the calibration as per (7), for m = 4. It can be seen that chosen here since the additional force feedback alleviates the
the function plateaus for a number of points n of 30 . For inconveniences caused by friction, which has a great impact
n ≥ 30, the observed standard deviation is approximately on the quality of an impedance control.
0.01 m for this parameter, and can reach up to 0.03 m for
others. Fig. 4 shows the quadratic mean of the residuals of A. Control Scheme
(8), which also reflects the impact of the number of points n
Furthermore, a velocity controller is used instead of a
on the solution accuracy and also plateaus around n = 30.
position controller because the resulting haptic rendering has
It is also pointed out that the residual error for n = 10 is
proven superior in previous experiments. While a closed-loop
not zero, which may seem erroneous at first glance. This
control over the velocity of the actuator does not replicate
is correct, however, and indicates that there is no exact
exactly the intended model, it is less sensitive to external
solution even when the system contains as many unknowns
parasitic forces than a position controller. The opposite can
as equations. Given a mechanism roughly the size of a 2 m
be said of a closed-loop position control. It replicates the
cube, we obtain approximately one to two percent error on
exact intended behaviour of the model but any physical per-
the positioning of the anchor points.
turbation generates erratic movements due to the controller
trying to catch up with the model. An unexpected collision
x coordinate of A3 (m)

1.94 into a physical object, exceeding the actuator capabilities or

leaving the workspace are examples of physical perturbations
1.92 that greatly degrade haptic rendering if closed-loop position
1.9 control is used and such situations may be very difficult
to avoid. All these perturbations still have an effect when
1.88 velocity control is used, but are less problematic because they
do not result in unstable movements of the user platform.
1.86 Humans seem far more sensitive to instability errors than to
10 20 30 40 50 velocity errors.
Number of calibration points n A force control feedback is another approach to haptic
control that could have been used. We discarded it because its
Fig. 3: Coordinate y of an anchor point as a function of the mass and damper virtual model need acceleration estimates,
number of points used in the calibration. which are too noisy when derived from encoder measure-
ments.
The general control scheme used for this mechanism is
·10−3 shown in Fig. 5. Forces measured by the sensor and applied
2.5
Root Mean Squares (m)

by the virtual environment are driving the physical model.

This mechanical model of the virtual environment prescribes
2 a Cartesian velocity for the platform to follow the virtual
object. The Jacobian matrix J transforms this Cartesian
1.5 velocity into joint velocity which is given as a command to
the low-level controller Gc , which is a closed-loop controller
1 on articular velocity as displayed by the feedback of ρ̇ in Fig.
5. The forward kinematics are used to feedback the Cartesian
0.5 position to the Jacobian and the virtual environment. It is
10 20 30 40 50 noted that, unlike in many other control schemes devised for
Number of calibrations points n cable driven mechanisms, the controller proposed here does
not require the management of internal cable tensions.
Fig. 4: Quadratic mean of the residual error to the least
squares problem for the mechanism as a function of the
number of points used in the calibration. Virtual p Forward ρ
env. Kinematics
+
IV. CONTROL
+ f ρ̇d
Model ẋ τs
F/T
Two classes of control schemes can be used in the control J Gc Robot
sensor
of haptic devices as described in [17]. Impedance control is ρ̇
the most widely used of these two classes, because it does
not require a force sensor. An impedance control law takes Fig. 5: Block diagram of the control scheme.

822

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on November 12,2020 at 05:17:08 UTC from IEEE Xplore. Restrictions apply.
 The chosen admittance model is that of a mass and B. Virtual Environment
damper, i.e., The virtual environment is the part of the control which
f = Mẍ + Cẋ (15) generates the forces acting on the virtual object represented
where f is the resultant force applied on the object by the by the user’s moving platform in the real world. According
environment and is further described in Section IV-B. The to our model of the virtual environment, the object evolves
terms of mass and damping take the form in either of three modes. The first mode occurs when the
object is moved freely by the user in the workspace. The
M = mp 1[3x3] and C = c1[3x3] , (16) second and third modes occur respectively when the object
where 1[3x3] is the 3×3 identity matrix. This gives a realistic is in contact with a smooth surface and a smooth edge in
model where the mass and damping are the same in all the virtual environment. Each of these situations has its own
directions. component of the resultant force f , i.e.,
The admittance model inputs a force f and outputs a f = fs + fg + fc (24)
velocity ẋ, which is computed as follows. The acceleration
is determined from (15) to be where fs , fg and fc are respectively the force reported by
the sensor, the non-contact external forces and the contact
ẍ(t) = M−1 (f − Cẋ(t)) . (17)
forces.
The velocity at the next time step is then computed as 1) Free Space Motion: Forces acting on the object while
Z t+∆t it is freely moving are forces induced by the user and non-
ẋ(t + ∆t) = ẋ(t) + ẍ(u)du, (18) contact external forces like gravity. Those are represented
t in (24) by the terms fs and fg , respectively. The force fs is
yielding that measured by a force sensor. For this approach to work,
Z t+∆t
low inertias of the sensor and handle are essential, so that
ẋ(t + ∆t) = ẋ(t) + M−1 (f − Cẋ(u))du. (19) the sensor does not pick up inertial forces. If this is not the
t case, a dynamic model of the handle would be necessary to
The Cartesian velocity computed by the admittance model account for their effects. External forces are forces that act in
is then used to control the mechanism. The Jacobian matrix a constant direction in the virtual environment and that can
J calculated from the Cartesian position p of the platform be defined by a static vector fg . A low damping coefficient
is used to compute the desired joint velocities ρ̇d , as per (3) c is normally given to the model to insure system stability,
to obtain : unless a specific scenario requires it otherwise. The mass mp
Z t+∆t ! of the model has to correspond to that of the moving object
−1 in the scenario of the virtual environment.
ρ̇d = J ẋ(t) + M (f − Cẋ(u))du . (20)
t 2) Contact with a Smooth Surface: To generate a contact
The low-level controller Gc use to command the actua- force fc , a contact situation must be defined. For point motion
tors is, in this case, a closed-loop; high-gain; proportional in 3D space, a point of contact and a normal direction are
velocity controller, namely necessary to define a contact without friction.
Consider a smooth surface defined by g(p) = 0, where
τ = Kp (ρ̇d − ρ̇), (21) g(p) < 0 represents the interior of the object and n a unit
vector normal to its surface. The contact force is defined by
where ρ̇d is the vector of desired joint velocities and ρ̇ is
two aspects : the normal n to the plane and the distance from
the vector of actual joint velocities.
P to the surface of the virtual object. The contact occurs
The control signals to the actuators are then computed
when g(p) < 0. The resulting force fc is calculated as a
from (20) and (21) as
! proportion of the penetration distance in the direction normal
Z t+∆t
to the plane, that is.
τ = Kp Jẋ(t) + J M−1 (f − Cẋ(u))du − ρ̇ .
t (
(22) 0 if g(p) ≥ 0,
fc = (25)
This control law can specify negative tension (compression) −Kc n g(p) if g(p) < 0.
to the actuator, which is not feasible by a cable-driven par-
allel mechanism. To circumvent this limitation, a saturation where Kc is a positive gain defining the stiffness of the wall
is placed on τ in order to keep a minimum tension in the and which usually takes a high value for rigid contacts.
cable, namely As discussed in [18], this method could easily create
( instability and energy leaks due to quantization effects for a
τ if τ > τ min high gain Kc . A damping coefficient needs to be added to
τs = , (23) limit this effect. The damping is added only in the direction
τ min if τ ≤ τ min
normal to the contacted surface and only while g(p) < 0,
where τ min is the minimum admissible tension in the cables so as to keep a consistent rendering in the other directions.
and is adjusted experimentally. This damping is added to the damping coefficient of (16),

823

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on November 12,2020 at 05:17:08 UTC from IEEE Xplore. Restrictions apply.
 which becomes V. EXPERIMENTAL RESULTS
C = c1[3×3] + nnT cw (26) Experiments were conducted on the mechanism of Fig.
1 with the admittance control scheme described above. A
where cw is a damping gain. This approach has the second series of scenarios are presented to show the capabilities of
benefit of allowing to adjust the degree of kinetic restitution this haptic feedback device.
of collisions, since it is equal to c + cw . First, as a characterization of the device bandwidth, it
3) Contact with Sharp Edges: When an object is de- was subjected to a sinusoidal velocity command of increas-
scribed by a patchwork of smooth surfaces connected by ing frequency, without any human intervention. The input
sharp edges, contacts at the edges are not properly defined command was well followed until 10 Hz, at which point
by the model of Section IV-B.2 and its direct application oscillations caused by unconstrained end-effector rotations
can result in poor haptic immersion. This occurs because become noticeable. These effects are minimised when a
the sole normal n of the closest surface is used. This effect person is holding the platform, but remain the limiting factor
is apparent when contacts are made with a virtual cube. The in the admissible bandwidth.
normal to each face is defined as pointing outward of the Also, a trajectory in a virtual environment containing
cube. In this manner, a contact with the cube occurs when different objects representing challenges for haptic rendering
all the distance terms are smaller than or equal to zero. The is accomplished. It contains a portion of free motion, a box
contact force is defined in the same manner as in (25). A for contacts with planes and sharp edges, and a sphere for
degradation of the haptic feedback may arise for this kind contacts with smooth surfaces. All surfaces are rigid, which
of situation, as depicted by the trajectory shown in Fig. 6(a). mean that their virtual stiffness is as high as possible without
As the distance to the vertical face becomes smaller than the instabilities. The trajectory as reconstructed from the encoder
distance to the horizontal face, the contact force fc suddenly measurements is presented in Fig. 7.
changes direction. This instantaneous rotation of 90◦ of fc The quality of a haptic rendering is hard to quantify.
is perceived by the user as an artificial push away from Qualitatively, it can be mentioned that the rendering provided
the vertical face. This unrealistic behaviour should thus be by the mechanism is excellent. A low mass mp and damping
avoided. coefficient c allow the platform to move with minimal forces.
To correct this undesirable effect, we propose to adjust The contacts are sharp and convincing, without vibration on
the direction of the normal n. Instead of taking the normal the surfaces. A degradation of the rendering occurs only
to the plane closest to point P , each normal to the cube is when pressing so hard on a surface that the limits of the
taken into account and weighted according to the inverse of actuators are reached. In such a case, the platform starts to
the distance to its corresponding plane, that is penetrate the virtual surface. A video attached to this article
presents the mechanism being manipulated along a trajectory
6
X nk similar to that presented in Fig. 7.
η= , (27)
| dk |p The behaviour of the haptic-feedback device was also
k=1
compared against that of the real word in a simple experi-
where the exponent p defines the smoothness of the edges. ment. This experiment consists in reproducing a ball bounc-
The adjusted normal is then normalised, ing on a hard surface. In a virtual environment artificially
η subjected to gravity, the object was dropped on a virtual
n= , (28)
k η k2 horizontal plane. The contact between the object and the hor-
izontal plane is modeled as a spring and damper. The object
to be used in (25) to generate fc . A trajectory similar to
that of Fig. 6(b) can then be obtained, which matches more
closely what a user would expect when touching a sharp
1.4
edge.
1.2
fs fs fs
fs 1
z (m)

d fs
p p
fc 0.8
fc
fc fc 1.2
0.6 1
fs fs 1.2 0.8
1 0.8 0.6 0.6
fc x (m)
fc y (m)
(a) Without adjusted normal (b) With adjusted normal
Fig. 7: Trajectory of user’s platform in a virtual environment
Fig. 6: Trajectories around a sharp corner from left to right. (lengths in metres).

824

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on November 12,2020 at 05:17:08 UTC from IEEE Xplore. Restrictions apply.
 mass, the contact stiffness and the contact damping were ACKNOWLEDGMENT
adjusted to resemble those of a bouncing ball. The object The authors would like to thank Boris Mayer-St-Onge,
trajectory was recorded through the encoder measurements. Thierry Laliberté and Simon Foucault for their help in
The height of the object is traced against time in Fig. 8. In building the mechanism presented in this paper.
a separate experiment, the height of each bounce of a real
freely-bouncing ball were recorded and are also traced in R EFERENCES
Fig. 8. [1] (2013, Sep.) Occulus VRTM . [Online]. Available:
http://http://www.oculusvr.com/
[2] T. H. Massie and K. J. Salisbury, “PHANToM haptic interface:
Height of the user platform a device for probing virtual objects,” in Proceedings of the 1994
International Mechanical Engineering Congress and Exposition, ser.
0.15 Maximum height of the ball Dynamic Systems and Control Division, vol. 55-1. Massachusetts Inst
of Technology, Cambridge, United States: ASME, 1994, pp. 295–299.
Height [m]

[3] L. L. Cone, “Skycam: an aerial robotic camera system,” BYTE, vol. 10,
0.10 pp. 122–132, 1985.
[4] G. Meunier, B. Boulet, and M. Nahon, “Control of an overactuated
0.05 cable-driven parallel mechanism for a radio telescope application,”
IEEE Transactions on Control Systems Technology, vol. 17, no. 5, pp.
1043–1054, 2009.
0.00 [5] R. Lindemann and D. Tesar, “Construction and Demonstration of a
9-String 6-DOF Force Reflecting Joystick for Telerobotics,” in NASA
International Conference on Space Telerobotics, 1989, pp. 55–64.
−0.05 [6] M. Ishii and M. Sato, “A 3D interface device with force feedback: a
0 0.5 1 1.5 virtual work space for pick-and-place tasks,” in Proceedings of IEEE
Time [s] Virtual Reality Annual International Symposium. Seattle, WA, USA:
IEEE, 1993, pp. 331–5.
Fig. 8: Height of the user platform while bouncing in a virtual [7] R. Xiao, J. Huang, and P. Wu, “Large space force feedback system
based on string,” in Proceedings of 2011 International Conference
environment in comparison to the bouncing height of a ball on Computer Science and Network Technology, vol. vol.4. Harbin,
in the real world. China: IEEE, Dec. 2011, pp. 2705–2708.
[8] Z. Chen, Y. Zhang, D. Wang, C. Li, and Y. Zhang, “iFeel6-BH1500:
a large-scale 6-DOF haptic device,” in 2012 IEEE International
As shown on this graph, the mechanism closely follows Conference on Virtual Environments Human-Computer Interfaces and
the real-world ball for the first bounce, but diverges slowly Measurement Systems (VECIMS). Tianjin, China: IEEE, 2012, pp.
121–125.
from the real trajectory in subsequent bounces. This result [9] P. Gallina, G. Rosati, and A. Rossi, “3-d.o.f. Wire Driven Planar Haptic
is expected because the control scheme does not guarantee Interface,” Journal of Intelligent and Robotic Systems, vol. 32, no. 1,
a perfect replication of the physics of the real world, but pp. 23–36, 2001.
[10] R. L. Williams II, “Cable-Suspended Haptic Interface,” International
focuses on smooth and stable rendering. Nevertheless, the Journal of Virtual Reality, vol. 3, pp. 13–21, 1998.
mechanism follows the trajectory of a real object closely [11] T. Morizono, K. Kurahashi, and S. Kawamura, “Realization of a virtual
enough to avoid disrupting the immersion of the user. A sports training system with parallel wire mechanism,” in Proceedings
of International Conference on Robotics and Automation, vol. 4, no.
video of this experiment shows both objects at the same time April. Albuquerque, NM, USA: IEEE, 1997, pp. 3025–3030.
and demonstrates this life-like behaviour. [12] C. Carignan and K. Cleary, “Closed-loop force control for haptic
simulation of virtual environments,” Haptics-e, vol. 1, no. 2, pp. 1–14,
2000.
VI. CONCLUSIONS [13] C. Gosselin, R. Poulin, and D. Laurendeau, “A Planar Parallel 3-DOF
Cable-Driven Haptic Interface,” in 12th World Multi-Conference on
A control law based on an admittance scheme and closed- Systemics, Cybernetics and Informatics, vol. 3, Orlando, FL, USA,
loop velocity control was presented and proven to provide 2008, pp. 266–271.
a realistic haptic rendering on the three-degree-of-freedom [14] P. Miermeister and A. Pott, “Auto calibration method for cable-
driven parallel robots using force sensors,” Latest Advances in Robot
cable-driven mechanism presented herein. Also, a simple Kinematics, pp. 269–276, 2012.
self-calibration method based on measurement redundancy [15] P. H. Borgstrom, B. L. Jordan, B. J. Borgstrom, M. J. Stealey, G. S.
was proven effective in an experiment providing a relative Sukhatme, M. A. Batalin, and W. J. Kaiser, “NIMS-PL: A Cable-
Driven Robot With Self-Calibration Capabilities,” IEEE Transactions
error from one to two percent on the mechanism geometry. on Robotics, vol. 25, no. 5, pp. 1005–1015, 2009.
Future work will include adjusting the low-level controller [16] J. Alexandre Dit Sandretto, D. Daney, M. Gouttefarde, and C. Baradat,
Gc to allow a wider dynamic bandwidth by, for example, “Calibration of a fully-constrained parallel cable-driven robot,” INRIA,
Tech. Rep. RR-7879, 2012.
including a feed-forward term taking into account inertial [17] M. Otis, V. Duchaine, G. Billette, S. Perreault, C. Gosselin, and
forces. Other improvements include the possibility of in- D. Laurendeau, “Cartesian Control of a Cable-Driven Haptic Mecha-
teracting with moving and movable objects in the virtual nism,” in Advances in Haptics, Mehrdad Hosseini Zadeh, Ed. Intech,
2010, ch. 4, pp. 75–103.
environment, as well as adding four cables to increase the [18] R. B. Gillespie and M. R. Cutkosky, “Stable user-specific haptic ren-
number of degrees of freedom to six. Having six degrees of dering of the virtual wall,” in Proceedings of the ASME International
freedom will allow the platform to generate torque, extending Mechanical Engineering Congress and Exhibition, vol. 58, 1996, pp.
397–406.
the system to more complex virtual objects and collisions.
The last step will consist in placing the mechanism inside a
CAVE environment, to immerse the user into a full 3D both
visual and haptic feedbacks.

825

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on November 12,2020 at 05:17:08 UTC from IEEE Xplore. Restrictions apply.