Sie sind auf Seite 1von 11

872

IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 4, AUGUST 2008

A HumanExoskeleton Interface
Utilizing Electromyography
Christian Fleischer and Gunter Hommel

AbstractThis paper presents a humanmachine interface to


control exoskeletons that utilizes electrical signals from the muscles of the operator as the main means of information transportation. These signals are recorded with electrodes attached to the
skin on top of selected muscles and reflect the activation of the
observed muscle. They are evaluated by a sophisticated but simplified biomechanical model of the human body to derive the desired
action of the operator. A support action is computed in accordance to the desired action and is executed by the exoskeleton. The
biomechanical model fuses results from different biomechanical
and biomedical research groups and performs a sensible simplification considering the intended application. Some of the model
parameters reflect properties of the individual human operator and
his or her current body state. A calibration algorithm for these parameters is presented that relies exclusively on sensors mounted
on the exoskeleton. An exoskeleton for knee joint support was designed and constructed to verify the model and to investigate the
interaction between operator and machine in experiments with
force support during everyday movements.
Index TermsElectromyography (EMG), exoskeleton, human
machine interface, torque amplifier.

I. INTRODUCTION
ESEARCH on exoskeletons is not a new topic to the scientific community. Already in the 1970s, the group around
Vukobratovic [1] investigated possible constructions and control strategies for disabled people. Since then, numerous groups
have experimented with exoskeletons, motivated by the large
number of potential applications: the Berkeley lower extremity
exoskeleton (BLEEX), for example, is a military exoskeleton to
aid soldiers carrying heavy loads [2], and the hybrid assistive
leg is an actuated body suit for both legs [3], in the latest version
extended for both arms. It is designed to support elderly people and as a device for rehabilitation. The powered lower limb
orthosis described in [4] is developed to assist during motor re-

Manuscript received October 12, 2007; revised February 5, 2008. This paper
was recommended for publication by Associate Editor K. Yamane and Editor
H. Arai upon evaluation of the reviewers comments.
The authors are with the Institute for Computer Engineering and Microelectronics, Technische Universitat Berlin, Berlin 10587, Germany (e-mail:
fleischer@cs.tu-berlin.de; hommel@cs.tu-berlin.de).
This paper has supplementary downloadable multimedia material available
at http://ieeexplore.ieee.org. provided by the author. This material includes
one video (TUPLEE-Video.mpg) demonstrating the TUPLEE-exoskeleton with
force support. The operator is sitting on a chair, getting up, walking a few steps,
climbs a small stair, turns on the platform of the stair, walks downstairs and
back to the chair where he sits down (all in a laboratory). After that, walking
in a staircase in an indoor environment is shown. In the laboratory setup, data
curves of the knee angle, the estimated contribution of the operator, and the
torque produced by the actuation are shown online. The video requires a video
player capable of playing standard MPEG2-files. The size is 70 MB. Contact
fleischer@cs.tu-berlin.de for further questions about this work.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TRO.2008.926860

habilitation after neurological injuries by relearning typical gait


patterns. Many more exoskeletons have been developed over the
years, showing the relevance of this research and the width of
the application spectrum, for example, to act as carrying aids
for factory workers [5], to regain dexterity of hand and finger
movements [6][14], and to support shoulder movement [15].
Further potential applications include haptic interfaces to learn
handling processes, virtual reality, and entertainment. Especially
in recent years, the reduction of the size and weight of actuators
and power supplies gave a motivating push to the developments.
But whatever the application of such an exoskeleton is, there
is need for an interface between the operator and the device as
soon as online control of the exoskeleton should be allowed.
Such an interface needs to be intuitive to minimize the process
of learning the handling, and to allow the operator to concentrate
on fulfilling a task in cooperation with the exoskeleton rather
than on the control.
In most of the aforementioned research projects, the dynamics
of the operator together with the exoskeleton are evaluated to
compute the support torque for the joints. In case of the BLEEX
for example, the evaluation of the inverse dynamic model of the
exoskeleton results in the joint torques that are currently applied.
By contributing a certain share of these torques through the
actuation of the exoskeleton, the operator is relieved of some of
the load [2]. In other projects the exoskeletons add fixed shares
of the torque that is required to maintain a static pose with the
current joint configuration in [6], [16]. The drawback of these
control systems is that to perform a movement, the operator
has to be able to initiate the movement before he or she can
receive support from the system. The control systems are not
directly coupled to the operator and the utilized dynamic models
are very sensitive to kinematic measurements and to changes in
the masses of operators, exoskeletons, payloads, and to external
contact forces that have not been modeled or have not been
accurately measured [17], [18].
An appropriate alternative is to connect the control system to
biological signals of the operator that are directly linked to the
desire of movement. Since some movements are generated in
the brain, some in the brain stem, and some in the spinal cord, it
is convenient to record these signals at the final output: the muscles. Prior to the actual contraction, a signal, called EMG1 signal,
is emitted by the muscle that can be detected by either needle
electrodes, inserted into the muscle, or surface electrodes, fastened on top of the muscle that should be observed. This signal
appears during activation of the muscle approximately 2080
ms [19], [20] before the resulting contraction begins, allowing
the signal evaluation to begin before the movement. Even if the
1 Abbreviation

1552-3098/$25.00 2008 IEEE

for electromyography.

FLEISCHER AND HOMMEL: HUMANEXOSKELETON INTERFACE UTILIZING ELECTROMYOGRAPHY

muscles are too weak to perform any movement, a support can be


given. Furthermore, the EMG signals are emitted unconsciously
by the operator and no additional mental load is created.
Evaluation of EMG signals to control support devices is most
widely spread in the area of rehabilitation or for functional
support of disabled people. Since EMG signals of patients are
weak and hard to interpret, evaluation is mostly performed with
classification algorithms, not using physiological models. Upon
recognition of certain EMG patterns, predefined trajectories
are replayed, making it a feasible approach for hand prostheses
control as, for example, in [21][23]. Continuous EMG
evaluation without a physiological model is performed, for
example, in [24] for an anklefoot orthosis where air pressure
is linearly related to the EMG signal, in [15] where shoulder
movement is computed with a neuro-fuzzy controller, and
in [25] for a full body exoskeleton where EMG signals are
linearly related to joint torques.
One rare exception is the work presented in [9], where EMG
signals of the biceps brachii and triceps brachii have been evaluated with a Hill-type muscle model to estimate the elbow joint
moment. A moment controller was fed to control a two-link
exoskeletal arm to lift an external load with the hand. A comparison between this model and a black box model with a neural
network is presented in [26].
The problem of complex physiological models is that the
more muscles are incorporated into the model, the more
parameters have to be determined. All EMG-related parameters
are subject-dependent and can change from day to day due
to varying conditions of the skin and body state. A nonlinear
least squares algorithm is used to determine the parameters for
each subject in [27]. This is possible with a complex external
measurement setup that allows accurate reference recordings
during different movements. The parameters of the Hill-based
state equation have been solved in [9] for a single extensor
and flexor muscle of the human upper arm. The tendon slack
lengths and optimal muscle fiber lengths have been determined
in [28] for four elbow flexors together with a parameter for
muscle stress by recording the torque of the muscles under
maximum voluntary contractions and electrical stimulation
in different angular configurations. A nested Nelder-Mead
simplex optimization was utilized to determine all parameters.
In this paper, an exoskeleton, called Technische Universitat
Berlin Powered Lower Extremity Exoskeleton (TUPLEE), is
developed that is driven by a biomechanical model that evaluates the EMG signals of the operator to derive the desired action.
The exoskeleton supports the thigh muscles during flexion and
extension of the knee joint and is more suited for healthy people who are capable of controlling their muscles and can take
full advantage of the flexible control system while receiving
the support. The body model and the calibration are described
for this particular exoskeleton so that proper graphic descriptions can be given, although they can be applied to other body
parts and exoskeletons as well. In contrast to the approaches
described earlier, our calibration method allows determination
of all parameters without requiring any external measurement
setup aside from the exoskeleton itself, making the practical
application of this system easy.

873

Fig. 1. Control structure of the system. The body model computes the joint
torque the operators muscles are producing by evaluation of the EMG signals.
This torque is multiplied by a support ratio and passed to the torque control loop
that controls the actuator to produce the desired torque.

The following Section II describes the control structure before


in Section III, the biomechanical model is introduced. Section IV
explains calibration of model parameters, before Section V describes the exoskeleton construction, and experiments justifying
the model complexity with the actuated exoskeleton are presented in Section VI. Section VII summarizes the discussion
and a short conclusion is given in Section VIII.
II. CONTROL STRUCTURE
The concept of the control system is shown in Fig. 1: the
EMG signals are read from the operators muscles and brought
into the body model that estimates the resulting torque in the
joint. This torque is multiplied by the support ratio giving the
target torque for the torque control loop. This loop includes
an actuator model that computes the current torque produced
by the actuator from the force sensor and joint angle readings.
The error between current torque and target torque is passed
to the controller that sets the appropriate control signals for
the actuation. This evaluation is performed continuously. No
position controller and no predefined trajectories are used. The
whole controller acts as a torque amplifier.
It is assumed here that a linear amplification results in a
system behavior that the operator can take into account easily
and that the shape of the muscle activation pattern is changed
as little as possible. It is desired that only the amplitude of the
muscle activation is reduced, resulting in a short training phase
to get used to the behavior of the system.
III. HUMAN BODY MODEL
The human body model is responsible for estimating the
torque of the operator from the EMG signals of all observed
muscles. It is a combination of models from biomechanical and
biomedical research groups, but simplified to fit the needs of this
particular application, and to allow calibration with low effort.
The chosen degree of complexity is justified in experiments in
Section VI-A. The order of computation is explained inside out:
at first, the EMG signals of all regarded muscles are converted

874

IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 4, AUGUST 2008

passive force FPm , when the muscle is stretched:


F m = FAm + FPm .

(2)

The force of the contractile element is calculated by the product


of the muscle activation a(u) and the maximum force the muscle
can produce, which is the maximum isometric force4 at optimal
muscle fiber length Fom , and the active forcelength function
fA (lm ), as defined in [30]:
Fig. 2. Schematic of the muscletendon model with a contractile element
(CE) producing the force FAm ,a passive element (PE) producing FPm , and the
resulting muscle force F m as a sum of FAm and FPm that acts at an angle to
the direction of pull, resulting in the force of the musculotendinous unit F m t .
l m is the fiber length and l m t the length of the whole unit.

into the activation levels of the muscles before the resulting


muscle forces are computed. These forces are transformed into
the individual joint torques and summed. The computation is
shown for a single but arbitrary muscle, and the index i for the
variables and parameters is omitted.
A. EMG Signal to Muscle Activation
The muscle activation reflects the level of active force production of a muscle. It ranges between 0 and 1 and can be
expressed as a function of the neural activation of a muscle,
which, in turn, is the degree of stimulation the muscle receives
from the nervous system, expressed through the postprocessed
EMG signal u. Based on the definition in [29], the activation
a(u) of a muscle is computed by

FAm = fA (lm )Fom a(u)

B. Muscle Activation to Muscle Force


Once we have obtained the activation of the muscle, we
can compute the resulting force F m , using a simplified muscle model that is shown in Fig. 2 (adapted from [30]). It consists
of two elements: a contractile element producing the active muscle force FAm , and a parallel elastic element that produces the
2 Since this function is applied online, a definite maximum cannot be determined in advance. Due to potential submaximal muscle activations during
calibration of R, u is allowed to exceed R in the following experiments. In
contrast to [29], u is not normalized in advance.
3 Common values are between 410 Hz [30]. Our experiments have shown an
improved system behavior for lower frequencies (refer to Section VI-C).

m
lm = l
lom

(3)

where lm is the current muscle fiber length and lom the length
at which the muscle can produce maximum force. fA (lm ) describes the ability of the muscle to produce force under different
muscle fiber lengths [31].
The passive force is calculated as a product of the maximum
isometric force Fom , and the normalized passive forcelength
curve fP , as given in [32] by
FPm = fP (lm )Fom

with

m
lm = l .
lom

(4)

To calculate the current length of the muscle fibers lm , a


complex muscle and bone model has to be used that describes
the pathway of the muscletendon complex taking into account
the necessary joint angles. Such a model is published in [33]
allowing computation of the length of the musculotendinous
unit lm t , taking into account the necessary joint angles.
As can be seen in Fig. 2, lm t can be expressed through
lm t = lt + lm cos

eA u R 1
(1)
a(u) =
eA 1
where u is the postprocessed EMG value, R an expected
maximum2 (average of maxima of all trials during calibration)
of the signal u, and A a nonlinear shape factor defining the
curvature of the function with A < 0. For A 0, the function
approximates a linear relationship.
The raw EMG signal is postprocessed with dc offset elimination, full-wave rectification, subtraction of the measured offset
when the muscle is relaxed, and low-pass filtering with a secondorder Butterworth filter with a cutoff frequency of 1.6 Hz3 , resulting in a phase delay of about 130 ms for common activation
patterns. Cross-talk effects have not been considered, because
the muscles of the thigh are large and neighboring muscles have
a similar task.

with

(5)

where l is the length of the tendon and the current pennation


angle that changes with the muscle length. According to [34],
the tendon is rather stiff: the strain is only about 3% of the
tendon length for maximum muscle force and is neglected here.
With this simplification, it is shown in [35] that the length of
the muscle fibers can be computed with the pennation angle at
optimal muscle fiber length o , through

lm = (lom sin o )2 + (lm t lt )2
(6)
which is not depending on the current pennation angle anymore.
But this equation contains the length of the tendon lt , which is
not known exactly. Since the tendon strain is omitted, values
for the tendon slack length from literature [33] can be used.
But unfortunately, the model is very sensitive to this quantity
that varies between subjects. As a consequence, a tendon slack
length scale st is introduced to adapt the tendon slack length
from literature lst to the individual operator:
lst = st lst .

(7)

C. Muscle Force to Joint Torque


According to Fig. 2, the force of the musculotendinous unit
F m t is calculated by
F m t = F m cos
4 Force

that is produced without length change of the muscle.

(8)

FLEISCHER AND HOMMEL: HUMANEXOSKELETON INTERFACE UTILIZING ELECTROMYOGRAPHY

where can be computed as derived in [35] by


 m

l sin o
= arcsin o m
.
l

875

(9)

The moment arms of the ith muscle around the joint ri are
approximated constant values derived from applying the tendon displacement method described in [36] to lim t . The torque
contribution Ti of muscle i is calculated by
Ti = ri Fim t

(10)

where Fim t is the force of the ith musculotendinous unit. The


total joint torque T is the sum of the individual contributions:
T =

Fig. 3. Concept of the calibration. The joint torque T R is distributed among all
muscles allowing independent calibration using the relationship between EMG
values and the torque contribution of the muscle (indicated by the dots).

B. Calibration Concept
Ti .

(11)

IV. CALIBRATION OF PARAMETERS


The biomechanical model contains a large number of parameters that, in theory, have to be adapted to the individual
operator. In order to be able to perform the calibration with
sensors mounted on the exoskeleton (for a mobile, easy-to-use
system), only the most important parameters are calibrated, and
parameters that are well established in the biomechanical and
biomedical communities are taken from literature. Parameters
of muscle i selected for optimization are:
1) the shape of the EMG-to-force function Ai ;
2) the expected maximum EMG signal Ri ;
m
m
, Fo,i
;
3) the maximum isometric force at lo,i
t
4) the scale of the tendon slack length si .
The scales sti are geometry parameters and are only required
to be calibrated once per subject. The EMG-to-force parameters
m
have to be calAi and Ri and the maximum isometric force Fo,i
ibrated once for every experimental session, because they are
affected by the current condition of the subject (skin conductivity, etc.). The index i is omitted in the following whenever we
refer to only one arbitrary muscle.

During the trials described in the previous section, the EMG


signals of all muscles, the knee angle, and the force at the actuator FR produced by all muscles are recorded. Data selection and
storage have been performed as in [37]. As a result, one table
per muscle per trial exists where the table is selected through
the knee angle (corresponding to trial k) and the postprocessed
EMG signal u maps linear to the entry index v (typically 50
100 entries per table). Every entry contains one value for every
recorded quantity that is averaged with new values whenever
the table entry is selected during data storage.
The maximum expected EMG signal R of a particular muscle is the maximum taken from the topmost entries of the tables
of this muscle. The force in all entries FR is converted into a
knee torque TR , then stored in each entry and used as a reference for the calibration. These reference values are distributed
among all muscles according to their activation as described in
Section IV-C (refer to Fig. 3), so that the parameter calibration can be performed for all muscles independently using the
relationship between the postprocessed EMG values of different muscle activations and the corresponding force output of
the muscle. After that the geometry parameters are calibrated
(described in Section IV-D) followed by the determination of
the EMG parameters (described in Section IV-E) utilizing the
geometry parameters.

A. Calibration Setup

C. Force Determination of Individual Muscles

The operator is sitting on a chair with the exoskeleton leg


not having any contact with the environment below the knee
joint. The actuator is not powered but locked, fixing the knee
at an arbitrary angle. When the thigh muscles are relaxed, the
force measured by the force sensor integrated into the actuator
is a result of gravitation acting on the exoskeleton, shank, and
foot. During each trial, the operator tries to extend or flex the
knee with maximum muscle force a few times against the locked
actuator (isometric contractions). The measured force is now an
overlay of all muscles and the influence of gravitation.
For the geometry calibration, these isometric trials have to
be performed several times at different knee angles to record
data with different muscle fiber lengths, whereas for the EMG
parameter calibration, a single trial is sufficient.
In the following, a knee angle of 0 represents a straight leg
and negative angles indicate knee flexion.

To calibrate all muscles independently, the force share for


the ith muscle Fim based on the torque TR has to be computed
(torque distribution in Fig. 3) for each table entry:
Fim =

Ti
ri cos i

with

wi
Ti = TR 
i wi

(12)

where ri is the moment arm, i the current pennation angle, and


wi the weight of the muscle. wi is computed by using the relative
physiological cross-sectional area of the muscle, %PCA, and the
current activation of the muscle:

%PCA ai (ui ), ifAi already calibrated
(13)
wi = %PCA ui ,
otherwise.
Ri
The activation function in the first case ai (ui ) is only available
if this calibration has been performed before and Ai from the

876

IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 4, AUGUST 2008

Fig. 4. Establishing the relationship between the postprocessed EMG signal


and the muscle activation for each muscle.

previous run is available. Otherwise the activation function is


approximated by a linear relationship (second case).
D. Geometry Calibration
The goal of the geometry calibration is to optimize the tendon
slack length scale st for every muscle. The muscle activation
for table entry v of table k, ak ,v , can be computed as the inverse
muscle model from (2)(4) by
ak ,v =

(Fkm,v /Fom ) fP (lm )


fA (lm )

with

m
lm = l
lom

Fig. 5. Muscle activation derived from the reference values plotted against
corresponding EMG values for different knee angles without optimization of
the tendon slack length scale (left). The geometry calibration has minimized the
standard deviation in vertical direction, and the EMG calibration has computed
the continuous calibrated function (right).

(14)

where Fkm,v is the force share when the postprocessed EMG


signal was equal u and the knee angle was k (data table k, entry
index v). It depends on st , affecting computation of lm , and
Fom as shown in Fig. 4. The standard deviation of the resulting
values from (14) as a function of st , v (st ), for a particular v
corresponding to a certain level of muscle activation is computed
across all trials of this muscle by

 K 1
1 
t
(ak ,v a
v )))2
(15)
v (s ) =
K

Fig. 6. The exoskeleton used for the experiments. The mini-PC and the power
supply are not shown here.

k =0

where a
v is the average of all ak ,v for a particular v and K is the
number of trials. The scale st now has to be optimized in such
a way that the average of the standard deviations for all EMG
values is minimized:

(st ) =

1
V

V
1

error function of the optimization is


E(A, Fom ) =

K
1 V
1


a(uk ,v )Fom Fkm,v

(17)

k =0 v =0

v (st ) min

(16)

v =0

where V is the number of elements in the tables. The minimization has been performed through subspace search with a
fixed step size of 0.01 for st within an interval of [0.85,1.25].
Refer to Fig. 5 and Section VI-A for examples of the geometry
calibration. The left diagram in this figure can be interpreted as
a visualization of all tables of a single muscle: the horizontal
axis shows the mapping from the postprocessed EMG signal
to an entry (points where values can be found), whereas every
discrete curve represents a trial and points correspond to the
ak ,v computed for the individual entries. Equation (15) can be
interpreted as the standard deviation of all points lying on a
vertical line above a certain activation.
As long as Fom is not known from prior calibration iterations,
the term for fP in (14) has to be omitted, and Fom is substituted
by a constant value greater than zero.
E. EMG-to-Force Calibration
The parameters A and Fom are determined by a twodimensional curve-fitting algorithm as shown in Fig. 5 (right)
to obtain a closed form of the activation function as of (1). The

where A is the EMG parameter in the activation function, uk ,v


is the postprocessed EMG signal from table k entry v, and Fkm,v
is the corresponding force share computed with the data from
the same entry. We have assumed 5 A < 0, 100 N< Fom <
2500 N, and performed a two-dimensional search.
After the EMG calibration is performed, the geometry calibration can be repeated with the same input data to refine the
torque distribution with the actual activation functions and to
incorporate the influence of the passive muscle force. But experiments presented in [35] show that this repetition does not
improve the results significantly. Additional iterations let the
resulting parameters oscillate in very small intervals around average values due to the fixed step sizes during the search.
V. EXOSKELETON HARDWARE
As can be seen in Fig. 6, the exoskeleton consists of an
orthosis that covers the right thigh, shank, and foot, and a linear
actuator that is able to perform knee flexion and extension. The
exoskeleton was manufactured according to the anatomy of one
particular subject to fit closely and is fastened with four straps
at thigh and shank.

FLEISCHER AND HOMMEL: HUMANEXOSKELETON INTERFACE UTILIZING ELECTROMYOGRAPHY

Fig. 7. Attachment of electrodes to the soft tissue of the orthosis for the RF,
VM, VL, SM, ST, and BF (left). Approximate sensor placement (right).

A. Actuation
The actuator consists of a ball screw driven by a 90 W Maxon
RE35 dc motor, and creates a maximum force of 1700 N and
a maximum linear velocity (without load) of 100 mm/s. The
force produced by the actuator is measured by a GS XFTC300
force sensor integrated into the tip of the actuator, and a Philips
KMZ41 Hall sensor, mounted on the knee joint, measures the
current joint angle.
B. EMG Sensors
Six Delsys 2.3 differential surface electrodes are integrated
in soft tissue of the thigh brace (refer to Figs. 6 and 7) and measure the muscle activations of three extensor muscles: the rectus
femoris (RF), vastus medialis (VM), vastus lateralis (VL), and
three flexor muscles: the semimembranosus (SM), semitendinosus (ST), and the biceps femoris (BF). The correct electrode
placement is very important (recommendations can be found
in [38], illustrated in Fig. 7, right), but has to be determined
only once for the operator. Since the electrodes are attached to
the exoskeleton and the position of the thigh brace is held fixed
by the axes of the knee and ankle joints, the electrodes are automatically placed at the same position on the skin whenever it is
donned. The electrodes are connected to a reference electrode
to minimize interferences from outside and have an inbuilt amplifier with a gain of 1000 V/V and a bandpass filter from 20 to
450 Hz.
C. Data Processing Unit and Safety
All sensors are digitized with converter circuits and connected
to a mini-PC with a Pentium-M 1.7 GHz via serial peripheral
interface (SPI) bus. Signals from all sensors are sampled with
1 kHz, downsampled and evaluated with 100 Hz. The low-level
torque control loop runs at a frequency of 1 kHz. A hardware
watchdog monitors the heartbeat of the computer and switches
to a hardware P-controller in case of failure. The P-controller
controls the actuator to not create force at the force sensor (to
avoid blocking a rescue movement). The weight of the exoskeleton without PC and power supply is approximately 5 kg.
VI. EXPERIMENTS AND DISCUSSION
All experiments have been performed with a single subject
(male, 32 years, 1.88 m, 82 kg). The subject was the author

877

himself and all experiments have been performed in accordance


to local ethical guidelines. The experiments are divided into
two categories: the first two experiments justify the model complexity and show the performance of the calibration. Values are
presented for individual experiments as well as averages over
a larger number of trials. After that, experiments showing the
exoskeleton giving support during everyday movements with
different support ratios are presented. An objective function to
quantify the performance of the system in this context is currently not available, because only the general task is known,
but not the exact desired trajectory. For these experiments, the
performance is evaluated with the following scheme: every experiment is presented with the knee angle trajectory and torque
curves of the operators contribution and the support from the
exoskeleton. The resulting knee angle trajectories with various
support ratios are compared to the trajectory without force support for the general shape, and it is searched for unusual dents
in the trajectories and corresponding muscle activations. These
dents are an indicator for unusual muscle activation that is amplified by the exoskeleton. These rapid changes in the muscle
activation patterns indicate that the operator is surprised by the
external force, or feels unwell and tries to counteract the force
produced by the exoskeleton.
The procedure of the experiments is described in the respective sections together with the discussion of the results.
A. Geometry Model Validation
This experiment is performed to analyze certain properties of
the biomechanical model (to justify the complexity), to evaluate the performance of the geometry calibration, and to obtain
values for the tendon slack length scales sti .
1) Procedure: The setup during this experiment is described
in Section IV-A. The isometric contractions of the extensor muscle have been performed at the following knee angles: 101 ,
86 , 70 , 47 , and 27 (the trials are labeled E0E4 in
the same order in Fig. 8). The flexor muscle trials have been
performed at 95 , 73 , 64 , 47 , and 33 (F0F4). In
each trial, the respective muscles have been activated four to
five times for a period of approximately 23 s.
The data evaluation has been performed in five versions: the
first version omits the musculotendinous model at all. The knee
torque contribution of a muscle is calculated by
m
ai (ui )
Ti = ri Fo,i

(18)

where ai (ui ) is the activation function of muscle i with the corresponding parameters. The second version includes the musculotendinous model, using the unmodified tendon slack lengths
from literature (sti = 1.0 for all muscles). The third version
includes the model without the passive force and without the
pennation angle. The fourth version only neglects the pennation angle, and the fifth version uses the complete model from
Section III. Table I shows the minimum averaged standard deviations as of (16) (dimensionless) for the different versions of
the biomechanical model.
2) Discussion: It can be seen that the complete model can
reduce the values by more than 50% compared to the absence

878

IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 4, AUGUST 2008

TABLE II
OPTIMIZED sti FROM EXPERIMENTAL SESSIONS WITH AT LEAST FOUR DAYS
IN-BETWEEN, AVERAGES (
sti ), AND STANDARD DEVIATIONS ( s t )
i

Fig. 8. Consistency of the EMG-to-muscle activation relationship for the RF


(top row), VM (middle row), and BF (bottom row) for different geometry
models. Left column: without the model, middle column: with the uncalibrated
model, right column: with the complete model (compare to Table I).

TABLE I
OF DIFFERENT VERSIONS OF THE GEOMETRY CALIBRATION
MINIMUM

the uncalibrated model, and the right column shows the complete
calibrated model for the RF (top row), VM (middle row), and
BF (bottom row). Obviously, the consistency is improved very
much, since in the right column, the curves have a very low
divergence. This is especially notable for the extensor muscles
and the trials where the leg is slightly flexed (trials labeled E3
and E4). A similar effect can be seen for the biceps femoris,
especially for trial F0 where the fibers are very short: without
the geometry model, muscle forces would be overestimated.
It can be concluded that it is important to include the geometry
model and to perform a proper parameter calibration.
B. Geometry Consistency Across Different Sessions
It has to be investigated if the calibrated geometry parameters
are indeed consistent across different sessions and independent
of the variable physical condition of the subject.
1) Procedure: To prove this, five experiments as described
in Section VI-A have been performed with at least four days
in-between. Results from these sessions, named S1 to S5, are
summarized in Table II. It shows the scale of each muscle for
every session, together with the average of the scales sti , and the
standard deviation of the scales s ti .
2) Discussion: It can be seen that the scales have a very low
deviation across the sessions. This further increases the trust in
the model. Similar results have been shown in [27].

of the model. This is especially important for the extensor muscles, since they are producing large forces during the considered
movements, and bad estimations result in a system behavior that
is hard to predict for the operator. The results also show that
inclusion of the musculotendinous model without a proper calibration can produce results far worse than without the model at
all. This is due to the strong effect of the forcelength relationship on the force output, and the sensitivity of the model to the
tendon slack length and the associated scale.
The improvement of the calibrated model compared to not
using the model can best be viewed with the muscle force plotted
over the postprocessed EMG signal, as shown in Fig. 8. The
left column shows the EMG-to-force relationship without the
musculotendinous model, the middle column shows results with

C. Sit-to-Stand Movement
This experiment is performed similar to the everyday movement of getting up from a chair, except that some additional
constraints have been applied to make the results more comparable. It is also used to determine a feasible low-pass frequency
for the EMG signal postprocessing.
1) Procedure: The movement was started in an upright sitting position. Getting up was performed without the arms holding onto something. Both feet were placed side by side and
the legs were parallel during the whole movement. The same
weight was put on both feet to minimize the influence of varying ground reaction forces. The movement was performed as
naturally as possible and as similar as possible to reduce the
effect of different postures on the required muscle activations.

FLEISCHER AND HOMMEL: HUMANEXOSKELETON INTERFACE UTILIZING ELECTROMYOGRAPHY

Fig. 9. Sit-to-stand experiment with different low-pass cutoff frequencies:


1.6 Hz (left), 4.0 Hz (right). From top to bottom, the support from the exoskeleton
is increased and the operator contributes less torque.

The low-pass cutoff frequency for the EMG postprocessing was


set for the first three trials to 1.6 Hz (Fig. 9, left column) and
for the second three trials to 4.0 Hz (Fig. 9, right column). Each
diagram shows the knee angle, the torque contributions of the
operators muscles estimated from the EMG, and the torque
produced by the actuator. The support ratio was set to 0.0 (top
row), 0.5 (middle row), and 1.0 (bottom row).
2) Discussion: It can be seen that the shape of the knee
angle trajectory is similar over all trials, indicating that the
desired movement could be performed successfully. A very important aspect is that with increased support ratio, the operator
can reduce the overall activation of his muscles (area under the
curve) and benefit from the support, although with differences in
quality: the joint trajectories in the right column (low-pass frequency 4.0 Hz) show obvious bumps with increasing support.
The movement has not been performed smoothly. This originates from the waviness of the EMG signals. Since the support
is directly depending on the prediction of the operators contribution, the resulting support is very unsteady. The mechanical
coupling of the exoskeleton with the operator leads to a feedback, which he tried to compensate unconsciously, resulting in
an undesired oscillation. In diagram 2c (see Fig. 9), the movement was performed more dynamic instead of energy reduced
compared to 2b, requiring slightly more muscle force.
This feedback effect can also be noticed in the first series
of trials in the left column: at first, the operator is activating
his muscles very strongly, out of long-learned experience, to a
degree that is required if no external support is given. This leads
to a very steep inclination of the knee trajectory in the initial
phase compared to the trajectory without support. As soon as
the locomotor system recognizes that the resulting movement

879

Fig. 10. Stair climbing experiment with different support ratios. It can be seen
that the operator can reduce his torque contribution with increased support from
the exoskeleton. Refer to the text for details on the interpretation.

is faster than desired, the muscle activation is reduced to slow


down the movement. Since this results in a decreased support
that is not considered by the locomotor system, the movement is
slower than desired that is countered by a following increase of
the muscle activation. The resulting knee joint trajectory shows
small bumps as a result of this untrained interaction between
the human and the exoskeleton. It is best seen in diagram 1c
of Fig. 9. This interaction is performed without conscious effort, and cannot be simply suppressed. Additional training may
reduce this effect.
For the following experiments, a low-pass cutoff frequency
of 1.6 Hz was utilized.
D. Stair Climbing
This experiment is very important because the cooperation of
the human with the machine can be analyzed in phases where
higher support from the exoskeleton can be contributed, as well
as smaller forces have to be applied to allow positioning the foot
over the next step. The torque curves and knee joint trajectories
from the experiments are shown in Fig. 10.
1) Procedure: The movement is performed similar to the
daily activity of climbing a stair without using a handrail or
similar support. The stair has four steps and the movement was
begun in an upright standing position and has been initiated
with the supported leg. The knee angle trajectory is explained
for the unsupported movement in the top row: at t < 2.0 s, the
subject is standing in front of the stair, both feet side-by-side.
The foot is raised from the ground until the following minimum
at t 3.5 s when the foot is at its highest point and brought
over the first step. At t 4.1 s, the foot is put onto the step,

880

IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 4, AUGUST 2008

and the operator is leaning forward to bring his weight over the
leading foot. The extensor muscles start to contract to push the
subject up the stair. The unsupported leg is raised and put onto
the second step (at t 6.0 s). The weight is moved over this
foot and the movement is repeated. The fifth step begins with a
flexion at t 9.5 s to raise the leg and to bring down the foot
onto the platform beside the other at t > 10.5 s.
2) Discussion: As can be seen, the support the exoskeleton
contributes to the movement gradually increases with the support ratio. For lower ratios, this support can be integrated into the
movement, and the muscle activity is reduced. Unfortunately,
for a ratio of 1.0, this support cannot be utilized comfortably.
The joint angle trajectories show many bumps, for example, at
t 3.0 s and t 7.5 s in the bottom diagram. These bumps
are a result of small muscle activities that can also be seen in
the other diagrams, but are amplified here. In these phases, the
supported leg has no floor contact and small variations in the
torques lead to large accelerations. The same variations have
less effect during phases with floor contact.
Small and short muscle activity seems to be hard to adapt
spontaneously to external influences, like the support of the
exoskeleton. Amplification of these stereotyped patterns results
in undesired effects. During push-up phase, this support could
be integrated more easily into the movement, and the muscle
activity could be reduced. It seems as if the locomotor system
is used to modulating the muscle activity during the push-up
phase because of different loads a human carries everyday. The
phase in which the leg is freely moved without external contact
rarely needs to be adapted.
E. Movement Combination
The last experiment shows the smooth transitions between
movements and the flexibility of the control algorithm.
1) Procedure: The support ratio was set to 0.5. It begins with
the subject sitting on a chair. He: 1) stands up; 2) walks three
steps with each foot; 3) climbs four steps of a stair; 4) turns
on the platform; 5) descends the stair; 6) walks three steps; 7)
turns; and 8) sits down (refer to Fig. 11).
2) Discussion: As can be seen, the transitions between the
movements appear smoothly, and significant support is contributed during the movements that require large muscle forces:
sit-to-stand, climbing the stairs up and down, and stand-to-sit
movements. In-between, the walking can be performed in a
quite normal fashion, but without the exoskeleton providing a
significant amount of torque.
We have recorded a video showing combined movements
together with output from the control system. This is available
at http://ieeexplore.ieee.org.
VII. GENERAL DISCUSSION
System behavior: The experiments presented earlier clearly
show the advantages of the simple control scheme coupled to
the biomechanical model as well as the limitations of this approach. Able-bodied users can take advantage of the flexible
support and are not limited to certain trajectories or movement
patterns. Since the human is inside the control loop, he or she can

Fig. 11. Combination of investigated movements. It can be seen that the


transitions between the individual movements appear very smoothly. Refer to
the text for details on the procedure of this experiment.

regulate the muscle activations to perform the desired movement


while receiving the support. Through feedback in the body, he
or she can take into account the external support and can reduce
the own force contribution to the movement when the actuator
gives support. But higher amplification factors demand lower
latencies of the system (exoskeleton and human) and the capability of the operator to adapt his or her normal movement
patterns to the modified circumstances quickly. The output of
the model was predictable enough for the operator in most cases
so that he could take the substantial external support from the
exoskeleton into account (unconsciously) and reduce his own
muscle activations while performing the desired movement. A
significant amount of torque was taken over by the exoskeleton,
although for high support ratios, the interaction between the exoskeleton and the operator became problematic if it resulted in
fast movements. In these cases, the wavy shape of the muscle
activations and the resulting joint trajectory are an indication
for the operator not feeling comfortable. The operator could not
benefit completely from the large support and was overreacting
to the feedback.
Limitations due to the human body: The overreaction of the
human to higher support was performed unconsciously and cannot be immediately suppressed with increased concentration on
the task. The latency of this reaction (signal propagation delay
inside the human body) can lead to oscillations that cannot be
avoided in general due to the integral part in the controlled process (acceleration to joint angle) and is depending on where the
reactions originate from in the human body. Most of them seem
to come from the lower levels of the locomotor system so that
only additional training may produce more suitable reactions. In
this context, it has to be investigated if there is a feasible alternative to modulating the operators estimated torque contribution
by multiplication with the support ratio to compute the support
torque. For example, it could be more desirable to have lower
support for smaller forces from the operator, which are mostly
used for fine-tuned movements that are rather disturbed through
additional support, but the support should be set extra large for
higher torques from the operator.

FLEISCHER AND HOMMEL: HUMANEXOSKELETON INTERFACE UTILIZING ELECTROMYOGRAPHY

Limitations due to the control algorithm: As this control


scheme has no knowledge about the movement that will result
from the combined effort of the operator and the actuation
(because no model of the world exists) and the desired trajectory
is also unknown, it is not easily possible to suppress the oscillations algorithmically. Furthermore, the simplicity of the general
concept is traded for the possibility to include algorithms that observe the global behavior to control postural stability or suppress
inappropriate movements. These topics are important when
considering the exoskeleton as a support device for disabled or
elderly people who may not be able to control postural stability
on their own. The control system heavily relies on actively
participating operators, but depending on the intended movement and the disabilities of a patient, this demand may be too
high.
Extensions: The experiments have been performed with rather
slow movements due to the limitation through the velocity of
the actuator. Further experiments have to show if more details
have to be integrated into the biomechanical model when it
should be used for more dynamic movements, as, for example,
the muscle activation dynamics (as described in [39]) modeling
the delay between neural activation and force production of the
muscle as reported, for example, in [19] and [20], or the force
velocity relationship, taking into account the length change of
the muscle fibers per unit of time (described, for example, in
[34]). Furthermore, in literature (for example, [27] and [30]), it
is reported that the optimal muscle fiber length changes with the
level of activation.
Calibration: The calibration algorithm is closely related to
model complexity. It has to calibrate 4 n parameters, where
n is the number of muscles included in the model. In contrast
to work in [27], for example, the algorithm can only use values from isometric measurements as described in Section IV-A,
to simplify the preparation time and procedure of the calibration (since it has to be performed whenever the exoskeleton is
donned). In this setup, many muscles are cooperating in a similar fashion making it impossible for an all-in-one optimization
algorithm to distinguish between the individual muscle contributions that are summed. But during different nonisometric
tasks, the muscles are behaving in different cooperation patterns
so that the muscles cannot simply be fused together in, for example, a single flexor and a single extensor muscle. For that
reason, the torque distribution has been implemented. Although
it is a very crude estimation, it served well to allow the determination of the individual parameters. A positive side effect is
that the algorithm becomes very fast due to splitting the whole
set of parameters into groups, reducing the parameter space for
the individual optimizations. Computation of the geometry calibration takes around 30 s (depending on the number of trials to
evaluate) and only a few seconds for the EMG calibration to be
performed on the mini-PC used for controlling the exoskeleton
(mainly consumed while reading the stored dataset).
VIII. CONCLUSION
In this paper, an exoskeleton system for supporting the operator with extra force in the knee joint is presented. The inter-

881

face between the operator and the exoskeleton is based on the


evaluation of electrical signals emitted by the muscles during
their activation. The control system is based on a force amplifying scheme that turned out to be robust enough for real-life
experiments, requiring only a minimum of sensors. This simplicity is traded for the possibility to include certain algorithms,
for example, to enforce postural stability, because the resulting
movement is not known in advance.
With our experiments, we have shown that a healthy operator
can successfully perform movements required for applications
as described in Section I in cooperation with the exoskeleton
while receiving substantial support.
For now, these experiments have only been performed for a
single healthy subject. But we are confident that it can be used
by other healthy subjects as well, because the individual components of the biomechanical model have been developed and
verified for a large number of subjects elsewhere in literature.
The algorithms have not been adapted to our particular operator
at any place.
We are looking forward to perform the experiments with other
subjects to have a closer look at the interaction of the human
and the exoskeleton that we expect to be different between
operators. Research will also uncover whether the support ratio
can be further increased for support functions other than the
linear and bring forward improvements of the overall safety as
well as investigations on how the algorithms have to be adapted
to apply them to disabled operators.
ACKNOWLEDGMENT
The authors would like to thank the company Otto Bock
Healthcare for support.
REFERENCES
[1] M. Vukobratovic, B. Borovac, D. Surla, and D. Stokic, Biped Locomotion:
Dynamics, Stability, Control, and Application. New York: SpringerVerlag, 1990.
[2] H. Kazerooni and R. Steger, The Berkeley lower extremity exoskeleton,
J. Dyn. Syst., Meas., Control, vol. 128, pp. 1425, 2006.
[3] H. Kawamoto and Y. Sankai, Power assist method based on phase sequence driven by interaction between human and robot suit, in Proc. 13th
IEEE Int. Workshop Robot Human Interactive Commun., 2004, pp. 491
496.
[4] G. Sawicki, A. Domingo, and D. Ferris, The effects of powered anklefoot orthoses on joint kinematics and muscle activation during walking in
individuals with incomplete spinal cord injury, J. Neuroeng. Rehabil.,
vol. 3, p. 3, 2006.
[5] S. Kawai, H. Yokoi, K. Naruse, and Y. Kakazu, Study for control of a
power assist device. Development of an EMG based controller considering
a human model, in Proc. 2004 IEEE/RSJ Int. Conf. Intell. Robots Syst.
(IROS), vol. 3, pp. 22832288.
[6] M. Ishii, K. Yamamoto, and K. Hyodo, Stand-alone wearable power
assist suit-development and availability, J. Robot. Mechatron., vol. 17,
no. 5, pp. 575583, 2005.
[7] K. Low, X. Liu, C. Goh, and H. Yu, Locomotive control of a wearable
lower exoskeleton for walking enhancement, J. Vib. Control, vol. 12,
no. 12, pp. 13111336, 2006.
[8] T. Nakamura, Z. Wang, K. Saito, and K. Kosuge, Wearable antigravity
muscle support system utilizing human body dynamics, Adv. Robot.,
vol. 20, no. 11, pp. 12371256, 2006.
[9] J. Rosen, M. Brand, M. Fuchs, and M. Arcan, A myosignal-based powered exoskeleton system, IEEE Trans. Syst., Man, Cybern., vol. 31, no. 3,
pp. 210222, May 2001.

882

[10] H. Kobayashi and H. Suzuki, A muscle suit for the upper bodydevelopment of a new shoulder mechanism, in Proc. IEEE Workshop
Adv. Robot. Social Impacts, 2005, pp. 149154.
[11] K. Nagai, I. Nakanishi, H. Hanafusa, S. Kawamura, M. Makikawa, and
N. Tejima, Development of an 8 DOF robotic orthosis for assisting human
upperlimb motion, in Proc. 1998 IEEE Int. Conf. Robot. Autom., vol. 4,
pp. 34863491.
[12] Y. Kim, J. Lee, S. Lee, and M. Kim, A force reflected exoskeleton-type
masterarm for humanrobot interaction, IEEE Trans. Syst., Man, Cybern.
A, Syst., Humans, vol. 35, no. 2, pp. 198212, Mar. 2005.
[13] N. Tsagarakis and D. Caldwell, Development and control of a softactuated exoskeleton for use in physiotherapy and training, Auton.
Robots, vol. 15, no. 1, pp. 2133, 2003.
[14] K. Naruse, S. Kawai, H. Yokoi, and Y. Kakazu, Development of wearable
exoskeleton power assist system for lower back support, in Proc. 2003
IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), vol. 4, pp. 36303635.
[15] K. Kiguchi, K. Iwami, M. Yasuda, K. Watanabe, and T. Fukuda, An
exoskeletal robot for human shoulder joint motion assist, IEEE/ASME
Trans. Mechatron., vol. 8, no. 1, pp. 125135, Mar. 2003.
[16] J. Pratt, B. Krupp, C. Morse, and S. Collins, The RoboKnee: An exoskeleton for enhancing strength and endurance during walking, in Proc.
2004 IEEE Int. Conf. Robot. Autom. (ICRA), vol. 3, pp. 24302435.
[17] H. Kazerooni, R. Steger, and L. Huang, Hybrid control of the Berkeley
lower extremity exoskeleton (BLEEX), Int. J. Robot. Res., vol. 25,
no. 5/6, pp. 561573, 2006.
[18] C. Fleischer and G. Hommel, Predicting the intended motion with EMG
signals for an exoskeleton orthosis controller, in Proc. IEEE/RSJ Int.
Conf. Intell. Robots Syst., 2005, pp. 20292034.
[19] P. Cavanagh and P. Komi, Electromechanical delay in human skeletal
muscle under concentric and eccentric contractions, Eur. J. Appl. Physiol., vol. 42, no. 3, pp. 159163, 1979.
[20] S. Zhou, D. Lawson, W. Morrison, and I. Fairweather, Electromechanical
delay in isometric muscle contractions evoked by voluntary, reflex and
electrical stimulation, Eur. J. Appl. Physiol., vol. 70, no. 2, pp. 138145,
1995.
[21] J. Chu, I. Moon, and M. Mun, A real-time EMG pattern recognition
system based on linear-nonlinear feature projection for a multifunction
myoelectric hand, IEEE Trans. Biomed. Eng., vol. 53, no. 11, pp. 2232
2239, Nov. 2006.
[22] H. Huang and C. Chen, Development of a myoelectric discrimination
system for a multi-degree prosthetic hand, in Proc. 1999 IEEE Int. Conf.
Robot. Autom., vol. 3, pp. 23922397.
[23] A. Ajiboye and R. Weir, A heuristic fuzzy logic approach to EMG pattern
recognition for multifunctional prosthesis control, IEEE Trans. Neural
Syst. Rehabil. Eng., vol. 13, no. 3, pp. 280291, Sep. 2005 (see also IEEE
Trans. Rehabil. Eng.).
[24] D. P. Ferris, K. E. Gordon, G. S. Sawicki, and A. Peethambaran, An improved powered ankle-foot orthosis using proportional myoelectric control, Gait Posture, vol. 23, no. 4, pp. 425428, Jun. 2006.
[25] H. Kawamoto and A. Sankai, Comfortable power assist control method
for walking aid by HAL-3, in Proc. 2002 IEEE Int. Conf. Syst., Man
Cybern., Ibaraki, Japan, vol. 4, p. 6.
[26] J. Rosen, M. Fuchs, and M. Arcan, Performances of hill-type and neural network muscle modelsToward a myosignal based exoskeleton,
Comput. Biomed. Res., vol. 32, no. 5, pp. 415439, 1999.
[27] D. Lloyd and T. Besier, An EMG-driven musculoskeletal model to estimate muscle forces and knee joint moments in vivo, J. Biomech., vol. 36,
pp. 765776, 2003.
[28] T. Koo, A. Mak, and L. Hung, In vivo determination of subject-specific
musculotendon parameters: Applications to the prime elbow flexors in
normal and hemiparetic subjects, Clin. Biomech., vol. 17, no. 5, pp. 390
399, 2002.
[29] J. Potvin, R. Norman, and S. McGill, Mechanically corrected EMG for the
continuous estimation of erector spinae muscle loading during repetitive
lifting, Eur. J. Appl. Physiol. Occup. Physiol., vol. 74, pp. 119132,
1996.
[30] T. Buchanan, D. Lloyd, K. Manal, and T. Besier, Neuromusculoskeletal
modeling: Estimation of muscle forces and joint moments and movements
from measurements of neural command, J. Appl. Biomech., vol. 20,
pp. 367395, 2004.
[31] A. Gordon, A. Huxley, and F. Julian, The variation in isometric tension
with sarcomere length in vertebrate muscle fibres, J. Physiol., vol. 184,
no. 1, pp. 17092, 1966.

IEEE TRANSACTIONS ON ROBOTICS, VOL. 24, NO. 4, AUGUST 2008

[32] L. Schutte, Using musculoskeletal models to explore strategies for improving performance in electrical stimulation-induced leg cycle ergometry, Ph.D. dissertation, Stanford Univ. Stanford, CA, 1992.
[33] S. Delp, J. Loan, M. Hoy, F. Zajac, E. Topp, and J. Rosen, An interactive
graphics-based model of the lower extremity to studyorthopaedic surgical
procedures, IEEE Trans. Biomed. Eng., vol. 37, no. 8, pp. 757767, Aug.
1990.
[34] F. Zajac, Muscle and tendon: Properties, models, scaling, and application
to biomechanics and motor control, Crit. Rev. Biomed. Eng., vol. 17,
no. 4, pp. 359411, 1989.
[35] C. Fleischer, Controlling exoskeletons with EMG signals and a biomechanical body model, Ph.D. dissertation, Institut fur Technische Informatik und Mikroelektronik, Technische Universitat Berlin. Berlin,
Germany, 2007.
[36] K. An, K. Takahashi, T. Harrigan, and E. Chao, Determination of muscle
orientations and moment arms, J. Biomech. Eng., vol. 106, no. 3, pp. 280
282, 1984.
[37] C. Fleischer and G. Hommel, Calibration of an EMG-based body model
with six muscles to control a leg exoskeleton, in Proc. IEEE Int. Conf.
Robot. Autom., 2007, pp. 25142519.
[38] H. Hermens, B. Freriks, R. Merletti, D. Stegeman, J. Blok, G. Rau, C. Disselhorst Klug, and G. Hagg, European Recommendations for Surface Electromyography: Results of the SENIAM Project. Enschede, The Netherlands: Roessingh Research and Development, 1999.
[39] F. Zajac and J. Winters, Modeling musculoskeletal movement systems:
Joint and body segmental dynamics, musculoskeletal actuation, and neuromuscular control, Multiple Muscle Systems, J. M. Winters and S. L.Y. Woo, Eds. New York, Springer-Verlag, 1990.

Christian Fleischer was born in Berlin in 1974. He


received the Diploma and Ph.D. degrees (with distinction) in computer engineering from the Technische
Universitat (TU) Berlin, Berlin, Germany, in 2001
and 2007, respectively.
He developed medical software and software for
the automotive industry. In 2003, he joined the RealTime Systems and Robotics Group, TU Berlin, where
he was engaged in research on control systems for
exoskeletons. His current research interests include
humanmachine interfaces, humanmachine interaction, and autonomous robots.

Gunter
Hommel received the Diploma degree
in electrical engineering and the Ph.D. degree in
computer science from the Technische Universitat
(TU) Berlin, Berlin, Germany, in 1970 and 1978,
respectively.
He joined the Faculty of Electrical Engineering,
TU Berlin, where he later joined the Faculty of Computer Science. In 1978, he joined the Nuclear Research Center, Karlsruhe, where he was engaged in
the field of real-time systems. During 1980, he was
the Head of a research group in Software Engineering, German National Research Center for Information Technology (GMD),
Bonn. In 1982, he was appointed as a Professor in the Department of Computer
Science, TU Munich, where he was engaged in the fields of real-time programming and robotics. Since 1984, he has been a Professor in the Department of
Computer Engineering, TU Berlin, where he heads the Real-Time Systems and
Robotics Group. In 2004, he was appointed as an Advisory Professor at Shanghai Jiao Tong University.
Prof. Hommel received an Honorary Doctorate (Dr. h.c.) from Moscow State
University of Instrument Engineering and Computer Science in 2002. In 2005,
he was nominated as the Director of the TU BerlinShanghai Jiao Tong University Research Laboratories for Information and Communication Technology,
Shanghai.

Das könnte Ihnen auch gefallen