torque control of a push-pull cable driven powered orthosis ...torque control of a push-pull cable...
Post on 26-Sep-2020
6 Views
Preview:
TRANSCRIPT
Vrije Universiteit Brussel
Torque Control of a Push-Pull Cable Driven Powered Orthosis for the CORBYS PlatformRodriguez Guerrero, Carlos; Grosu, Victor; Grosu, Svetlana; Leu, Adrian; Ristic-Durrant,Danijela; Vanderborght, Bram; Lefeber, DirkPublished in:International Conference on Rehabilitation Robotics
Publication date:2015
Document Version:Final published version
Link to publication
Citation for published version (APA):Rodriguez Guerrero, C., Grosu, V., Grosu, S., Leu, A., Ristic-Durrant, D., Vanderborght, B., & Lefeber, D.(2015). Torque Control of a Push-Pull Cable Driven Powered Orthosis for the CORBYS Platform. In InternationalConference on Rehabilitation Robotics (pp. 61-66). Singapore: IEEE.
General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright ownersand it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portalTake down policyIf you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediatelyand investigate your claim.
Download date: 25. Jan. 2021
Torque Control of a Push-Pull Cable DrivenPowered Orthosis for the CORBYS Platform.
Carlos Rodriguez Guerrero∗, Victor Grosu∗, Svetlana Grosu∗, Adrian Leu†, Danijela Ristic-Durrant†, Bram Vanderborght∗ and Dirk Lefeber∗.
∗Robotics and Multibody Mechanics Group, Vrije Universiteit Brussel, Brussels Belgium†Institute of Automation, University of Bremen, Germany
Corresponding email: carodrig@vub.ac.be
Abstract—Rehabilitation robotics is a growing field whichis on the verge of exploring novel actuation technologies thatallows the designers to build assistive devices with large powerto weight ratios without compromising the transparency of thesystem. In this paper a novel push-pull cable driven technologyimplemented in the CORBYS rehabilitation system as a solutionfor a proximally actuated device is presented. A novel torquecontrol strategy enhanced with a machine learning compensa-tion method is proposed to deal with the inherent complexitiesof the system. Experiments will show results obtained on thepowered exoskeleton part of the platform.
I. INTRODUCTION
We have recently experienced in our times a boom in
the so called rehabilitation robotics domain, where robots
are designed to physically interact with patients and relief
therapists from the physical effort needed to carry the thera-
pies and giving them a more ’decision making’ supervisory
and scientific role. Several commercial and non-commercial
devices have appeared in the recent years such as the Au-
toAmbulator (ReoAmbulator) Healthsouth [1]), the Lokomat
[2], the WalkBot [3], [4] LOPES [5], ALEX [6], KNEXO
[7], WalkTrainer [8], ALTACRO [9].
In spite of this boom, a clear advancement over con-
ventional therapy stays out [10],[11],[12]. Pennycott [13]
made a review study on the outcome of robot-assisted gait
rehabilitation. They propose the inclusion of more degrees of
freedom (DOF) that allows more aspects of the gait such as
balance to be trained, and robots with greater transparency
for a more natural human robot interaction. They also pointed
out: ’These (robots) will provide an important opportunity in
the near future to intensively investigate the clinical impacts
of these various concepts.’ To tackle both, added DOF and
transparency at the same time, seems to have a fundamental
problem in the way that the addition of more active DOF,
faces the intrinsic problem of adding more actuators to the
structure and therefore adding more weight to it most likely
hindering the transparency and available output torque of the
system.
Decoupling the actuation system from the effector can
give freedom to design more powerful rehabilitation systems
without the inconvenience of the added weight due to the
inclusion of bigger motors that translates into added inertia
on the links. This type of actuation is so called proximal
Fig. 1. Subject walking on a treadmill with the CORBYS gait rehabilitationsystem while the system is being controlled in zero force mode.
actuation and aims to provide high power to weight ratios by
placing the actuators over a grounded structure rather than
directly on the joints and has been already implemented in
lower limb rehabilitation robots such as LOPES [5], LOPES
II and now the CORBYS platform [14]. In LOPES, the basic
idea was to detach the motor from the robot frame by the
use of flexible Bowden cables and series elastic elements
so that the exoskeleton legs could move unhindered, while
little weight is added to the robot construction, compared
to the relatively heavy electro motors [5]. Bowden cables
however, are not able to withstand compression forces, so all
the structure should ensure that all the cables remain tight
whatever the configuration of the structure [15]. CORBYS
uses instead push-pull cable (PPC) technology which can
stand both, tension and compression forces.
CORBYS - Cognitive Control Framework for Robotic
Systems is a colaborative project funded by the european
commission under the 7th framework program [16], [14].
CORBYS developed new principles for the way that human
and robot interact and share their cognitive capabilities.
Structurally, CORBYS includes a combination of a mobile
platform, a lower limb exoskeleton and a pelvis module that
acts like a weight support system and drives the exoskeleton
61978-1-4799-1808-9/15/$31.00 c©2015 IEEE
up, down, left and right, see Fig. 1.
This paper focuses on the compliant control of the ex-
oskeleton’s PPC novel actuation system that drives the active
degrees of freedom of the robotic orthosis in the sagittal
plane. The main contribution in this paper, is that of pre-
senting a novel control architecture that can be used to
deal with the major drawbacks of using PPCs for power
transmission. This architecture features a model compensa-
tion system based on machine learning. Section 2 includes
a brief description of the PPC actuated system and shows
data of the dynamics found at the hip and knee joints of
the powered orthosis. It also presents the joint space control
strategies used to make the PPC driven system act as a
torque source which is the base for any further haptic strategy
like impedance control. Section 3 describes experiments to
validate the proposed controller. In section 4 and 5 we present
and discuss the results achieved and finally conclusions are
presented in section 6.
II. MATERIALS AND METHODS
A. The CORBYS platform
The CORBYS powered orthosis assists the patients lower
limb joint motions. The configuration of the DOFs of the
CORBYS orthosis was implemented to follow natural human
limb kinematics. There are 6 DOFs at each leg: 3 DOFs
in the hip, 1 DOF in the knee, and 2 DOFs in the ankle
joints. The hip and knee joint motions on the sagittal plane
(flexion and extension), as well as ankle joint motions in the
sagittal plane (plantarflexion and dorsiflexion), are selected
as the active DOFs based on the biomechanics properties of
human walking, while the hip DOFs in the frontal and the
transverse plane (adduction/abduction and internal/external
rotation) as well as the ankle DOF in the frontal plane
(eversion and inversion) are passive. The orthosis of the
CORBYS system is attached to a mobile platform via a 4
DOFs pelvis interface mechanism (consisting of the pelvis
link and the linear unit), which enables pelvis rotation in the
frontal and transverse planes as well as the pelvis vertical
and side to side movement. Together with the 4 DOFs of the
pelvis interface mechanism, the CORBYS orthosis provides
a total of 16 DOFs for the body movement, which makes
it the first orthotic device with such a number of DOFs.
Additionally, the orthosis was designed in order to satisfy
the requirement of adaptability to the patients height and
weight, as well as the requirement of the range of motion
of the biomechanical joints.
The particular novelty of the CORBYS orthosis is that the
hip joint provides all 3 DOFs of the human hip joint where
only the rotation in sagittal plane is actuated which, however,
does not influence the rotations in the transverse and frontal
planes. This is achieved via a common center of rotation of
three hinge-joints. This common center of rotation is aligned
with the biomechanical hip joint of the human. As another
novelty, an angle independent weight compensation of the
orthosis in the frontal plane is implemented, as it is needed
Fig. 2. Representation of the CORBYS PPC actuation system. Actuatorsare placed on frame and power is transferred by push pull cables to the joint.
for the passive abduction/adduction joint. The movements of
the orthosis joints in the sagittal plane are controlled by a
push-pull control cable actuation system [14][17]. There are
three actuators (PRL+ modules, SCHUNK) per orthosis leg
that actuate the hip, knee and ankle joints in the sagittal plane.
B. PPC actuation system
The exoskeleton’s actuation system, see Fig. 2, The
actuators are placed on the mobile platform while the PPC
cables are flexible links to the joints used to transfer the
rotational movement of the motors to specifically designed
orthotic joints. This has the advantage of keeping the
powered orthosis weight low while the necessary torques are
applied through the PPC cables to the patients joints. In the
CORBYS system, PPCs of different diameters and lengths
are used, as different moments are required for actuation
of orthosis leg joints. Therefore, the rotary movement of
the actuator has to be transformed to a linear movement in
order to actuate PPC. This transmission is provided using
a mechanical lever construction where the actuation force
depends on the lever length and position. The transmitted
force changes with the lever angular position. This means
the wider the operation range of the actuator, the smaller
is the applicable force in PPC direction. The force applied
by the actuator is reduced by the system geometry and the
friction in the PPC. The efficiency of the PPC is essential
for the function of the powered orthosis.
The actuators for the active joints are placed in a proximal
way (related to the structure) back on the mobile platform
while the PPC is a flexible link to the joints. The static
(unmovable) part of the PPC is mounted to the demonstrator
base frame. The distance from the smart actuators center to
the static attachment of the PPC depends on the necessary
stroke of the PPC. This stroke is given by the kinematic
62 2015 IEEE International Conference on Rehabilitation Robotics (ICORR)
design of the orthotic joints. Because, each joint (hip, knee,
and ankle) has a different kinematic design various stroke
parameters will be attributed.
Fig. 3. Sketch of mechanism that transforms push-pull cable forces intotorques and vice versa
Figure 3 shows the skeleton diagram of the actuation
mechanisms. In order to get the torque to force relation that
the mechanism provides, we can start by assuming P1 and
P2 have always the same velocity. This will be the starting
and key point to get to our analytical solution. Figure 4
shows the mechanism decomposition where:
R = Distance between the joint and the slider.R1 = Length of the crank link.R2 = Length of the connector link.P = Slider velocity.τ1 = Joint Torque.F = Force component on the direction of R2.ω1 = Motor angular velocity.θ1 = Angle between R1 and the slider mechanism axis.θ2 = Angle between R2 and the slider mechanism.
Fig. 4. The skeleton diagram of the actuation mechanisms. This helps tovisualize the analytical solution that translates force-cell readings with thetorque at the crank link.
Assuming no important losses in the mechanism one can
safely state that the power transmitted from R2 to R1 is
constant and thus:
τ1ω1 = PF (1)
Solving the equations by using the principle of virtual
work, and solving for the angle θ2, we can get to the
following relation:
θ2 = arcsinR1(1− sinθ1)
R2(2)
Having the angle θ2, we can now find τ1, which is the
torque we are interested in with the following formula:
F cos(θ2) =τ1
R1(sin θ1 − cos θ1 sin θ2)(3)
Which leads to:
τ1 = F cos(θ2)(R1(sin θ1 − cos θ1 sin θ2)) (4)
Knowing F cos(θ2) is the load cell reading, R1 is a
constant value, θ1 is read from the joint encoder and θ2 is
computed from the motor position, we can easily obtain τ1from equation (4).
C. Torque Control of the CORBYS orthosis
Typically, electric motors have a linear relation between
the armature current and the torque output which makes
it straightforward to implement impedance control loops
whenever the power transmission between the motor and
the joint is direct, and there is nothing between the mo-
tors and the joints. The CORBYS system, however, uses
a proximal actuation system by means of push-pull cables
and the SCHUNK’s PRL+ units have large harmonic drives
which makes the system non backdrivable. The use of PPC’s
and non-backdrivable motors have different consequences for
the control. While there exists well known techniques such
as admittance control that can turn backdrivable systems
into force/position controlled systems by using force sensor
feedback and an inner position/velocity loop that tries to
render the virtual admittance output. This is not a suitable
solution in the CORBYS setup as we can’t just simply use
the internal position/velocity loop embedded in the PRL+ unit
since is the joint what we are aiming to control. What we
have instead is a non-collocated problem, as the sensor that
closes the loop is not sensing the actuators output directly.
That means that in order to control a virtual admittance at
the joint, we need to implement the inner position/velocity
loop at the joint side first. However, implementing a stable
position/velocity controller on a non collocated system with
considerable backlash and connected by a PPC system proved
not to be an easy task since most of the times it resulted in an
unstable system due to the big discrepancies on the motion
of the source reflected on the load and also in this case, the
resolution of the encoders on the motor side and the joint
side. Moreover the dynamics of the CORBYS PPC system
proved to be highly nonlinear.
In order to have an idea of how the system dynamics
at the joint level looked alike, we have performed some
experiments to gather data that maps the configuration of the
robot to the torques read at the joint levels. Figure 5(a) and
2015 IEEE International Conference on Rehabilitation Robotics (ICORR) 63
Fig. 5. Angle Vs Torque characteristics for hip and knee joints. Several nonlinear dynamics can be observed.
Figure 5(b) show the relationship between angles and torques
at the hip and knee level in response to a slow sinusoidal
trajectory commanded to the motor. In those figures we can
see complex nonlinear behavior such as mechanical backlash,
internal friction and flexibility mainly due to the influence of
the PPC internal bending.
D. Joint space torque control
In this paper, we propose to create a fast and robust torque
control that can later be used by an outer impedance or proxy
based sliding mode controller to render assistive/corrective
force fields. Figure 6 shows a block diagram of the imple-
mentation of the proposed joint space torque controller for
CORBYS.
In order to deal with the backlash and non-collocated
control problem, we have implemented a PI+D dual loop,
a variant of a PID for systems with backlash [18].
In this dual PID loop originally proposed for position
control systems, the proportional and integral terms operate
on the signal coming from the joint load, forming the main
feedback loop, while the derivative term acts on motor load
feedback, which is the best source of information for stabi-
lizing the system. The tuning of the PI+D torque controller
was made experimentally by using a square wave input
(alternating step signal) as a virtual desired torque, and tuning
the gains to improve bandwidth and disturbance rejection. It
is important to point out that the harmonic drives behavior
curves and therefore the output torque of the PRL+ unit is
Fig. 6. General proposed torque control block diagram. The system iscomprised by a dual loop torque controller, a velocity limiter controllerand a neural network robot model compensator.
dependent not only on the input current and the current/torque
relation, but also on the gearbox efficiency, which depends in
turn on the speed and on the gearbox temperature which will
require an extra sensor. Therefore the tuning was made while
the gearbox was warm after been exposed to some use (1
hour) as it increases its efficiency making it more responsive
and therefore changes the stability margins.
Another important implementation issue that is faced in the
CORBYS system, was the speed limit safety implementation
that the PRL+ units have by default. The maximum velocity
of the motors is 45 deg/s. For safety reasons, the motors
activate their brakes if this velocity is exceeded. Therefore,
a velocity limiter that observes the velocity, and counteracts
the commanded current to clamp the velocity to a desired
maximum level was designed. This was implemented by
means of a P controller that acts on the error signal between
the maximum and read velocities, with a counter-current
output that adds to the torque controller and prevents the
motor to get to an undesirable level.
E. Robot model compensation
One of the most important characteristics of any phys-
ically interactive robot is that of being able to properly
render reactive force fields. In order to achieve this, the
low level control system ideally should behave like a perfect
force/torque source independent of the system dynamics. This
means that in order to achieve transparency, we need to have
a good model compensation.
As the system dynamics resulted to be heavily non linear,
a parametric model to compensate for each joint would
most likely result in a very complex yet inaccurate solution.
Therefore a multi-layer perceptron (MLP) was trained to
learn a model to compensate for the hip and knee own
dynamics coming from both, position and velocity dependent
terms like gravity, Coriolis, centripetal and viscous friction.
Inertia was not compensated as there is no acceleration
feedback in CORBYS. The ankle was excluded since the
impact of the model compensation on the overall behavior
of the joint was not significant enough to justify for the two
64 2015 IEEE International Conference on Rehabilitation Robotics (ICORR)
extra inputs and one output on the MLP and since it is the
final link of the kinematic chain, there is no coupling dynamic
effects that propagates further to the structure.
The MLP was trained in MATLAB by using the neural
network toolbox algorithm and then exported to C++ using
the coder toolbox for optimized code generation. The training
data set was obtained by driving sinusoidal trajectories on
both hip and knee with amplitudes that covered all the
range of motion for both. Notice that the excitation function
frequency for the knee was 10 times bigger than that on the
hip to cover as much as possible the whole configurations of
the robot.
The results on the error from the real data and the MLP
estimation for the hip joint can be seen in Figure 7. The
resulting error (Fig 7b) from the difference between the
obtained data and the model estimation output for the hip
(Fig 7a) was in mean=-0.0067 and std=2.63 Nm while for
the knee was in mean=0.0012 and std=2.8501 N.m.
Fig. 7. On top, Hip torque data together with the perceptron prediction.At the bottom, the mean error between real hip data and estimated hip dataover 12000 samples at 100Hz (120 seconds). NN stands for neural network.
III. EXPERIMENTS
Two experiments were performed to test different aspects
of the proposed controller design. On the first one, we wanted
to see the tracking response of the torque controller and
specially the backlash rejection. For that, we commanded a
sinusoidal wave to the knee’s torque controller set point and
registered the commanded current to the motor, the motor’s
velocity, the joint’s position and joint’s torque feedback.
For the second experiment, the torque controller’s set point
was set to 0 and the maximum velocity was set to 41 deg/s,
while we applied alternating external forces to the joint by
hand to the knee joint back and forth, to accelerate the link
and see the effects of the velocity clamper over the current
output and the motor velocity signal.
IV. RESULTS
Figure 8 shows results on the first experiment, showing
a low tracking error between the commanded torque signal
and the read torque. Besides good torque tracking, we can
also see how the backlash rejection (black box in Figure
8) provides a smooth motion profile by looking at the
position signal. It is also visible the effect of the motor being
momentarily accelerated by an over current to deal with the
non contact zone (zero torque zone due to the backlash)
and provide a smooth output. In spite of that short burst of
current, the commanded current provided by the controller is
typically below the normal operation levels of the motor (10
A commanded current) which is important for durability of
the machine.
Fig. 8. Torque controller tracking and backlash rejection response (blackbox). We can see the controller’s fast reaction in response to the discontinuityin the torque signal.
Figure 9 show the action of the velocity clamper. We can
see the effectiveness of the limiter as we applied external
forces to the joint to accelerate it into a fast motion.
Fig. 9. Velocity clamper controller with a set point of 41 deg/s. Torque’scomputed, counter and output currents are displayed to show the effects ofthe limiter on the current sent to the motor.
2015 IEEE International Conference on Rehabilitation Robotics (ICORR) 65
V. DISCUSSION
The proximal actuation system proved to be an excellent
way of decoupling the motors from the structure allowing
the usage of bigger and more powerful motors. The use of
this technology however comes at its costs as the PPC’s
presented considerable losses due to the friction created by
the internal bending between the inner and outer cables.
Besides that, backlash was also induced by the PPC’s internal
bending which presented a challenge in the control design.
This however has been successfully solved by implementing
the control strategy proposed in section 2.D where the PI part
tends to drive the motor faster based on the joint’s torque
error signal, while the D part keeps the system stable in
response to the error signal on the motor’s side torque and the
outer loop command. The response of the proposed controller
was satisfactory in dealing with the mechanical design of the
system and we believe it can be extrapolated to any device
that uses PPC actuation with force control. It is also visible
that there is some ripple effect when the limit velocity is
reached. Although this phenomenon is not critical, it could
be compensated by using a PD controller instead of a pure P
action. However, since the feedback for the velocity limiter is
the motor’s velocity, and this is obtained by pure backwards
differentiation, a D controller could induce considerable noise
and may result unpractical.
VI. CONCLUSIONS
Using PPC’s as a form of transmission proved to be an
effective way of dealing with systems that requires high
power ratios with low added inertia at the end effector such
as the one present in the CORBYS powered orthosis. It
is important nevertheless to be aware of the lengths and
bending angles of the PPCs since we have noticed most of the
undesirable dynamic effects that we had in the platform came
from the excessive long cables as the longer they are the more
internal space for bending it will be and thus, more friction,
backlash and losses there will be. The use of PPC’s induces
several challenges from the control point of view such as
backlash and complex dynamics that may be difficult to
model using a classical robot parametric approach. However,
as shown in this paper, the use of machine learning to
compensate for the robot dynamics proved to be an excellent
and easy to use solution albeit the lack of flexibility that
an algebraic model gives. We are studying the possibility to
include an on-line learning algorithm in the near future using
an accelerated version of a Gaussian Process Ensemble [19]
designed under the CORBYS project. This could allow us to
re-train the system when needed as the system’s properties
may change due to machines wear or when considerably
changing the robot’s links to fit a new patient.
ACKNOWLEDGMENT
This research was supported by the European Commission
as part of the CORBYS (Cognitive Control Framework for
Robotic Systems) project under contract FP7 ICT-270219.
The views expressed in this paper are those of the authors,
and not necessarily those of the consortium.
REFERENCES
[1] “AutoAmbulator,” www.healthsouth.com/experience-healthsouth/the-healthsouth-difference/leading-technology/autoambulator, 2014. [On-line]. Available: www.healthsouth.com/experience-healthsouth/the-healthsouth-difference/leading-technology/autoambulator
[2] R. Riener, L. Lunenburger, I. Maier, G. Colombo, and V. Dietz,“Locomotor Training in Subjects with Sensori-Motor Deficits: AnOverview of the Robotic Gait Orthosis Lokomat,” Journal of Health-care Engineering, vol. 1, no. 2, pp. 197–216, 2010.
[3] J.-H. Jung, N.-G. Lee, J.-H. You, and D.-C. Lee, “Validity andfeasibility of intelligent Walkbot system,” Electronics Letters, vol. 45,no. 20, p. 1016, 2009.
[4] D. H. Kim, Y.-I. Shin, K.-L. Joa, Y. K. Shin, J. J. Lee, and S. J. H. You,“Immediate effect of Walkbot robotic gait training on neuromechanicalknee stiffness in spastic hemiplegia: A case report,” NeuroRehabilita-tion, vol. 32, no. 4, pp. 833–838, 2013.
[5] J. F. Veneman, R. Ekkelenkamp, R. Kruidhof, F. C. T. van der Helm,and H. van der Kooij, “A Series Elastic- and Bowden-Cable-BasedActuation System for Use as Torque Actuator in Exoskeleton-TypeRobots,” The International Journal of Robotics Research, vol. 25, no. 3,pp. 261–281, 2006.
[6] D. Zanotto, P. Stegall, and S. K. Agrawal, “ALEX III: A novel roboticplatform with 12 DOFs for human gait training,” Proceedings - IEEEInternational Conference on Robotics and Automation, pp. 3914–3919,2013.
[7] P. Beyl, K. Knaepen, S. Duerinck, M. Van Damme, B. Vanderborght,R. Meeusen, and D. Lefeber, “Safe and Compliant Guidance by aPowered Knee Exoskeleton for Robot-Assisted Rehabilitation of Gait,”International Journal of Advanced Robotic Systems, vol. 25, no. 5, pp.513–535, Jan. 2011.
[8] Y. Stauffer, Y. Allemand, M. Bouri, J. Fournier, R. Clavel, P. Metrailler,R. Brodard, and F. Reynard, “Pelvic motion measurement during overground walking, analysis and implementation on the WalkTrainer reed-ucation device,” in IEEE/RSJ International Conference on IntelligentRobots and Systems, 2008, pp. 2362–2367.
[9] B. Brackx, C. Rodriguez Guerrero, V. Grosu, R. Van Ham, M. VanDamme, B. Vanderborght, and D. Lefeber, “Design of the gait reha-bilitation robot ALTACRO : A powered exoskeleton using compliantactuation,” Currently in review, p. 1, 2015.
[10] E. Swinnen, S. Duerinck, J.-P. Baeyens, R. Meeusen, and E. Kerckhofs,“Effectiveness of robot-assisted gait training in persons with spinalcord injury: a systematic review,” Journal of Rehabilitation Medicine,vol. 42, no. 6, pp. 520–526, 2010.
[11] E. Swinnen, D. Beckwee, R. Meeusen, J.-P. Baeyens, and E. Kerckhofs,“Does robot-assisted gait rehabilitation improve balance in strokepatients? A systematic review.” Topics in stroke rehabilitation, vol. 21,no. 2, pp. 87–100, 2014.
[12] N. Koceska and S. Koceski, “Article: Review: Robot Devices forGait Rehabilitation,” International Journal of Computer Applications,vol. 62, no. 13, pp. 1–8, 2013.
[13] A. Pennycott, D. Wyss, H. Vallery, V. Klamroth-Marganska, andR. Riener, “Towards more effective robotic gait training for strokerehabilitation: a review,” J Neuroeng Rehabil, vol. 9, p. 65, 2012.
[14] S. Slavnic, D. Ristic-Durrant, R. Tschakarow, T. Brendel, M. Tutte-mann, A. Leu, and A. Graser, “Mobile robotic gait rehabilitation systemCORBYS - overview and first results on orthosis actuation,” pp. 2087–2094, 2014.
[15] B. Dehez and J. Sapin, “ShouldeRO, an alignment-free two-DOFrehabilitation robot for the shoulder complex,” (ICORR), 2011.
[16] “CORBYS project webpage.” [Online]. Available:http://www.corbys.eu/
[17] S. Grosu, C. Verheul, C. Rodriguez-guerrero, B. Vanderborght, andD. Lefeber, “Towards the elaboration of 3D dynamic model for push/ pull cable ( PPC ) actuation system,” in IMSD/AMSD, 2014.
[18] “Motion control made easy.” [Online]. Available:www.electromate.com/db support/attachments/Backlash Compensa-tion.pdf
[19] C. Glackin, C. Salge, M. Greaves, D. Polani, and D. Risti, “GaitTrajectory Prediction using Gaussian Process Ensembles,” in IEEE-RAS International Conference on Humanoid Robots, Madrid, 2014.
66 2015 IEEE International Conference on Rehabilitation Robotics (ICORR)
top related