HAL Id: hal-01302398
https://hal.sorbonne-universite.fr/hal-01302398
Submitted on 14 Apr 2016
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Upper-limb robotic exoskeletons for neurorehabilitation:
a review on control strategies
Tommaso Proietti, Vincent Crocher, Agnes Roby-Brami, Nathanael Jarrasse
To cite this version:
Tommaso Proietti, Vincent Crocher, Agnes Roby-Brami, Nathanael Jarrasse. Upper-limb robotic
exoskeletons for neurorehabilitation: a review on control strategies. IEEE Reviews in Biomedical
Engineering, 2016, PP (99), pp.1. �10.1109/RBME.2016.2552201�. �hal-01302398�
1
Upper-limb robotic exoskeletons for
neurorehabilitation: a review on control strategies
Tommaso Proietti, Vincent Crocher, Agn
`
es Roby-Brami, and Nathana
¨
el Jarrass
´
e
Abstract—Since the late 90s, there has been a burst of research
on robotic devices for post-stroke rehabilitation. Robot-mediated
therapy produced improvements on recovery of motor capacity;
however, so far, the use of robots has not shown qualitative benefit
over classical therapist-led training session, performed on the
same quantity of movements. Multi degrees of freedom robots,
like the modern upper-limb exoskeletons, enables a distributed
interaction on the whole assisted limb and can exploit a large
amount of sensory feedback data, potentially providing new
capabilities within standard rehabilitation sessions.
Surprisingly, most publications in the field of exoskeletons
focused only on mechatronic design of the devices while little
details were given to the control aspects. On the contrary,
we do believe a paramount aspect for robots potentiality lays
on the control side. Therefore the aim of this paper is to
provide a taxonomy of currently available control strategies
for exoskeletons for neurorehabilitation, in order to formulate
appropriate questions towards the development of innovative and
improved control strategies.
Index Terms—Upper-limb robotic exoskeletons, rehabilitation,
post-stroke robotic therapy, control strategies.
I. INTRODUCTION
Since the end of the nineties, there has been a burst of
research on and development of robotic devices for rehabil-
itation, particularly for the neurorehabilitation of post-stroke
patients. Stroke is indeed the second leading cause of death in
the world and the leading cause for acquired disability in adults
[1], [2]. Stroke survivors are usually left with disabilities,
mainly motor impairments on upper-limb movements and loss
of hand dexterity, both partially recoverable by undergoing
rehabilitation [3]. Rehabilitation has been proven to be effec-
tive when it is both intense and involving for the patient [4],
and robotics is one possible solution to provide intensity, by
increasing the number of repetitions which a therapist could
impose, as well as motivation thanks to technology appeal,
virtual reality, and gaming [5], [6].
The first robots used for upper-limb rehabilitation, also
referred as manipulanda, were only able to guide the motion
of patient’s hand in the plane (for example the MIT-Manus and
MIME platforms [7]). Extensive clinical testing on the InMo-
tion ARM© robot (the commercialized version of the MIT-
Manus) confirmed the improvements of the motor capacity of
the impaired arm after robotic therapy [8]. However, so far,
the use of robots did not show extra qualitative benefits (i.e.
T. Proietti, A. Roby-Brami, and N. Jarrass
´
e are with Sorbonne Universit
´
es,
UPMC Univ. Paris 06, CNRS, UMR 7222, INSERM, the Institute of Intelli-
gent Systems and Robotics (ISIR), 4 place Jussieu, 75005, Paris, France
V. Crocher is with Department of Electrical and Electronic Engineering,
University of Melbourne, Australia.
improved functional recovery), over conventional therapist-led
training session, performing the same quantity of motions [9],
[10].
The willingness to address 3D movements and to work at
the joint level encouraged the researchers to design robotic
exoskeletons for neurorehabilitation. These multiple degrees
of freedom (DOF) structures, allowing distributed physical
interaction with the whole limb, could potentially provide
new capabilities within standard rehabilitation sessions. Nu-
merous upper-limb exoskeletons have therefore been recently
developed (around the mid-2000s) but their effects have still
been little studied up to now, especially clinically. Indeed, the
first and only commercially available upper-limb exoskeleton
for rehabilitation (the ArmeoPower© by Hocoma, based on
ARMinIII robot) was only released at the end of 2011 [11].
The two major features of the robotic exoskeletons are their
abilities to apply forces distributed along the assisted limbs
and to provide reliable joint measurements [10]. Although the
physical interaction through multiple interaction points raises
interest and fundamental questions from the control point
of view, most of the existing control approaches have been
developed by only considering the end-point interactions.
While many reviews on robotic exoskeletons are available,
most of them are focused on their mechanical features (for
example [12], [13], [14]). We do believe that the key charac-
teristic of exoskeletons addressing neurorehabilitation stands
in their control strategies which, on top of intrinsic mechanical
behaviour of the devices (inertia, friction, backdriveability,
etc.), dictate the human-robot interactions. Reviews on high-
level control strategies for neurorehabilitation robots, including
both manipulanda and exoskeletons, exist ([15] and [16]), but
since these were not targeting the specificities of exoskeletons,
many devices and control approaches are missing.
The aim of this paper is therefore to disclose the real
interaction capabilities of upper-limb robotic exoskeletons for
neurorehabilitation, i.e. their control laws and physical inter-
action with humans, in order to give researchers in the field an
overview of existing works, their possibilities and limitations
along with possible orientations for future developments.
Review methodology
Our review, which rises searching in the main scientific
databases (in particular PubMed, ClinicalTrials, IEEE Xplore
Digital Library, Science Direct, and Google Scholar) different
combinations of some keywords (upper-limb, rehabilitation,
robot, exoskeleton, shoulder, elbow, wrist, arm, therapy, as-
sisted, training, stroke), moves through about 100 papers
2
days after stroke weeks after stroke months after stroke
neurorehabilitation
Passive mode
unresponsive
impaired limb
Active mode
responsive
impaired limb
acute phase chronic phasesub-acute phase
Fig. 1. General neurorehabilitation timeline. Robot controllers should adapt
to the correct phase of the therapy of the stroke survivor. Passive and Active
are traditional terms to indicate the involvement of the subject in the training.
published before March 2016. Besides, we also conducted a
free search in the above-mentioned databases for references
listed in keywords-based findings to include larger context. For
a purpose of feasibility, we focused our analysis on the subset
of minimum 3 degrees of freedom (DOF) actuated upper-limb
exoskeletons for rehabilitation, with at least a control over two
upper-limb joints among shoulder, elbow, and wrist. Besides,
this decision reflects the idea of characterizing the available
control strategies for multi-DOF devices, which apply forces
along the assisted limb rather then acting only at the end-
effector. Furthermore, a review including most of the planar
robots is already available in literature [16]. In the following
review, we briefly illustrate every available control solution,
providing more details on the ones which were clinically
tested.
II. AVAILABLE CONTROL STRATEGIES
32 robotic exoskeletons for upper-limb rehabilitation have
been listed, considering only the devices with 3-DOF or
more controlling the motions of minimum two out of the
shoulder-elbow-wrist joint groups (see Table I). Four devices
([54], [55], [56], [57]) composed by multiple robots acting
on multiple contact points were not included in this review.
Among the 32 exoskeletons, to the authors’ knowledge, 10
devices were tested on post-stroke impaired subjects, but only
6 test provided comparisons with impaired subject control
groups which tested different rehabilitation therapies. Besides
this, roughly one third of the existing structures seems not to
be currently subject of study, or rather, no results are appearing
for these devices on the main journals and conferences for at
least three years.
Inspired by the categories presented in [16], we defined
three main global strategies for robotic-mediated rehabilita-
tion: assistance, correction and resistance, see figure 2.
• Assistance means that the robot is supporting the weight
of the impaired limb and providing forces to complete
the task. If the patient does not produce any effort, task
completion can be still achieved depending on the level
of assistance.
• Correction defines the rehabilitation situation in which the
robot is only acting when the patient is not performing the
movement correctly, forcing the impaired limb to recover
a desired inter-joint coordination.
• Resistance represents the techniques in which the robot
opposes forces to the motion (potentially increasing the
current error, for example) in order to make the task more
DOF clinical
project name nationality year a p type pHRI testing
supported motion of shoulder-elbow-wrist
ARAMIS [17] Italy 2009 6 0 e 2-sfh [17]
ARMinIV [18] Switzerland 2007 7 0 e ufh [19]
ArmeoPower
1
[11] Switzerland 2011 6 0 e ufh
1
ARMOR [20] Austria 2008 8 4 e 2-uffh [20]
BONES+SUE [21] USA 2008 6 0 p ufh [22]
BOTAS [23] Japan 2013 6 0 e 2-ufh
ETS-MARSE [24] Canada 2010 7 0 e ufh
EXO-UL7
2
[25] USA 2011 7 0 e 2-ufh [25]
IntelliArm [26] USA 2007 7 2 e ufh
NTUH-ARM [27] Taiwan 2010 7 2 e uf
RUPERT IV [28] USA 2005 5 0 p sufh
SRE [29] UK 2003 7 0 p fh
SUEFUL 7 [30] Japan 2009 7 1 e uffh
supported motion of shoulder-elbow
- [31] France 2009 4 0 e uf
- [32] New Zealand 2014 5 0 e uh
ABLE [33] France 2008 4 0 e uf [34]
ALEx [35] Italy 2013 4 2 e ufh
AssistOn-SE [36] Turkey 2012 6 1 e ufh
CAREX [37] USA 2009 5 0 e suf
CINVESRobot-1 [38] Mexico 2014 4 0 e uf
L-Exos [39] Italy 2002 4 1 e ufh [40]
LIMPACT [41] Netherlands 2008 4 6 h uuff
MEDARM [42] Canada 2007 6 0 e uf
MGA [43] USA 2005 5 1 e uh
MULOS [44] UK 2001 5 0 e uff
Pneu-WREX [45] USA 2005 4 0 p ufh [46]
RehabExos [47] Italy 2009 4 1 e ufh
supported motion of elbow-wrist
6-REXOS [48] Sri Lanka 2015 4 2 e fh
MAHI EXO-II [49] USA 2006 5 0 e ufh
3
MAS [50] Japan 2008 4 0 p ufh
ULERD [51] Japan 2013 3 4 e uufh
Wrist Gimbal [52] USA 2013 3 0 e fh
TABLE I
EXOSKELETONS FOR UPPER LIMB REHABILITATION (3-DOF SYSTEMS
CONTROLLING AT LEAST TWO JOINTS OUT OF THE SHOULDER-ELBOW-WRIST
CHAIN). DOF: A-ACTIVE THUS ACTUATED, P-PASSIVE THUS MECHANICAL ONLY.
TYPE OF ACTUATION: E-ELECTRICAL, P-PNEUMATIC, H-HYDRAULIC. PHYSICAL
HUMAN-ROBOT INTERFACE (FIXATION LEVELS) : 2-TWO ARM EXOSKELETON,
S-SHOULDER, U-UPPER ARM, F-FOREARM, H-HANDLE. DOUBLE LETTERS INDICATES
DOUBLE INTERFACES.
1
BASED ON ARMINII, THE ONLY COMMERCIALIZED
EXOSKELETON FOR THE CLINICAL ENVIRONMENT.
2
BASED ON CADEN-7 (2006)
[53].
3
ONGOING TEST, CLINICALTRIALS.GOV IDENTIFIER: NCT01948739.
complex for the subject, and to train his ability to correct
the movement and to adapt to external perturbations.
Contrary to assistance, correction does not assist the patient
in achieving the task. Obviously, pure correction is an ideal
case of neurorehabilitation therapy, as well as the former cate-
gorization. More often, the therapy involves several strategies
combined [10].
While resistive controllers exist for manipulanda (for ex-
ample, resistive force-field in [58], or error augmentation in
[59]), to the authors’ knowledge, to date there are no available
resistive controllers developed for exoskeletons. This could
be a consequence of the fact that exoskeletons often target
early stage chronic patients who rarely have recovered enough
motor capabilities to undergo resistive therapies. However,
most control strategies developed for manipulanda could be
translated to exoskeletons, by considering only the end-effector
control; though this solution would not take the full advantages
of dealing with multi-contact systems like the exoskeletons.
Existing controllers for exoskeletons are mostly combined
3
Resistive
Assistive
Task
Corrective
Assistive mode
1. Passive control
2. Triggered passive control
3. Partially assistive control
Passive trajectory tracking, Passive mirroring, Passive stretching
Impedance/Admittance control, Attractive force-field,
Model-based assistance, Offline adaptive control
Corrective mode
1. Tunneling
2. Coordination control
Resistive mode
no controls developed yet
Fig. 2. The three global strategies for robotic-mediated rehabilitation and the current implementations on robotic exoskeletons.
assistive-corrective controllers but there exists a variety of
implementations of each category by using different control
techniques.
A. Assistive modes
Within this type of control approaches, three groups of assis-
tive strategies were determined — passive, triggered passive,
and partially assistive control — although there exist solutions
which are mixed strategies. The distinction between passive
and partially assistive controls is thin, since if the patient
is not participating in the task, both controllers would react
similarly. Also, triggered assistance generally refers to solution
only different to initiate the assistance: once triggered, these
controllers usually exhibit either passive or partially assistive
behaviours.
1) Passive control: The simplest way to control an ex-
oskeleton is to control the motion rigidly along a desired
reference trajectory through position feedback control with
high corrective gains. In rehabilitation this passive technique
is common at early stages of the post-stroke therapy, when
the impaired limb is usually unresponsive, and thus passive
mobilization is the only feasible solution to achieve any result.
Nonetheless the feedback controller gains have to be tuned
carefully and the exoskeleton has to show a minimum of
compliance not to hurt the subject in the presence of trajectory
errors due to excessive muscle contraction, spasticity, or other
pathological synergies. Such compliance between the robot
and the human body can also be introduced mechanically,
at the fixation level with, for example, elastic straps [28] or
mechanical fuses placed serially [57].
. Passive trajectory tracking: Passive control can be
achieved by adopting different techniques. The simplest is
the use of a proportional-integral-derivative (PID) feedback
control which usually regulates the position or the interaction
force along a known reference (for example, a trajectory or a
force field model), and can be applied either at the joint or at
the end-effector level. Examples of these joint controllers are
shown in [38], [52], [60], [43], [61], [47], [62], [41].
More complex approaches have been recently developed to
improve the quality of the physical interaction during passive
mobilization of patient’s limb. In [24], the authors developed
the Sliding Mode Control with Exponential Reaching Law
(SMERL), a non-linear control technique which minimizes the
tracking error on a state space projection. In [63] a fuzzy
logic technique is adopted. In order to deal with uncertain-
ties and disturbances from the environment, adaptive fuzzy
approximators estimated the dynamical uncertainties of the
human-robot system, and an iterative learning scheme was
utilized to compensate for unknown time-varying periodic
disturbances. Preliminary results on healthy people showed
better performances, in terms of tracking error and average
control input, compared to classical PID control and fuzzy
logic techniques used alone.
Different methods exist for defining reference trajectories.
For passive strategies, these references are often created by
recording the physiotherapist inputs on the subject limb at-
tached to the exoskeleton, during a teaching phase. In this
phase, the robot is generally set in a transparent mode (usually
achieved by adopting a feedforward term to compensate for the
gravity and the dynamics of the robot) to limit any resistance
to perform motions. Once recorded, the exoskeleton is ready to
replay the trajectory with its feedback controller. Teach-and-
replay is presented in [64]. In [60] teach-and-replay mode was
tested on four chronic stroke patients during 8 weeks of train-
ing. Encouraging preliminary results were obtained (increase
of Fugl-Meyer Assessment score – FMA, meaning improve-
ment of motor functionality of the paretic arms, in addition
to positive transfer to ADLs). Nevertheless, no comparisons
with control group receiving traditional therapy were provided.
Similarly to teach-and-replay, record-and-replay uses healthy
limb motions, recorded within the robot, to create reference
trajectories. This strategy was for example used in [17].
Instead of using external inputs or “healthy” trajectories
as references, some research groups have been trying to
directly detect patient motion intention by measuring muscle
activity through surface-EMG. In [65] EMG and two offline-
trained time delayed neural networks (one for the shoulder
and one for the elbow) were used to estimate and predict the
resulting torques at the joint level. By using this prediction, a
reference joint position trajectory was computed and fed to a
PD controller with gravity and friction compensation. In [66]
4
an EMG-based version of the SMERL controller was provided.
In this case, EMG signals, transformed to a desired position
reference through a muscle model, acted as the reference for
the position feedback controller.
A third strategy to define reference trajectories is to de-
termine cost functions and to use optimization algorithms.
Generally the cost function aims at minimizing the jerk. This
solution is mostly used for triggered passive control (for
example in [23]) or for partially assistive controls (examples
in [67] or [68]). Due to the complexity and variety of pos-
sible actions of the upper-limb, generally the movements of
intermediate joints either occur as a consequence of the end-
effector movement in the task space (for example in [37]),
or are constrained along specific dedicated trajectories, which
are synchronized within the joint space (for example in [69]).
However, because of the redundancy of human arm, there are
no commonly accepted solution to compute joint trajectories
from task space motions [70].
Beside standard passive control approach, two other passive
strategies exist. Those either target two-arm exoskeletons
(passive mirroring) or utilize a different rehabilitative training,
i.e. the passive stretching.
. Passive mirroring: Few exoskeletons have two arms
(four devices in Table I, column pHRI). For these robots, pas-
sive mirroring is a further strategy. It consists of synchronous
passive mimicking of the behaviour of the healthy limb, in
a master-slave configuration, as in [17] or in [20]. In [17] a
clinical pilot study on 14 impaired subjects (21±6 days since
stroke), based on an average of 31 sessions over 2 months,
resulted in applicability and tolerability of the robot-mediated
treatment, with similar improvements to other robot-assisted
rehabilitation therapies. In [71] passive-mirroring was imple-
mented on a single-arm robot in a master-slave configuration
with an external haptic device to control the master side of
the system.
. Passive stretching: A different rehabilitative passive
training is the so-called passive stretching [26]. With this
strategy, individual joints were passively stretched by the robot
in order to identify their individual angle-resistance torque
relationships. These relationships were then used to coordinate
the passive stretching of multiple joints together. Feasibility
test performed in 3 stroke patients showed a reduction of cross-
coupled stiffness after a 40 min stretching session.
2) Triggered passive control: A slight variation of assistive
modes consists in approaches in which the user triggers
the exoskeleton assistance. This technique is frequently used
to introduce brain-machine interface (BMI) into the control
loop. These approaches are directly adapted from the field of
assistance to patients with non-recoverable impairments (like
tetraplegia). In fact, after the triggering (mainly a selection
of available targets), the exoskeleton is usually controlled
passively along pre-determined trajectories.
In [23] passive recorded-trajectory replication through BMI
control is shown. The end-effector trajectory were computed
by minimum-jerk optimization. BMI trajectory-replay trig-
gering was implemented through SSVEP (steady-state visual
evoked potential). SSVEP can be observed mainly from the
visual cortex when a person is focusing his visual attention
on a flickering stimulus. Pre-clinical testing (12 healthy and
3 upper cervical spinal cord injured subjects) were performed
to show the capacity of the system to be used by subjects to
activate different trajectory reproductions.
In order to extend patient’s control over the task, gaze track-
ing methods can also be used in addition to BCI-driven control
like in [72]. An eye-tracker together with a target-tracking
module (a Kinect camera) gave the reference position of the
target to reach to the exoskeleton controller; the BCI module
estimated user’s motion intention, modulating maximum joint
jerk, acceleration, and speed within an admissible predefined
set of values. Then a PD feedback control (helped by classical
gravity and friction compensations) performed the reaching
task. Pre-clinical testing with 3 healthy and 4 chronic stroke
patients showed that all subjects were able to operate with the
exoskeleton.
In [73] a Motor Imagery based Brain Computer Inter-
face (MI-BCI) is used to control an eight DOF exoskeleton.
Three chronic stroke survivors were able to perform passive
controlled tasks (arm motion towards a target, grasping and
releasing of an object) by producing MI of the reaching task
before the movement. Two seconds of MI activity collected,
triggered the exoskeleton to replay pre-recorded trajectories.
In [74] authors developed a similar approach, by controlling an
ArmeoPower exoskeleton through MI-BCI. 9 healthy subjects
and 2 stroke survivors were able to move the exoskeleton along
pre-defined trajectories, and these motions were brain-state
dependent, meaning that motion was performed only during
MI phases.
3) Partially assistive control: The effectiveness of pure
passive motions for stimulating neuroplasticity is known to
be limited [75], since the patient is not involved in making
any effort to perform the task. On the other side, assistance
is needed in order to reduce failures at least at the beginning
of the therapy, thus maintaining subject motivation, intensity
of the training, confidence in using the affected limb, and to
avoid negative reinforcement [6]. But as soon as the patient has
recovered a minimal amount of motor capacity, it is essential
for the robot to allow for shared control of the movements [76],
[77]. Indeed, as neurorehabilitation addresses issues related
to motor control, the devices must allow patients to express
whatever movement they can without suppressing any motor
capability [78].
Partial assistance, or assistance-as-needed [67], groups all
those control strategies which allow the impaired subject to
actively control the motion, supporting it based on perfor-
mance indexes. The most common solution to provide a partial
assistance to the impaired limb, is to increase the compli-
ance of the above-mentioned passive controllers. Instead of
rigid “industrial-type” position feedback controllers (with high
corrective gains), many controllers for exoskeletons rely on
more flexible impedance control [79], or its dual admittance
control, with reduced corrective gains to exhibit “human-like”
mechanical properties. These controllers allow to implement a
good compromise between tracking skills and compliance of
the robotic arm.
. Impedance/Admittance control: Impedance control is a
model-based force controller with position feedback while its
![TABLE I EXOSKELETONS FOR UPPER LIMB REHABILITATION (3-DOF SYSTEMS CONTROLLING AT LEAST TWO JOINTS OUT OF THE SHOULDER-ELBOW-WRIST CHAIN). DOF: A-ACTIVE THUS ACTUATED, P-PASSIVE THUS MECHANICAL ONLY. TYPE OF ACTUATION: E-ELECTRICAL, P-PNEUMATIC, H-HYDRAULIC. PHYSICAL HUMAN-ROBOT INTERFACE (FIXATION LEVELS) : 2-TWO ARM EXOSKELETON, S-SHOULDER, U-UPPER ARM, F-FOREARM, H-HANDLE. DOUBLE LETTERS INDICATES DOUBLE INTERFACES. 1BASED ON ARMINII, THE ONLY COMMERCIALIZED EXOSKELETON FOR THE CLINICAL ENVIRONMENT. 2BASED ON CADEN-7 (2006) [53]. 3ONGOING TEST, CLINICALTRIALS.GOV IDENTIFIER: NCT01948739.](/figures/table-i-exoskeletons-for-upper-limb-rehabilitation-3-dof-3e8kpwun.webp)