Articles

Control Strategies and

Artificial Intelligence in
Rehabilitation Robotics

Domen Novak, Robert Riener

n Rehabilitation robots physically sup- s robotics moved from industrial to service applica-
port and guide a patient’s limb during
motor therapy, but require sophisticated
control algorithms and artificial intelli-
gence to do so. This article provides an
A tions, engineers began looking for new tasks that could
be automated with robots. Industrial tasks had been a
perfect candidate for automation since they are physically
exhausting and require high precision. Motor rehabilitation
overview of the state of the art in this
area. It begins with the dominant para- seemed like a similarly appropriate robotics application. In
digm of assistive control, from imped- the course of rehabilitation, the patient must exercise by per-
ance-based cooperative controller forming limb motions thousands of times, and the therapist
through electromyography and inten- must physically support and guide the patient’s limb during
tion estimation. It then covers chal- these motions. Since therapists inevitably get exhausted, a
lenge-based algorithms, which provide
rehabilitation robot could support and guide the limb
more difficult and complex tasks for the
patient to perform through resistive con-
instead.
trol and error augmentation. Further- Numerous rehabilitation robots have been designed for
more, it describes exercise adaptation both the upper (figure 1) and lower limbs (figure 2). The two
algorithms that change the overall exer- most famous arm rehabilitation robots are the MIT-MANUS,
cise intensity based on the patient’s per- now sold as the InMotion ARM (Interactive Motion Tech-
formance or physiological responses, as nologies, USA) and the ARMin, now sold as the ArmeoPower
well as socially assistive robots that pro-
vide only verbal and visual guidance.
The article concludes with a discussion
of the current challenges in rehabilita-
tion robot software: evaluating existing
control strategies in a clinical setting as
well as increasing the robot‘s autonomy
using entirely new artificial intelligence
techniques.

Copyright © 2015, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602 WINTER 2015 23
Articles

Figure 1. The MIT-MANUS and the ARMin.

MIT-MANUS (left); ARMin (right). MIT-MANUS photo courtesy of H. I. Krebs, Massachusetts Institute of Technology.

(Hocoma AG, Switzerland). The most famous leg complete a motion within a desired time while cor-
rehabilitation robot is the commercially available recting any major errors (such as large deviations
Lokomat (Hocoma AG, Switzerland), with another from an optimal trajectory). The main characteristic
notable example being the Gait Trainer (Reha-Stim, of modern assistive controllers is that they only help
Germany). All of these, and many other robots, were as much as it is necessary for the patient to complete
developed in order to support and guide the patient’s a motion, an approach called patient-cooperative
limbs. However, appropriate hardware is not enough; control (Riener et al. 2005). This is similar to the work
both therapists and robots need to intelligently adapt of therapists in rehabilitation: they manually move
their support to ensure proper exercise. Mistakes the patient’s limb to accomplish a desired motion,
should be corrected, but the patient should exercise but let the patient move on his or her own whenev-
actively and intensely, so the support should not be er possible.
excessive. As summarized by Marchal-Crespo and Reinkens-
The first rehabilitation robot controllers did not meyer (2009), many rationales have been given for
adapt their support to the patient at all. They were such assistive controllers. Aside from allowing
very stiff, and essentially guided the patient’s limbs patients to perform more movements in a shorter
along a predefined trajectory with little care for what amount of time, they interleave active effort by the
the patient was doing or wanted to do. Clinical tests participant with stretching of the muscles and con-
found that patients put significantly less effort into nective tissue, they provide novel somatosensory
robot-aided exercise with such controllers than into stimulation that helps induce brain plasticity, and
therapist-aided exercise, and frequently just let the they may help teach patients to perform demon-
robot move their passive limbs without actively par- strated patterns. Although most of these rationales
ticipating in the motion (Israel et al. 2006, Ziherl et have not been extensively clinically verified (Mar-
al. 2010). This ”slacking“ process leads to slower neu- chal-Crespo and Reinkensmeyer 2009), assistive con-
romotor recovery (Casadio and Sanguineti 2012). To trol algorithms remain dominant, particularly
avoid it, the robot needs to adopt a control strategy impedance-based control.
that assists the patient only as needed: a cooperative
control strategy. Impedance-Based Assistance
The cooperative principle of impedance-based con-
Help Me Help You: Cooperative trollers is as follows: while a patient is moving along
a desired trajectory, the robot does not intervene, but
Assistive Control it corrects deviations from this trajectory by applying
Assistive controllers are the dominant control para- a force to the patient’s limb. This correcting force is
digm in rehabilitation robotics, and are used in the generated with a mechanical impedance. The first,
majority of commercial systems. They operate on the simplest controllers provided proportional position
level of the individual motion, helping the patient feedback: as the patient’s limb moves farther from the

24 AI MAGAZINE
 Articles

Figure 2. Lokomat and Gait Trainer.

The Lokomat (left) and Gait Trainer (right) leg rehabilitation robots. Photos courtesy of Hocoma AG, Switzerland, and Reha-Stim, Germany.

desired trajectory, the robot applies a proportionally patient. Several algorithms have been developed to
stronger force to the limb. Such force feedback is adapt reference trajectories automatically (Jezernik,
often combined with visual feedback that informs Colombo, and Morari 2004) or to make the trajecto-
the patient how he/she should move instead. This ries of certain limb joints more compliant than those
generally also requires a deadband around the trajec- of other joints (Stauffer et al. 2009). Furthermore,
tory so that the patient can make small deviations alternative impedance-generation techniques have
without being disturbed. The end effect feels some- been investigated, such as virtual objects (Ekke-
lenkamp et al. 2007). These are shapes generated by
what like a tunnel that the patient needs to follow.
the robot’s haptic interface; if the patient attempts to
An additional assistive force, sometimes dubbed a
move a limb into the physical space where the virtu-
“moving wall” or “flow force,” can push the patient
al object is, the haptic interface will push the limb
along the trajectory if he/she is too slow (relative to back, creating the illusion of an object. Such virtual
a desired velocity profile), providing another type of objects can thus physically support the patient in
assist-as-needed control. An example of such assist- reaching a goal.
as-needed control is shown in figure 3 for arm reha- Movements in rehabilitation should also be self-
bilitation. initiated for better motor learning (Lotze et al. 2003).
A major problem with such reference trajectories is Many impedance controllers, therefore, employ trig-
that they are difficult to adapt to an individual gered assistance: the robot is entirely passive until the

WINTER 2015 25
Articles

kg
bt
kt basket basket

object being carried

reference trajectory reference guidance point

tunnel wall placing point

Figure 3. The Virtual Tunnel and Moving Wall in an Arm Reaching Task.
A spring-damper system prevents the patient from deviating from the reference trajectory (left) while a second spring-
damper system (right) leads the patient along the trajectory (adapted from Mihelj et al. [2012]).

patient has started moving the limb. Once a motion robot apply torque to a joint proportionally to the
has begun, the robot uses an assistive force and push- EMG of the muscle used to control the corresponding
es the patient’s limb toward the desired position, but joint of the human limb (Song et al. 2008, Stein et al.
only if the patient is not moving sufficiently quickly 2007). This avoids the problem of reference trajecto-
or smoothly on his/her own. If the patient’s own ries, which are difficult to adapt to an individual
movement is adequate, the robot becomes passive patient and may thus constrain him/her: since EMG-
again. A more complex possibility is the assist-as- based assistance augments the patient’s own move-
needed paradigm where the assistance parameters ments, it does not provide any constraints.
(for example, robot stiffness, assistive force gain) are This lack of constraints, however, presents a differ-
adapted over a continuous interval based on recent ent weakness. Patients in motor rehabilitation do not
motor performance (Wolbrecht et al. 2008). When only exhibit voluntary motions, but also pathologi-
performance is high, assistance is gradually decreased cal movements such as tremor and spasms. Patho-
due to a forgetting factor in the control algorithm. logical EMG is not easily separated from voluntary
Such assist-as-needed control has been found to sig- EMG, leading to potential augmentation of patho-
nificantly improve motor recovery compared to clas- logical motions. For this reason, some authors have
sic impedance-based control (Cai et al. 2006) or con- suggested that EMG may not be a suitable control sig-
ventional therapy (Reinkensmeyer et al. 2012). nal with patient populations such as stroke (Cesqui et
al. 2013). Other studies, however, have demonstrated
Electromyography-Based Assistance significant benefits of training with EMG-controlled
Surface electromyography (EMG) is the measurement robots (Song et al. 2013). It is likely that the use of
of a muscle’s electrical activity from the surface of the EMG in rehabilitation is appropriate for some
limb. By measuring the activation of different mus- pathologies and muscles but not others, and this
cles, a robot could provide precise assistance to dif- should be investigated further.
ferent parts of the limb. The first implementation
was EMG-triggered assistance, where the robot begins Assistance Based on Other Information
providing assistance once sufficient EMG activity is Impedance-based controllers provide assistive forces
detected (Krebs et al. 2003). A more advanced coop- based on motion of the impaired limb while EMG-
erative control approach is to augment the activity of based controllers provide them based on muscle
individual muscles: for example, have an exoskeletal activity on the impaired limb. However, assistance

26 AI MAGAZINE
 Articles

Figure 4. Bimanual Training with the HapticMaster Robot.

Photo courtesy of Matic Trlep, University of Ljubljana.

can also be provided based on information from oth- cyclic process, it should be possible to predict the
er parts of the body: the motion of the unimpaired motion of one leg based on the motion of the other
limb, the eyes, and even brain activity related to leg. If only one leg is impaired due to hemiparesis,
movement. The exploitation of this information is the healthy motion pattern for the impaired leg
sometimes referred to as intention detection, though could also be extracted from the motion of the
the term remains contested. It remains largely exper- unimpaired leg, and a rehabilitation robot can then
imental, but has great potential for future rehabilita- provide impedance-based assistance based on this
tion robots. estimated healthy pattern.
Complementary Limb Motion Estimation In principle, since the entire body is coordinated
Complementary limb motion estimation (Vallery et during gait, the motion of the rest of the body can
al. 2009) is an approach mainly used for the lower also be used to more accurately generate a motion
limbs. Essentially, since human gait is a coordinated pattern for the impaired leg. A recent study has thus

WINTER 2015 27
Articles

Figure 5. EEG Combined with the ARMin Rehabilitation Exoskeleton.

Photo from the authors’ joint research with José del R. Millán and Tom Carlson, Ecole Polytechnique Federale de Lausanne, Switzerland.

used sensors embedded in a walking cane to control 4), the motion again becomes coordinated. In such
a mobile leg exoskeleton with such complementary bimanual training, it is then possible to apply coop-
motion estimation (Hassan et al. 2014). erative control by measuring the forces applied to the
Bimanual Training object by the unimpaired arm and replicating them
The motion of both arms is generally not as coordi- on the impaired side using a rehabilitation robot
nated as the motion of the legs during gait, so com- (Lum et al. 2006). The gains achieved in such biman-
plementary limb motion estimation cannot be ual training have been shown to transfer to other,
directly transferred to the arms. However, if both unimanual motions, though at least some motion of
arms are used to manipulate the same object (for the impaired arm is necessary for training (Trlep,
example, lift a box or turn a steering wheel — figure Mihelj, and Munih 2012).

28 AI MAGAZINE
 Articles

Electroencephalography moves, experts improve by overcoming increasingly

If the patient’s arm is completely paralyzed, there is challenging opponents and obstacles. This has led to
no way to perform cooperative control using motion the development of alternative control algorithms
measurements. While the robot could simply move where the robot challenges rather than assists the
the patient’s arm to achieve some passive exercise, patient.
this does not achieve efficient motor learning; it
would be much better to perform self-initiated move- Vive la résistance: Resistive Control
ments. Luckily, even completely paralyzed patients Resistive control algorithms, as the name suggests,
can think about moving their arm, even if no motion generate forces that constantly resist any movements
is achieved. If we can detect these thoughts, we can that patients try to make, forcing them to apply a
trigger the robot to move in response, achieving self- larger force in order to perform the motion. In sports,
initiated exercise. an analogous approach would be weight lifting: the
While reading thoughts may sound like science fic- weights generate a resistive force. Such resistive train-
tion, brain activity can be noninvasively measured ing has been proposed for a variety of arm and leg
using electroencephalography (EEG): measurements rehabilitation robots (Lam et al. 2008, Lambercy et al.
of the brain’s electrical activity along the scalp. Since 2007), but until recently did not achieve widespread
it is known which regions of the brain are responsi- use in commercial robots. However, this is likely to
ble for motor planning, increased EEG activity in change, as a recent multicenter clinical study (Klam-
these regions would indicate that the patient wants roth-Marganska et al. 2014) has shown that better
to perform a motion even if he/she is completely par- strength training is sorely needed in arm rehabilita-
alyzed. While EEG does not allow precise motor tion robots — a perfect opportunity for resistive con-
intentions to be identified, it has been used to pro- trol.
vide triggered assistance: once the rehabilitation
robot detects increased EEG activity, it activates assis- It’s Not A Bug, It’s A Feature:
tance along a predefined trajectory. An example pho- Error Augmentation
to of EEG in an arm exoskeleton is shown in figure 5. Unlike resistive control, which aims to improve
Such EEG-triggered assistance has actually been strength, error augmentation aims to improve coor-
shown to lead to significantly better rehabilitation dination and precision. It does this by identifying
outcome than equivalent amounts of robotic assis- and magnifying the subject’s deviations from a
tance applied at random times (Ramos-Murguialday desired movement trajectory. This can be done by
et al. 2013). having the robot push the patient with a disturbing
Advanced prototypes of rehabilitation exoskele- rather than assistive force, requiring him/her to over-
tons have combined EEG triggers with additional come the disturbances to achieve the goal (Patton et
sensors that indicate the type of motion to be per- al. 2006). Alternatively, the visual feedback given to
formed. For example, the approach of Frisoli et al. the patient can be distorted by, for example, intro-
(2012) uses an eye tracker to measure what object the ducing a rotation between physically performed and
patient is looking at. Once increased EEG activity is visually displayed movements (Patton et al. 2013).
detected, the robot begins assisting a movement This forces the patient to learn the mapping between
toward that object. However, the use of EEG has met physical and displayed movements by trial and error.
with some skepticism in the rehabilitation commu- Error augmentation and visual distortions in par-
nity due to a relatively long setup time and perceived ticular may seem counterintuitive at first, as it is not
unreliability of brain activity measurements, and immediately obvious how forcing patients to over-
more studies are needed to show that the clinical use- come such additional challenges is helpful for reha-
fulness of EEG outweighs its drawbacks. bilitation. However, error-driven learning has been
emphasized as crucial to learning skills in human
motion. Furthermore, a similar principle of error-dri-
Challenge-Based Control ven learning is seen in artificial learning systems such
Assistive control algorithms have been criticized by as neural networks. Studies with healthy subjects
studies that suggest that physically guiding the have indeed shown that error augmentation leads to
motion during a task can actually reduce motor better, more robust movement patterns than assistive
learning, a phenomenon referred to as the guidance control (Patton et al. 2006). Improved coordination
hypothesis (Marchal-Crespo and Reinkensmeyer has been shown to persist even when the error aug-
2008). While severely impaired patients definitely mentation is removed.
require physical guidance, patients with a lower lev- Until quite recently, most error augmentation
el of impairment may benefit more from control studies had been done with healthy subjects or in a
algorithms that provide a greater challenge, forcing single session, limiting their acceptance in therapy.
patients to adapt to more and more complex situa- Now that multisession studies with patients have
tions. This is similar to training in sports: while confirmed the usefulness of error augmentation
beginners need to be shown how to perform basic (Abdollahi et al. 2014), it is likely to become a stan-

WINTER 2015 29
Articles

dard feature of the next generation of commercial The use of such measurements in rehabilitation
rehabilitation robots. robotics is known as biocooperative control.
Physiological responses can be directly fed into a
controller that adapts exercise parameters to, for
On a Higher Level: example, keep heart rate close to a desired value
Exercise Adaptation (Koenig, Omlin, et al. 2011). Alternatively, machine-
learning algorithms, such as neural networks or dis-
The controllers in the Cooperative Assistive Control
criminant analysis, can be used to infer workload
and the Challenge-Based Control sections work on
from multiple physiological responses and adapt the
the level of each individual motion, changing their
exercise difficulty accordingly. Such analysis of phys-
assistance or challenges during the motion based on
iological responses has been shown to effectively
how well the patient is performing at that moment.
complement performance measurements in both
But we can go a step farther and adapt the difficulty
upper (Guerrero et al. 2013, Novak et al. 2011) and
of the overall exercise depending on how well the
lower (Koenig, Novak, et al. 2011) extremity rehabil-
patient has been doing over the course of the exer-
itation robots, though it is still unclear whether phys-
cise.
iological measurements provide enough additional
Performance-Based Adaptation information to offset the sensor costs and additional
preparation time.
The principle of performance-based exercise adapta-
tion is simple: if the patient is performing well with
respect to a certain performance criterion, make the Socially Assistive
exercise harder to perform by increasing the motion
complexity, required range of motion, required
Rehabilitation Robots
velocity, and so on. Conversely, if the patient is large- Finally, an entirely different type of rehabilitation
ly unsuccessful at achieving motions, make the exer- robot should be mentioned: those that do not phys-
cise easier by decreasing the above parameters. This ically support the patient, but instead demonstrate
should ideally ensure an appropriate moderate chal- motions to be performed and provide simultaneous
lenge for patients and increase their motivation to verbal guidance, acting as an exercise coach. These
exercise. An early implementation was demonstrated are relatively rare in motor rehabilitation and have
by Colombo et al. (2007), and increasingly more mainly been studied by the group of Maja Mataric for
complex strategies were proposed later (Cameirão et healthy older adults (Fasola and Mataric 2012). They
al. 2010; Chemuturi, Amirabdollahian, and Dauten- utilize artificial intelligence techniques that allow
hahn 2013). However, most of these strategies have them to deliver appropriate verbal instructions at the
only been tested over a single session; for the great- right time as well as, for example, switch between dif-
est benefit, they should be able to guide the patient ferent personalities — one patient may prefer caring
over multiple sessions, gradually shaping therapy. gentle guidance while another may prefer a military
A new, interesting challenge for performance- drill instructor robot that brooks no argument
based exercise adaptation has recently arisen in the (Tapus, Tapus, and Mataric 2008).
form of multipatient exercise games where patients While socially assistive robots are unlikely to see
compete or cooperate with each other in order to broad use with populations such as stroke or spinal
achieve a goal (Novak et al. 2014). The exercise cord injury, where physical support is crucial, they
parameters must then be adapted to suit each exer- have provided several lessons that also apply to
cising patient even though different patients may motor rehabilitation robots, such as patients’ subjec-
have different needs. Several adaptation algorithms tive preferences for different types of guidance and
have been suggested for such games (Caurin et al. encouragement. Researchers have now begun adapt-
2011, Maier et al. 2014), but are so far relatively basic ing elements of socially assistive robots for classic
and have seen very little testing with patients. rehabilitation robots, creating robots that give verbal
instructions as they provide impedance-based assis-
Blood, Sweat, and Tears: tance (Mihelj et al. 2012).
Biocooperative Control
Just because a patient is successfully performing the
exercise does not mean that exercise difficulty should
Discussion
be increased; he/she may be already overloaded and Large multicenter clinical studies of rehabilitation
struggling just to keep up. Alternatively, a patient robots have shown that robots can deliver effective
who is failing at the exercise may be enjoying the rehabilitation with several advantages over manually
challenge and could get annoyed if difficulty were assisted therapy (Klamroth-Marganska et al. 2014, Lo
decreased. To obtain additional subject-specific infor- et al. 2010). However, the biggest studies focused
mation, we can measure the patient’s physiological only on impedance-based assistance; other control
responses such as heart rate and skin conductance, strategies such as error augmentation have gained
which reflect both physical and mental workload. limited clinical acceptance and are rarely seen in

30 AI MAGAZINE
 Articles

commercial robots. One major challenge will be to home or group settings. In the future, this combina-
determine how different control strategies can be tion of new artificial intelligence methods and better
combined over the course of rehabilitation. For knowledge of existing control strategies will increase
example, it may be beneficial for patients to progress the therapeutic advantage of rehabilitation robots
from assistive control to resistive control and error and lead to their widespread adoption in many dif-
augmentation as their motor skills improve. As ferent settings.
another example, some patients might benefit from
EEG- or EMG-based assistance while others should
avoid it entirely. However, these two examples Acknowledgments
remain educated guesses at the moment. A major This work was supported by the Swiss National Sci-
step forward would be to create a set of guidelines ence Foundation through the National Centre of
(based on clinical evidence) for how different control Competence in Research (NCCR) Robotics and by
strategies should be combined in order to achieve the Clinical Research Priority Program “NeuroRe-
optimal rehabilitation outcome. hab,” University of Zurich.
These guidelines may not be only used by thera-
pists; they could be built into the rehabilitation robot References
itself to give it a greater level of autonomy. At the Abdollahi, F.; Case Lazarro, E. D.; Listenberger, M.; Kenyon,
moment, it is always the therapist’s task to adapt the R. V.; Kovic, M.; Bogey, R. A. Hedeker, D.; Jovanovic, B. D.,
exercise and switch between different control meth- Patton J. L. 2014. Error Augmentation Enhancing Arm
Recovery in Individuals with Chronic Stroke. Neurorehabil-
ods. In the future, the rehabilitation robot could per-
itation and Neural Repair 28(2): 120–128. dx.doi.org/
form such adaptation autonomously based on the 10.1177/1545968313498649
patient’s diagnosis, impairment level, and perform-
Cai, L. L.; Fong, A. J.; Otoshi, C. K.; Liang, Y.; Burdick, J. W.;
ance. The robot’s autonomy could be further and Roy, R. R. 2006. Implications of Assist-As-Needed
enhanced by teaching it to detect compensatory Robotic Step Training After a Complete Spinal Cord Injury
motions: motions where the patient, for example, on Intrinsic Strategies of Motor Learning. Journal of Neuro-
compensates for the inability to lift the arm by lifting science 26(41): 10564–10568. dx.doi.org/10.1523/JNEU-
the shoulder instead (Cirstea and Levin 2000). Such ROSCI.2266-06.2006
motions must be corrected by therapists because they Cameirão, M. S.; Badia, S. B. I.; Oller, E. D.; and Verschure,
impede the recovery process, create additional health P. F. M. J. 2010. Neurorehabilitation Using the Virtual Real-
problems, and lead to permanent adoption of patho- ity Based Rehabilitation Gaming System: Methodology,
logical movements. However, the artificial intelli- Design, Psychometrics, Usability, and Validation. Journal of
gence methods needed to detect and correct them are NeuroeEngineering and Rehabilitation 7(22 Sept 2010): 48.
beyond the current generation of rehabilitation dx.doi.org/10.1186/1743-0003-7-48
robots, and may also require additional sensors such Casadio, M., and Sanguineti, V. 2012. Learning, Retention,
as cameras. and Slacking: A Model of the Dynamics of Recovery in
Robot Therapy. IEEE Transactions on Neural Systems and
A combination of autonomous exercise adaptation
Rehabilitation Engineering 20(3): 286–296. dx.doi.org/10.
and detection of compensatory motions would 1109/TNSRE.2012.2190827
enable rehabilitation robots to be more efficiently
Caurin, G. A. P.; Siqueira, A. A. G.; Andrade, K. O.; Joaquim,
used in settings such as home exercise, where no R. C.; and Krebs, H. I. 2011. Adaptive Strategy for Multi-
therapist is present, or group exercise, where a single User Robotic Rehabilitation Games. In Proceedings of the
therapist supervises multiple patients and cannot ful- 2011 Annual International Conference of the IEEE Engineering
ly focus his/her attention on one patient. While the in Medicine and Biology Society, 1395–1398. Piscataway, NJ:
robot will always lack a certain human aspect, ele- Institute of Electrical and Electronics Engineers. dx.doi.org/
ments of socially assistive robotics could partially 10.1109/IEMBS.2011.6090328
compensate for this weakness, creating an intelligent Cesqui, B.; Tropea, P.; Micera, S.; and Krebs, H. I. 2013.
and affable robotic therapist. EMG-Based Pattern Recognition Approach in Post Stroke
Robot-Aided Rehabilitation: A Feasibility Study. Journal of
NeuroeEngineering and Rehabilitation 10: 75.
Conclusions dx.doi.org/10. 1186/1743-0003-10-75
Chemuturi, R.; Amirabdollahian, F.; and Dautenhahn, K.
Several control strategies have been developed for
2013. Adaptive Training Algorithm for Robot-Assisted
rehabilitation robots, from very simple assistive con- Upper-Arm Rehabilitation, Applicable to Individualised
trol to complex error augmentation methods and and Therapeutic Human-Robot Interaction. Journal of Neu-
task difficulty adaptation based on physiological roengineering and Rehabilitation 10(28 Sept. 2013): 102.
responses. However, these strategies need to be fur- dx.doi.org/10.1186/1743-0003-10-102
ther evaluated to determine how they can be most Cirstea, M. C., and Levin, M. F. 2000. Compensatory Strate-
effectively used and combined in clinical practice. gies for Reaching in Stroke. Brain 123(5): 940–953.
Furthermore, rehabilitation robots themselves need dx.doi.org/10.1093/brain/123.5.940
to be augmented with higher-level decision making Colombo, R.; Pisano, F.; Mazzone, A.; Delconte, C.; Micera,
so that they can operate more autonomously in S.; Carrozza, M. C.; Dario, P.; and Minuco, G. 2007. Design

WINTER 2015 31
Articles

Strategies to Improve Patient Motivation During Robot-Aid- Lam, T.; Wirz, M.; Lünenburger, L.; and Dietz, V. 2008.
ed Rehabilitation. Journal of Neuroengineering and Rehabilita- Swing Phase Resistance Enhances Flexor Muscle Activity
tion 4:(19 Feb 2007): 3. dx.doi.org/10.1186/1743-0003-4-3 During Treadmill Locomotion in Incomplete Spinal Cord
Ekkelenkamp, R.; Veltink, P.; Stramigioli, S.; and van der Injury. Neurorehabilitation and Neural Repair 22(5): 438–446.
Kooij, H. 2007. Evaluation of a Virtual Model Control for dx.doi.org/10.1177/1545968308315595
the Selective Support of Gait Functions Using an Exoskele- Lambercy, O.; Dovat, L.; Gassert, R.; Burdet, E.; Teo, C. L.;
ton. In Proceedings of the 10th IEEE International Conference on and Milner, T. 2007. A Haptic Knob for Rehabilitation of
Rehabilitation Robotics, 693–699. Piscataway, NJ: Institute of Hand Function. IEEE Transactions on Neural Systems and
Electrical and Electronics Engineers. dx.doi.org/10.1109/ Rehabilitation Engineering 15(3): 356–366. dx.doi.org/10.
icorr.2007.4428501 1109/TNSRE.2007.903913
Fasola, J., and Mataric, M. J. 2012. Using Socially Assistive Lo, A. C.; Guarino, P. D.; Richards, L. G.; Haselkorn, J. K.;
Human-Robot Interaction to Motivate Physical Exercise for Wittenberg, G. F.; Federman, D. G.; Ringer, R. J., Wagner, T.
Older Adults. Proceedings of the IEEE 100(8): 2512–2526. H.; Krebs, H. I.; Volpe, B. T.; Bever Jr., C. T.; Bravata, D. M.;
dx.doi.org/10.1109/JPROC.2012.2200539 Duncan P. W.; Corn, B. H.; Maffucci, A. D.; Nadeau, S. E.;
Frisoli, A.; Loconsole, C.; Leonardis, D.; Banno, F.; Barsotti, Conroy, S. S.; Powell, J. M.; Huang, G. D.; and Peduzzi, P.
M.; Chisari, C.; and Bergamasco, M. 2012. A New Gaze-BCI- 2010. Robot-Assisted Therapy for Long-Term Upper-Limb
Driven Control of an Upper Limb Exoskeleton for Rehabil- Impairment After Stroke. New England Journal of Medicine
itation in Real-World Tasks. IEEE Transactions on Systems, 362(19): 1772–83. dx.doi.org/10.1056/NEJMoa0911341
Man, and Cybernetics — Part C: Applications and Reviews Lotze, M.; Braun, C.; Birbaumer, N.; Anders, S.; and Cohen,
42(6): 1169–1179. dx.doi.org/10.1109/TSMCC.2012. L. G. 2003. Motor Learning Elicited by Voluntary Drive.
2226444 Brain 126(4): 866–872. dx.doi.org/10.1093/brain/awg079
Guerrero, C. R.; Fraile Marinero, J. C.; Turiel, J. P.; and Lum, P. S.; Burgar, C. G.; Van der Loos, M.; Shor, P. C.; Maj-
Muñoz, V. 2013. Using “Human State Aware“ Robots to mundar, M.; and Yap, R. 2006. Mime Robotic Device for
Enhance Physical Human-Robot Interaction in a Coopera- Upper-Limb Neurorehabilitation in Subacute Stroke Sub-
tive Scenario. Computer Methods and Programs in Bomedicine jects: A Follow-Up Study. Journal of Rehabilitation Research
112(2): 250–259. dx.doi.org/10.1016/j.cmpb.2013.02.003 and Development 43(5): 631–642. dx.doi.org/10.1682/JRRD.
Hassan, M.; Kadone, H.; Suzuki, K.; and Sankai, Y. 2014. 2005.02.0044
Wearable Gait Measurement System with an Instrumented Maier, M.; Ballester, B. R.; Duarte, E.; Duff, A.; and Ver-
Cane for Exoskeleton Control. Sensors 14(1): 1705–1722. schure, P. F. M. J. 2014. Social Integration of Stroke Patients
dx.doi.org/10.3390/s140101705 Through the Multiplayer Rehabilitation Gaming System. In
Israel, J. F.; Campbell, D. D.; Kahn, J. H.; and Hornby, T. G. Games for Training, Education, Health, and Sports, Lecture
2006. Metabolic Costs and Muscle Activity Patterns During Notes in Computer Science Volume 8395, 100–114. Berlin:
Robotic- and Therapist-Assisted Treadmill Walking in Indi- Springer. dx.doi.org/10.1007/978-3-319-05972-3_12
viduals with Incomplete Spinal Cord Injury. Physical Thera- Marchal-Crespo, L., and Reinkensmeyer, D. J. 2008. Haptic
py 86(11): 1466–1478. dx.doi.org/10.2522/ptj.20050266 Guidance Can Enhance Motor Learning of a Steering Task.
Jezernik, S.; Colombo, G.; and Morari, M. 2004. Automatic Journal of Motor Behavior 40(6): 545–556.
Gait-Pattern Adaptation Algorithms for Rehabilitation with Marchal-Crespo, L., and Reinkensmeyer, D. J. 2009. Review
a 4-Dof Robotic Orthosis. IEEE Transactions on Robotics and of Control Strategies for Robotic Movement Training After
Automation 20(3): 574–582. dx.doi.org/10.1109/TRA.2004. Neurologic Injury. Journal of Neuroengineering and Rehabilita-
825515 tion 6(16 June 2009): 20. dx.doi.org/10.1186/1743-0003-6-
Klamroth-Marganska, V.; Blanco, J.; Campen, K.; Curt, A.; 20. dx.doi.org/10.3200/JMBR.40.6.545-557
Dietz, V.; Ettlin, T.; Felder, M.; Fellinghauer, B.; Guidali, M.; Mihelj, M.; Novak, D.; Milavec, M.; Ziherl, J.; Olensek, A.;
Killmar, A.; Luft, A.; Nef, T.; Schuster-Amft, C.; Stahel, W.; and Munih, M. 2012. Virtual Rehabilitation Environment
and Reiner, R.. 2014. Three-Dimensional, Task-Specific Using Principles of Intrinsic Motivation and Game Design.
Robot Therapy of the Arm after Stroke: A Multicentre, Par- Presence: Teleoperators and Virtual Environments 21(1): 1–15.
allel-Group Randomised Trial. Lancet Neurology 13(2): 159– dx.doi.org/10.1162/PRES_a_00078
166. dx.doi.org/10.1016/S1474-4422(13)70305-3 Novak, D.; Mihelj, M.; Ziherl, J.; Olensek, A.; and Munih, M.
Koenig, A.; Novak, D.; Omlin, X.; Pulfer, M.; Perreault, E.; 2011. Psychophysiological Measurements in a Biocoopera-
Zimmerli, L; Mihelj, M.; and Riener, R. 2011. Real-Time tive Feedback Loop for Upper Extremity Rehabilitation. IEEE
Closed-Loop Control of Cognitive Load in Neurological Transactions on Neural Systems and Rehabilitation Engineering
Patients During Robot-Assisted Gait Training. IEEE Transac- 19(4): 400–410. dx.doi.org/10.1186/1743-0003-11-64
tions on Neural Systems and Rehabilitation Engineering 19(4): Novak, D.; Nagle, A.; Keller, U.; and Riener, R. 2014. Increas-
453–64. dx.doi.org/10.1109/TNSRE.2011.2160460 ing Motivation in Robot-Aided Arm Rehabilitation with
Koenig, A.; Omlin, X.; Bergmann, J.; Zimmerli, L.; Bolliger, Competitive and Cooperative Gameplay. Journal of Neuro-
M.; Müller, F.; and Riener, R. 2011. Controlling Patient Par- engineering and Rehabilitation 11(16 April 2014): 64.
ticipation During Robot-Assisted Gait Training. Journal of Patton, J. L.; Stoykov, M. E.; Kovic, M.; and Mussa-Ivaldi, F.
Neuroengineering and Rehabilitation 8(23 March 2011): 14. A. 2006. Evaluation of Robotic Training Forces That Either
dx.doi.org/10.1186/1743-0003-8-14 Enhance or Reduce Error in Chronic Hemiparetic Stroke Sur-
Krebs, H. I.; Palazzolo, J. J.; Dipietro, L.; Ferraro, M.; Krol, J.; vivors. Experimental Brain Research 168(3): 368–383.
Rannekleiv, K.; Volpe, B. T.; and Hogan, N. 2003. Rehabili- dx.doi.org/10.1007/ s00221-005-0097-8
tation Robotics: Performance-Based Progressive Robot- Patton, J. L.; Wei, Y. J.; Bajaj, P.; and Scheidt, R. A. 2013.
Assisted Therapy. Autonomous Robots 15(1): 7–20. Visuomotor Learning Enhanced by Augmenting Instanta-
dx.doi.org/10.1023/A:1024494031121 neous Trajectory Error Feedback During Reaching. PLOS One

32 AI MAGAZINE
 Articles

8(1): e46466. dx.doi.org/10.1371/ journal.pone.0046466

Ramos-Murguialday, A.; Broetz, D.; Rea, M.; Läer, L.; Yilmaz,
Ö.; Brasil, F. L.; Liberati, G.; Curado, M. R.; Carcia-Cossio, E.;
Vyziotis, A.; Cho, W.; Agostini, M.; Soares, E.; Soekadar, S.;
Caria, A.; Cohen, L. G.; Birbaumer, N. 2013. Brain-Machine
AAAI-16
Interface in Chronic Stroke Rehabilitation: A Controlled
Study. Annals of Neurology 74(1): 100–108. dx.doi.org/10.
1002/ana.23879
Registration Is
Reinkensmeyer, D. J.; Wolbrecht, E. T.; Chan, V.; Chou, C.;
Cramer, S. C.; and Bobrow, J. E. 2012. Comparison of Three-
Dimensional, Assist-As-Needed Robotic Arm/Hand Move-
Now Open!
ment Training Provided with Pneu-Wrex to Conventional
Tabletop Therapy after Chronic Stroke. American Journal of
Physical Medicine and Rehabilitation 91(11): 232–241.
dx.doi.org/10.1097/ PHM.0b013e31826bce79 Students—Be sure to arrive early
Riener, R.; Lünenburger, L.; Jezernik, S.; Anderschitz, M.; for AAAI-16!
Colombo, G.; and Dietz, V. 2005. Patient-Cooperative
Strategies for Robot-Aided Treadmill Training: First Experi-
mental Results. IEEE Transactions on Neural Systems and In order to participate fully in all student activities at
Rehabilitation Engineering 13(3): 380–94. dx.doi.org/10.1109/ AAAI-16, be sure to arrive by Friday, February 12,
TNSRE.2005. 848628 2016 in Phoenix.
Song, R.; Tong, K.; Hu, X.; and Li, L. 2008. Assistive Control The first day will include the AAAI Workshop Pro-
System Using Continuous Myoelectric Signal in Robot-Aid- gram, Tutorial Forum, and AAAI/SIGAI Doctoral Con-
ed Arm Training for Patients After Stroke. IEEE Transactions sortium (attendance open to all students). A wel-
on Neural Systems and Rehabilitation Engineering 16(4): 371– come student reception will be held Friday evening,
379. dx.doi. org/10.1109/TNSRE.2008.926707 February 12, to help new students get acquainted
Song, R.; Tong, K.; Hu, X.; and Zhou, W. 2013. Myoelectri- with others, as well as all the upcoming programs.
cally Controlled Wrist Robot for Stroke Rehabilitation. Jour- Included below are the tutorial and workshop sched-
nal of Neuroengineering and Rehabilitation 10(10 June 2013): ules.
52. dx.doi.org/10.1186/1743-0003-10-52 For complete registration and hotel information,
Stauffer, Y.; Allemand, Y.; Bouri, M.; Fournier, J.; Clavel, R.; please see www.aaai.org/aaai16.
Metrailler, P.; Brodard, R.; and Reynard, F. 2009. The Walk-
trainer — A New Generation of Walking Reeducation Device
Online registration is available at
Combining Orthoses and Muscle Stimulation. IEEE Trans-
regonline.com/aaai16.
actions on Neural Systems and Rehabilitation Engineering 17(1):
38–45. dx.doi.org/10.1109/ TNSRE.2008.2008288
Stein, J.; Narendran, K.; McBean, J.; Krebs, K.; and Hughes,
R. 2007. Electromyography-Controlled Exoskeletal Upper-
Limb-Powered Orthosis for Exercise Training After Stroke.
American Journal of Physical Medicine and Rehabilitation 86(4):
255–261. dx.doi. org/10.1097/PHM.0b013e3180383cc5 M. 2010. Evaluation of Upper Extremity Robot-Assistances
Tapus, A.; Tapus, C.; and Mataric, M. 2008. User — Robot in Subacute and Chronic Stroke Subjects. Journal of Neuro-
Personality Matching and Assistive Robot Behavior Adapta- engineering and Rehabilitation 7(18 Oct. 2010): 52. dx.doi.
tion for Post-Stroke Rehabilitation Therapy. Intelligent Serv- org/10.1186/1743-0003-7-52
ice Robotics 1(2): 169–183. dx.doi.org/10.1007/s11370-008-
0017-4
Trlep, M.; Mihelj, M.; and Munih, M. 2012. Skill Transfer Domen Novak is an assistant professor at the Department
from Symmetric and Asymmetric Bimanual Training Using of Electrical and Computer Engineering at the University of
a Robotic System to Single Limb Performance. Journal of Wyoming. He previously obtained his diploma and Ph.D.
Neuroengineering and Rehabilitation 9(17 July 2012): 43. from the University of Ljubljana, and was a postdoctoral
dx.doi.org/10.1186/1743-0003-9-43
fellow at the Sensory-Motor Systems Lab of ETH Zurich,
Vallery, H.; van Asseldonk, E. H. F.; Buss, M.; and van der Switzerland. His current research interests include motor
Kooij, H. 2009. Reference Trajectory Generation for Reha- and cognitive rehabilitation, human-robot interaction, psy-
bilitation Robots: Complementary Limb Motion Estima- chophysiology, and virtual reality.
tion. IEEE Transactions on Neural Systems and Rehabilitation
Engineering 17(1): 23–30. dx.doi.org/10.1109/TNSRE.2008. Robert Riener is a full professor and head of the Sensory-
2008278 Motor Systems Lab at ETH Zurich, Switzerland. He holds a
Wolbrecht, E. T.; Chan, V.; Reinkensmeyer, D. J.; and double professorship at the Medical Faculty of the Univer-
Bobrow, J. E. 2008. Optimizing Compliant, Model-Based sity of Zurich, and is thus active in the Spinal Cord Injury
Robotic Assistance To Promote Neurorehabilitation. IEEE Center of the Balgrist University Hospital. His research
Transactions on Neural Systems and Rehabilitation Engineering focuses on human-machine interaction, particularly
16(3): 286–97. dx.doi.org/10.1109/TNSRE.2008.918389 human sensory-motor control and the design of novel user-
Ziherl, J.; Novak, D.; Olensek, A.; Mihelj, M.; and Munih, cooperative robotic devices and virtual reality technologies.

WINTER 2015 33