17th IEEE International Conference on Control Applications
2
Part of 2008 IEEE Multi-conference on Systems and Control
San Antonio, Texas, USA, September 3-5, 2008
Single Magnet Levitation by Repulsion using a Planar Coil Array
Peter Berkelman and Michael Dzadovsky
Abstract— This paper describes the controlled levitation of a
single disk magnet above an array of cylindrical coils, using
electromagnetic repulsion to generate actuation forces and
torques from coil currents, and an optical motion tracking
system to provide realtime position and orientation measure-
ments for feedback control. Our method is to measure the
electromagnetic forces and torques generated from a single
cylindrical coil current on a disk magnet as a function of
their relative positions beforehand, then combine the forces
and torques from multiple identical coils in a planar array to
derive matrix transformations from coil currents to forces and
torques, given the position of the magnet during levitation. If
the transformation matrix is invertable, then the coil currents
necessary to generate forces and torques for feedback control
can be calculated. With our setup, proportional-derivative feed-
back control applied in each direction of motion is sufficient for
stable magnetic levitation. Experimental results for levitation of
a single magnet are given for a simplified case with 3 coils with
the magnet motions partially constrained by an attachment to
a ball joint, and for 5 coils with the rotation of the magnet
about its vertical axis left uncontrolled to rotate freely.
I. INTRODUCTION
To achieve stable controlled magnetic levitation of a
rigid body, 6 degrees of freedom (DOF) of electromagnetic
actuation, noncontact position and orientation sensing, and
feedback control are required to produce a full 6-element
vector of forces and torques on the levitated body. Earn-
shaw’s theorem of 1842 [1] regarding electrostatic levitation
of rigid bodies is applicable to the electromagnetic case as
well; no static arrangement of permanent magnetic fields can
stably levitate a rigid body.
In this work we use a planar array of cylindrical coils to
levitate one or more magnets. This arrangement is distinct
from magnetic levitation systems such as magnetic bearings
and levitation systems for transportation in that multiple coils
generate forces on each magnet in all directions, rather than
using separate magnet and coil assemblies for actuation and
control of each degree of freedom of the levitated body.
In our system each coil generates forces and torques in
all directions and their combined actuation provides the
stabilizing forces and torques for levitation using feedback
from a motion tracking position sensor.
Magnetic levitation in general provides considerable bene-
fits to practical applications such as precision motion control,
vibration isolation, and haptic interfaces, as friction and other
dynamic nonlinearities such as hysteresis and cogging typical
of other actuation methods are eliminated.
This work was supported by NSF grant CNS-05515, a donation from Real
Data Inc., and the University of Hawaii-Manoa College of Engineering.
Peter Berkelman and Michael Dzadovsky are with the Department
of Mechanical Engineering, University of Hawaii-Manoa, 2540 Dole St,
Honolulu HI USA. Email: peterb@hawaii.edu
978-1-4244-2223-4/08/$25.00 ©2008 IEEE. 108
WeA04.
II. BACKGROUND
Research in magnetic levitation with multiple DOF can
be classified into the areas of Lorentz force systems for fine
positioning, teleoperation, and haptic interaction; magnetic
bearings; planar motor systems for fine positioning; and
systems for transportation.
Planar magnetic levitation systems for silicon wafer pho-
tolithography have been developed by Kim and Trumper [2],
[3], [4] based on planar Sawyer motors [5]. Their devices
provide up to 30 nm position repeatability and up to 50
mm motion in the horizontal plane, yet less than 1 mm
of motion in the vertical direction and milliradians-range
rotation. Hollis and Salcudean developed magnetic levitation
using Lorentz forces [6] rather than electromagnetic attrac-
tion or repulsion. Lorentz magnetic levitation devices have
been applied to compliant robot wrists [7], force-feedback
teleoperation [8], and haptic interaction [9]. Magnetic bear-
ings levitate rotating shafts using electromagnetic attraction
to eliminate friction. Various systems for magnetic levitation
trains have been developed [10] including electromagnetic
systems based on attraction, electrodynamic systems with
superconducting coils to generate stronger fields, and the In-
ductrack system [11] which uses induced currents in passive
coils to generate levitation and guidance forces.
Non-contact position feedback for magnetic levitation
control systems can be provided by Hall effect sensors, laser
interferometry, or combinations of LED markers and position
sensing photodiodes. It has also been shown to be possible
to levitate magnets without feedback control by using highly
diamagnetic [12] or superconducting materials, or by using
spin to stabilize the motion of the magnet [13], [14].
A maglev stage developed by Lai, Lee, and Yen [15]
also uses a planar array of cylindrical coils to levitate disk
magnets, however their system requires 15 magnets and 37
solenoid coils for stable levitation, as separate sets of coils
and magnets are used to stabilize each axis of translation and
rotation.
III. G ENERAL L EVITATION A PPROACH
Our current system levitates a single magnet using a
limited number of coils, but the method can be extended
without difficulty to arbitrarily large coil arrays and levi-
tated platforms with multiple magnets to increase the planar
motion range and lifting capacity, such as pictured in Fig. 1.
The coils currently in use in the planar array are approx-
imately 28 mm in diameter and 28 mm high, with 1000
turns of wire and copper cores for effective heat dissipation.
Neodymium-iron-boron disk magnets 37.5 mm in diameter
and 12.5 mm high with a mass of 125 g and a maximum
 horizontal force (N)
2.0
1.0
0 .0
−1.0
−2.0
30 20
10 −40 −20 0 20 40
height (mm) horizontal offset(mm)
Fig. 1. Sample base planar coil array and levitated magnet platform

vertical force (N)

4.0
3.0
2.0
1.0
0 .0
−1.0
30 20
10 −40 −20 0 20 40
height (mm) horizontal offset(mm)

torque (N−mm)
20

0
Fig. 2. Coil actuation force measurement setup including coil, magnet,
force/torque sensor, motion stage, and power supply −20

energy product approaching 50 MGOe are used for levitation. 30

Each coil produces a combination of forces and torques
in different directions on each magnet in its vicinity. The
rotational symmetry of the coils and magnets simplifies the
model of the forces and torques generated between them.
A. Control Hardware
An optical motion tracking sensor1 is used with infrared
LED markers fixed to a rigid body to provide position
feedback with a resolution of approximately 0.01 mm, at
a sample rate of 500 Hz with our current software interface.
This sensor is mounted to a rigid frame 1.5 m above the plane
of the levitation coils so that its range is approximately 1 m
in each direction in the plane. The controller is implemented
on a standard PC equipped with a 32-channel PCI analog
output card2 . A power supply and PWM current amplifiers3
generate the required coil currents from the PC analog output
signals.
B. Actuation Force & Torque Measurements
The forces and torques generated generated between a
single coil and magnet over a range of vertical separation
distances and horizontal offsets for a given actuation current
were measured using the setup shown in Fig. 2, where a 6-
axis force/torque sensor4 is mounted between the magnet and
an assembly of vertical and horizontal motion stages. The
vertical force, radial horizontal force, and torque generated
on the magnet were measured at 2 mm intervals of vertical
separation and radial offset and the resulting data are shown
in Fig. 3. The forces and torques generated as the roll
1 OptoTrak,
Northern Digital Inc.
2 PD2-AO-32/16, Ueidaq Inc.
3 4212Z, Copley Controls Inc.
4 Mini40, ATI Industrial Automation Inc.
20
10 0 20 40
−40 −20
height (mm) horizontal offset(mm)
Fig. 3. Force and torque from single magnet and coil with 1.6 A current
and pitch tilt orientation of the magnet is varied will be
measured during planned future work to enable magnets
to be stably levitated at a range of orientations as well as
positions. All the forces and torques generated were found to
be proportional to the coil current to a close approximation.
The shape of the torque surface of Fig. 3 is of particular
note due to its complex shape with multiple local maxima
and minima and zero crossings.
C. Coil current to force-torque transformation
The forces and torques generated on a single magnet from
the currents in multiple coils in the planar array can be
combined and expressed in matrix form as a linear matrix
transformation A from a vector of coil currents I to a force
and torque vector F :
F = AI, (1)
where the number of columns of A is equal to the number
of active actuator coils and the number of rows is equal
to the number of controlled DOF of the magnet. For the
feedback control necessary for stable levitation, the inverse
transform matrix A−1 is used to calculate the coil currents
needed to generate the forces and torques to satisfy the
control laws. Therefore, the A matrix must be invertable,
at least as many coils are needed as controlled DOF of the
magnet motion, and stable levitation is impossible for the
magnet positions which produce singular A transformation
matrices. Furthermore, increasingly high coil currents must

109
be produced as the magnet position approaches these singular
configurations, which are undesirable to avoid overheating
the coils and/or saturating the current amplifiers. As coil
currents of over 2.5 A were found to cause rapid overheating
of the coils, the coil currents generated by the feedback
control system were limited to ±2.5 A by the control
software.
The condition number of the A matrix may be used as a
quantitative measure of how close the magnet is to a position
which would generate a noninvertible, singular transform
matrix. Analysis of the matrix condition number across
ranges of possible magnet positions allows the best magnet
positions for stable control and the effectiveness of various
coil configurations to be evaluated. The matrix determinant
or the maximum element in the inverse matrix could also be
used as similar quantitative metrics in a manner similar to
manipulability and controllability measures in robotics [16].
D. Feedback Control
Decoupled proportional-derivative feedback control laws
are implemented for each degree of freedom of motion of
the levitated magnet. A force feedforward term equal to the
weight of the magnet is added to the vertical magnet position
control law to cancel the effect of gravity.
IV. M AGNETIC L EVITATION WITH BALL J OINT
C ONSTRAINT
As a preliminary step towards unconstrained, spatial mag-
netic levitation, the magnet was connected to a low-friction
ball joint by a short, lightweight rod to partially constrain
its motion. This arrangement constrains the magnet freedom
of motion to the 3 DOF consisting of the vertical motion
z, horizontal motion perpendicular to the connecting rod
x, and rotation around the connecting rod axis 6 y. The
magnet is thus restricted to planar rigid-body motion with
two translational and one rotational DOF, rather than spatial
rigid-body motion with 3 translational and 3 rotational DOF.
The derivation of the transform matrix and the analysis of
the coil configuration for this simplified planar rigid-body
motion case is presented here to illustrate the methods used
and the relationship between the arrangement of the coils, the
position of the magnet, and the performance and stability of
magnetic levitation in general.
A. Setup
A single row of 3 actuation coils with centers spaced 35
mm apart is used to control the 3 DOF of the magnet motion.
A thin disk of clear plastic with 4 infrared LED position
markers is attached to the top of the magnet for motion
tracking position feedback, as shown in Fig. 4.
B. Analysis
Given the horizontal and vertical magnet position and the
coil positions, and using the single coil and magnet force and
torque data of Fig. 3, it is straightforward to combine the
force and torque contributions from each coil on the magnet
to generate a 3x3 element transformation matrix which can
110
Fig. 4. 3 DOF levitation setup with ball joint
be used to calculate a vector of forces and torques from
a vector of coil currents. In this planar motion magnetic
levitation case, the force/torque and current vectors from
equation (1) are
   
fx i1
F = fz  , I = i2  , (2)
τy i3
and the transformation matrix A is
A=
 
u(mx − c1 , mz ) u(mx − c2 , mz ) u(mx − c3 , mz )
 v(mx − c1 , mz ) v(mx − c2 , mz ) v(mx − c3 , mz )  ,
w(mx − c1 , mz ) w(mx − c2 , mz ) w(mx − c3 , mz )
(3)
where the horizontal and vertical position of the magnet is
given as mx and mz , and the positions of the coil centers
are c1 , c2 , and c3 . The horizontal force generated on the
magnet by a single coil with a 1.0 A current as a function
of the horizontal and vertical separation of the magnet and
coil is u(x, z), the vertical force is v(x, z), and the torque is
w(x, z). The shapes of the u, v, and w functions are shown
in the top, middle, and bottom plots of Fig. 3 respectively,
although the u, v, and w functions are scaled to 1.0 A coil
currents rather than 1.6 A as pictured. The torque function
w(x, z) is expressed in units of N-cm so that the elements of
A are of the same approximate order of magnitude on each
row.
The condition number of the transformation matrix A, up
to a maximum of 20, is shown in Fig. 5 as a function of the
vertical and horizontal position of the magnet. This surface
plot indicates that singular configurations are produced when
the magnet is directly above any one of the three coils,
and the magnet can be levitated with lower maximum coil
currents and improved stability when it is approximately
midway between two neighboring coils. When the magnet is
directly above a coil the lifting force of the coil is maximized,
but unfortunately these configurations lead to inability to
control the horizontal position and tilt of the magnet unless
additional coils and redundant actuation schemes are used.
 x position (mm)
18.35

18.3

18.25
20
18.2
condition number
transform matrix

15 0 0.05 0.1 0.15 0.2 0.25 0.3

10 4.2

5 4

z position (mm)
0
8 3.8
6
4
2
3.6
10 40
8 20 3.4
6 0
4 −20
vertical 3.2
position (mm) 2 −40 horizontal position (mm)
3

y rotation (degrees)
Fig. 5. Quality measure for 3 coil configuration 0 0.05 0.1 0.15 0.2 0.25 0.3
−0.1
C. Results −0.15

Fig. 6 shows the vertical, horizontal, and rotational motion −0.2

of the magnet in response to a step input command from a −0.25
levitated height of 3.0 mm to 4.0 mm. The desired horizontal −0.3
0 0.05 0.1 0.15 0.2 0.25 0.3
position was -18.0 mm, which produces a transformation time (seconds)
matrix with a condition number of 4.0.
The gains used for the position feedback control are: Fig. 6. Ball joint magnetic levitation results

translation rotation
Kp 0.3 N/mm 5.25 N-mm/degree
Kd 0.003 N-sec/mm 0.0525 N-mm-sec/degree
where Kp is the proportional error and Kd is the derivative
error feedback gain.
The steady-state errors of 0.26 mm in x, 0.05 mm in z,
and -0.2 degrees in 6 y in the magnetic levitation control
results are due to small variations in the forces and torques
generated by different coils and small misalignments of the
individual coil positions and the motion tracking position
sensor zero position. The addition of error integral control
or adaptive control terms into the feedback control laws used
would help to drive the steady-state levitated magnet position Fig. 7. Sample 5 coil configurations for levitation
errors to approach zero.
A. Coil Configuration Analysis
V. F REE M AGNETIC L EVITATION The force/torque and current vectors from equation (1)
A single disk magnet can be levitated, but its rotation about for the case of levitation leaving the yaw rotation 6 z uncon-
its vertical axis cannot be controlled due to the symmetry of trolled, are given by:
its magnetic field. We therefore leave this DOF uncontrolled.
   
fx i1
Accordingly, 5 coils are sufficient to levitate the magnet, fy  i2 
provided that the magnet is allowed to rotate freely about
   
F = fz  , I = i3  , (4)
  
its vertical axis. The transformation matrix A for this case  τx  i4 
can be obtained from a given 3D magnet position, the set τy i5
of planar coil positions, and force/torque data of Fig. 3, by
extending the methods of section IV. and the transform matrix can be calculated as:
Six different configurations for the 5 coils were evaluated
A=
by calculating the condition numbers of their coil current  
to force/torque transformation matrices across a range of cos(θ1 )u(r1 , mz ) cos(θ2 )u(r2 , mz ) ···
 sin(θ1 )u(r1 , mz ) sin(θ2 )u(r2 , mz ) · · ·
magnet positions. The candidate configurations are shown  
in Fig. 7 and will be referred to as the ’Cross’, ’T’, ’L’,

 v(r 1 , mz ) v(r2 , mz )  , (5)
· · ·
− sin(θ1 )w(r1 , mz ) − sin(θ2 )w(r2 , mz ) · · ·
’Pentagon’, ’Hex’, and ’C’ configurations. The coil center
separation for adjoining coils in all configurations is 35 mm. cos(θ1 )w(r1 , mz ) cos(θ2 )w(r2 , mz ) ···

111
 Cross: min. cond. #: 6.91 ’T’: min. cond. #: 7.25 B. Setup
30 30
The ’L’ configuration from Fig. 7 was used for the
y position (mm)

y position (mm)
20 20 implementation of magnetic levitation experiments due to the
10 10 large region where its transform matrix condition number is
0 0 less than 10.0. The magnet and coil setup was similar to
−10 −10 previous case of section IV, but with two additional coils
−20 −20 aligned at a right angle to the previous three, and without the
−30 −30
ball joint and rod connected to the levitated magnet. A 3.15
−30−20−10 0 10 20 30 −30−20−10 0 10 20 30
mm aluminum plate was placed on top of the levitation coils
x position (mm) x position (mm) underneath the levitated magnet to provide additional passive
’L’: min. cond. #: 8.01 Pent.: min. cond. #: 10.20 damping of the motion of the magnet by generation of eddy
30 30
currents, thereby stabilizing levitation for higher proportional
control gains.
y position (mm)

y position (mm)

20 20
10 10 C. Results
0 0 The control gains used to obtain the levitation results
−10 −10 shown in Fig. 9 are the same as those used in section IV. The
−20 −20 magnet was levitated at the position of x = −22, y = 13,
−30 −30 and z = 4, where the condition number of the current to
−30−20−10 0 10 20 30 −30−20−10 0 10 20 30 force/torque transform matrix is 8.0. Static magnet levitation
x position (mm) x position (mm) at this position for periods greater than approximately 30
Hex: min. cond. #: 3.95 ’C’: min. cond. #: 8.86 seconds leads to overheating of the two coils farthest from
30 30 the magnet however, as these coil currents are greater than
2.0 A. Small steady-state errors of up to 0.3 mm and 0.5
y position (mm)
y position (mm)

20 20
10 10
degrees are present as in the previous results, due to small
0
position and magnetic field variations in the experimental
0
setup and the absence of integrating or adaptive control
−10 −10
terms. The levitated magnet, actuator coils, and the plastic
−20 −20
disk with LED position markers are shown in Fig. 10.
−30 −30
−30−20−10 0 10 20 30 −30−20−10 0 10 20 30 VI. C ONCLUSIONS AND F UTURE W ORKS
x position (mm) x position (mm)
A. Conclusions
4 6 8 10 12 14 16 The experimental results have demonstrated the basic
contour line colors
feasibility of the magnetic levitation control methods. These
Fig. 8. Quality measures for 5 coil configurations methods are applicable to any combination of actuation coils,
in any planar configuration, levitating any number of magnets
where the radial offset distances ri between the vertical axes connected together, so that the present system can be easily
of the magnet and each coil are given by: reconfigured, or extended to an arbitrary size and motion
q range. The methods are first used to measure and compare the
ri = (mx − cix )2 + (my − ciy )2 (6) levitation effectiveness of different coil configurations and
magnet positions, and then used to derive the force/torque
and the rotation angles θ to separate radial horizontal forces and coil current transformations to be used in feedback
and torques into x and y components are: control for levitation.

θi = tan−1 ((my − ciy )/(mx − cix )), (7) B. Future Works

provided that the tan−1 function returns the correct angle
0 < θ < 2π for all 4 possible quadrants determined by the
signs of the numerator and denominator.
The rotational symmetry of the coils and magnets simpli-
fies the force and torque generation models sinceponly the
vertical distance z and the radial distance r = x2 + y 2
need to be measured, rather than measuring both x and y.
The condition numbers of the A transform matrices as a
function of magnet x and y positions at a z levitation height
of 4 mm are shown in Fig. 8 for each coil configuration, for
areas where the A condition numbers are less than 16.0.
The current and planned further research is to improve
and extend these control methods for magnetic levitation for
higher loads, larger ranges of motion, and more dynamic
motion trajectories.
The actuation force and torque data of Fig. 3 will be
extended to measure changes in forces in torques as the
magnet tilt angles 6 x and 6 y or roll and pitch are varied by
up to ±90 degrees. This actuation data will enable levitation
of magnets at tilted angles, instead of with a flat orientation
only. A spherical magnet will be levitated as well as the
present disk magnet, and additional LED markers will be
used to provide position feedback as the magnet rotation

112
x pos (mm)
−22.2 by adding more coils in both direcions, to increase the
−22.3 motion range of the levitation system. Multiple magnets will
−22.4 be attached together in a single levitated platform so that
0 0.05 0.1 0.15 0.2 0.25
the yaw angle 6 z can be controlled and to increase the
y pos (mm)

13 lifting capacity of the system. With more coils available than

12.9 controlled motion DOF of the levitated platform, redundant
12.8 actuation methods will be developed and a commutation
12.7 system will be used to switch currents to different sets of
0 0.05 0.1 0.15 0.2 0.25
coils as the levitated magnet moves over the coil array area.
4.4
ACKNOWLEDGMENTS
4.2 The authors thank Ji Ma for his assistance programming
z pos (mm)

the software interfaces for the OptoTrak motion tracker.

4
R EFERENCES
3.8
[1] S. Earnshaw, “On the nature of the molecular forces which regulate the
constitution of the luminiferous ether,” Transactions of the Cambridge
3.6 Philosophical Society, vol. 7, pp. 97–112, 1842.
[2] W.-J. Kim and D. Trumper, “High-precision magnetic levitation stage
3.4 for photolithography,” Precision Engineering, vol. 22, pp. 66–77,
1998.
0 0.05 0.1 0.15 0.2 0.25 [3] W.-J. Kim, D. Trumper, and J. Lang, “Modeling and vector control
x rot (deg)

−0.3 of planar magnetic levitator,” IEEE Transactions on Industry Applica-

tions, vol. 34, no. 6, p. 12541262, November/December 1998.
−0.5 [4] D. Trumper, W.-J. Kim, and M. Williams, “Design and analysis
framework for linear permanent magnet machines,” IEEE Transactions
−0.7 on Industry Applications, vol. 32, no. 2, pp. 371–379, March/April
0 0.05 0.1 0.15 0.2 0.25 1996.
y rot (deg)

0.8
0.6
0.4
0 0.05 0.1 0.15
time (seconds)
Fig. 9. Magnetic levitation results
Fig. 10. Single levitated magnet
angles cause the present LED markers not to be visible to
the overhead motion tracking sensor.
To realize smooth trajectory following over larger ranges
of motion, the A transformation matrices and their inverses
for each will be calculated from magnet position at each
control update during motion. A linear interpolation of
force/torque data points will be implemented instead of
selecting the nearest data point. A model of the levitated
rigid-body dynamics will be incorporated into the trajectory
controller.
Finally, the coil array will be extended to a larger area
[5] W. Hinds and B. Nocito, “The sawyer linear motor,” in Proceedings
of the 2nd Symposium on Incremental Motion Control Systems and
Devices, 1973, pp. 1–10.
[6] R. L. Hollis and S. E. Salcudean, “Lorentz levitation technology: a
0.2 0.25 new approach to fine motion robotics, teleoperation, haptic interfaces,
and vibration isolation,” in Proc. 6th Int’l Symposium on Robotics
Research, Hidden Valley, PA, October 2-5 1993.
[7] R. L. Hollis, S. Salcudean, and A. P. Allan, “A six degree-of-
freedom magnetically levitated variable compliance fine motion wrist:
design, modeling, and control,” IEEE Transactions on Robotics and
Automation, vol. 7, no. 3, pp. 320–332, June 1991.
[8] S. Salcudean, N.M. Wong, and R.L. Hollis, “Design and control
of a force-reflecting teleoperation system with magnetically levitated
master and wrist,” IEEE Transactions on Robotics and Automation,
vol. 11, no. 2, pp. 844–858, December 1995.
[9] P. J. Berkelman and R. L. Hollis, “Lorentz magnetic levitation for hap-
tic interaction: Device design, function, and integration with simulated
environments,” International Journal of Robotics Research, vol. 9,
no. 7, pp. 644–667, 2000.
[10] H.-W. Lee, K.-C. Kim, and J. Lee, “Review of maglev train technolo-
gies,” IEEE Transactions on Magnetics, vol. 42, no. 7, pp. 1917–1925,
July 2006.
[11] R. Post and D. Ryutov, “The inductrack: A simpler approach to
magnetic levitation,” IEEE Transactions on Applied Superconductivity,
vol. 10, no. 1, pp. 901–904, March 2000.
[12] M. D. Simon, L. O. Heflinger, and A. K. Geim, “Diamagnetically
stabilized magnet levitation,” Americal Journal of Physics, vol. 69,
no. 6, pp. 702–713, June 2001.
[13] M. V. Berry, “The Levitron (TM): An adiabatic trap for spins,”
Proceedings: Mathematical, Physical, and Engineering Sciences, vol.
452, no. 1948, pp. 1207–1220, May 1996.
[14] T. B. Jones, M. Washizu, and R. Gans, “Simple theory for the levitron,”
Journal of Applied Physics, vol. 82, no. 2, pp. 883–888, July 1997.
[15] Y.-C. Lai, Y. Lee, and J.-Y. Yen, “Design and servo control of a single-
deck planar maglev stage,” IEEE Transactions on Magnetics, vol. 43,
no. 6, pp. 2600–2602, June 2007.
[16] T. Yoshikawa, “Manipulability of robotic mechanisms,” International
Journal of Robotics Research, vol. 4, no. 2, pp. 3–9, 1985.

113