Content uploaded by Daniel Ferris
Author content
All content in this area was uploaded by Daniel Ferris
Content may be subject to copyright.
RESEARCH ARTICLE
The effects of error augmentation on learning to walk
on a narrow balance beam
Antoinette Domingo •Daniel P. Ferris
Received: 5 March 2010 / Accepted: 31 August 2010
ÓSpringer-Verlag 2010
Abstract Error augmentation during training has been
proposed as a means to facilitate motor learning due to the
human nervous system’s reliance on performance errors to
shape motor commands. We studied the effects of error
augmentation on short-term learning of walking on a bal-
ance beam to determine whether it had beneficial effects on
motor performance. Four groups of able-bodied subjects
walked on a treadmill-mounted balance beam (2.5-cm
wide) before and after 30 min of training. During training,
two groups walked on the beam with a destabilization
device that augmented error (Medium and High Destabi-
lization groups). A third group walked on a narrower beam
(1.27-cm) to augment error (Narrow). The fourth group
practiced walking on the 2.5-cm balance beam (Wide).
Subjects in the Wide group had significantly greater
improvements after training than the error augmentation
groups. The High Destabilization group had significantly
less performance gains than the Narrow group in spite of
similar failures per minute during training. In a follow-up
experiment, a fifth group of subjects (Assisted) practiced
with a device that greatly reduced catastrophic errors (i.e.,
stepping off the beam) but maintained similar pelvic
movement variability. Performance gains were signifi-
cantly greater in the Wide group than the Assisted group,
indicating that catastrophic errors were important for short-
term learning. We conclude that increasing errors during
practice via destabilization and a narrower balance beam
did not improve short-term learning of beam walking. In
addition, the presence of qualitatively catastrophic errors
seems to improve short-term learning of walking balance.
Keywords Gait Task-specificity Rehabilitation
Movement variability
Introduction
Physical guidance is often given in rehabilitation settings
via the hands of a therapist. More recently, robotic devices
have been developed to provide physical guidance in
rehabilitation settings. The use of robotics has much
potential in rehabilitation because of their ease of use,
reliable measurement of performance and their capability
to deliver a high intensity and dosage of therapy
(Reinkensmeyer et al. 2004; Huang and Krakauer 2009).
However, to maximize rehabilitation outcomes, it is
important to first understand how best to use physical
guidance, robotic or otherwise (Marchal-Crespo and
Reinkensmeyer 2009; Reinkensmeyer and Patton 2009).
Several studies have shown that physical guidance
during practice hinders motor learning. For upper limb
movements, guidance given frequently during practice
improved performance, but once the guidance was
removed, the improvements were not present (Armstrong
A. Domingo D. P. Ferris
School of Kinesiology, University of Michigan,
3158 Observatory Lodge, 1402 Washtenaw Avenue, Ann Arbor,
MI 48109-2214, USA
D. P. Ferris
Department of Biomedical Engineering,
University of Michigan, Ann Arbor, MI, USA
D. P. Ferris
Department of Physical Medicine and Rehabilitation,
University of Michigan, Ann Arbor, MI, USA
A. Domingo (&)
818 West 10th Avenue, 3rd Floor, Vancouver,
BC V5Z 1M9, Canada
e-mail: antoinette.domingo@ubc.ca
123
Exp Brain Res
DOI 10.1007/s00221-010-2409-x
1970; Winstein et al. 1994). Similarly, we showed in a
recent study that error-reducing physical assistance given
during practice was detrimental to short-term learning
unassisted walking on a narrow balance beam (Domingo
and Ferris 2009). These findings are consistent with the
theory that error detection and correction are needed for
forming and updating internal models during motor learn-
ing (Wolpert and Ghahramani 2000). Internal models, or
neural representations of the body, task and environment,
are used to compare the expected movement to the actual
movement produced (Kawato 1999). When errors occur
(differences between the expected and actual movement),
the internal model is updated, and motor output is modified
to produce the correct movement. Over time, these errors
drive learning of a new internal model for new limb
dynamics or environment. Previous studies have shown
that motor learning is proportional to motor errors experi-
enced in upper limb tasks (Thoroughman and Shadmehr
2000; Scheidt et al. 2001). From this evidence, it could be
inferred that magnifying errors, rather than reducing errors,
that the subject experiences may increase the rate of motor
learning.
Previous motor adaptation studies have already shown
that amplifying errors improves motor learning of a new
task. Error augmentation of trajectory errors can enhance
learning of visuomotor rotations in an upper extremity
reaching task in healthy subjects (Wei et al. 2005). In
another study, robot-generated forces were applied to the
arm of individuals with stroke while the moving their arm
through a plane. After training, the individuals had
improved movement trajectories in directions where error
was amplified more than when error was reduced or was
zero (Patton et al. 2006). For motor learning in the lower
limb, Emken and Reinkensmeyer (2005) showed that error
amplification lead to faster formation of the internal model
in a novel walking task. No study as of yet has tested
whether error augmentation could be used to improve
motor learning of walking balance. This is an important
question because dynamic balance is a critical component
of gait control necessary for patients to safely practice
walking.
The purpose of this study was to determine whether
augmenting error during training affects short-term learn-
ing of walking balance. We studied able-bodied subjects
learning to walk on a treadmill-mounted balance beam
(beam-mill). Beam walking is similar to over ground
walking, but is more challenging to dynamic balance
because it exploits the lateral instability of walking
(Donelan et al. 2004; Schrager et al. 2008; Domingo and
Ferris 2009). We hypothesized that using error augmenta-
tion during training would improve motor short-term
learning of walking on the beam-mill more than practice
without error augmentation.
Previous error augmentation studies involved enhancing
motor adaptation to an altered environment by using
amplification of continuous trajectory errors. Unlike these
studies, we sought to enhance motor learning of an unal-
tered environment by augmenting the discrete error of
stepping off of the beam during practice. Two groups of
subjects practiced walking on the wide beam (2.5 cm) with
a destabilization device applied at the hips (Fig. 1a). The
destabilization device had the properties of a spring with
negative stiffness and augmented error by increasing the
number of times subjects stepped off of the beam. There
were two levels of spring stiffness used (Medium Desta-
bilization and High Destabilization groups). We tested a
third group that walked on the narrow beam as a form of
error augmentation. This group had also experienced error
augmentation during training (i.e., had stepped off the
beam more often) but the training was more similar to the
evaluation task than using the destabilization device. We
then compared these results to a group that practiced
walking on the wide beam without the destabilization
device (Wide group). All subjects from the four groups
were evaluated on the wide beam without the device pre-
and post-training. We hypothesized that subjects using
error augmentation during practice would have greater
performance gains than subjects that did not use error
augmentation. We based this hypothesis on the rationale
that error drives motor learning (Rumelhart et al. 1986;
Lisberger 1988; Dancause et al. 2002); therefore, aug-
menting error would lead to a faster rate of learning. We
also hypothesized that subjects walking on the narrow
beam during practice would have greater performance
gains than those that practiced with the destabilization
device on the wide beam. Walking on the narrow beam has
more similar task dynamics to the evaluation task of
walking on the wide beam because moments are still
generated at the foot to help maintain balance and no
additional external forces are introduced anywhere else in
the body.
We also wanted to investigate the relative roles of
making smaller control errors in movement (movement
variability of the pelvis) and larger catastrophic errors
(stepping off the beam) on learning to walk on the
beam-mill. In a second, follow-up experiment, we tested
another group of subjects that practiced walking on the
beam-mill while wearing a ‘‘kinematic channel’’ device
(Assisted group) that greatly decreased the number of
catastrophic errors experienced during practice but still
allowed a similar amount of movement variability as
walking on the wide beam without assistance (Wide
group). We hypothesized that the Assisted and Wide
groups would have similar performance gains, based on
results from the first experiment that showed perfor-
mance gains were better correlated to movement
Exp Brain Res
123
variability than to catastrophic errors experienced during
practice.
Methods
Subjects
We tested 50 able-bodied subjects (see Table 1for subject
characteristics). Subjects were medically stable and had no
history of major leg injuries. The University of Michigan
Institutional Review Board approved this study. All sub-
jects gave written informed consent in accordance with the
ethical standards laid down in the 1964 Declaration of
Helsinki prior to participating.
a
b
c
r
rp
s
F
p
Fs
r
rp
s
F
p
Fs
F
p, left
F
p, left
F
p, right
F
p, right
F
s, right
Position
F
F
s, left
Force
person
Centered
Pin joint
Fig. 1 Walking balance devices. aDestabilization device. A subject
walking on the beam-mill with the destabilization device used to
apply forces on the subject with springs that appeared to have
negative stiffness. This was accomplished by varying the moment
arms via an over-center linkage placed between the springs and the
subject. When the subject’s pelvis moved away from the center of the
beam, the device applied a proportional force onto the subject in the
same direction that the subject was moving. The inset graph shows a
simplified representation of the properties of the device. The thin gray
lines represent the forces due to the device at each side of the person,
where the heavy black line shows the net force due to both sides of the
device as a function of the subject’s pelvis position. The shaded area
represents the operating range of the device. Physical blocks were set
so that the device was would stop applying additional force soon after
the subject stepped off of the beam. bDetail of over-center linkage
used in destabilization device. The linkage provided a ‘‘moment
balance’’ between the moment produced by the spring force (F
s
) and
the moment produced from the cable tension that applied force to the
person (F
p
). When the linkage rotated as the subject moved from
being ‘‘centered’’ over the beam to ‘‘off-centered’’, the moment arms
changed greatly, resulting in an increase in F
p
.cA subject walking on
the treadmill-mounted balance beam with the kinematic channel
device. The assist device had straps that were set so that each subject
would have maximal space to move in the frontal plane but not so
much space that the subjects would be unable to right themselves as
they were beam walking
b
Table 1 Subject demographics
Group Gender Body mass (kg) Leg length (m)
MF
Narrow 2 8 60.3 ±10.5 0.88 ±0.032
Medium Destabilization 3 7 59.1 ±8.3 0.88 ±0.054
High Destabilization 2 8 60.7 ±8.9 0.87 ±0.048
Wide 4 6 64.6 ±14.7 0.89 ±0.063
Assisted 4 6 65.5 ±7.2 0.91 ±0.019
Exp Brain Res
123
Procedures
Five groups of 10 subjects walked on the beam-mill for a 3-
min pre-training evaluation, a 30-min training period (with
rest breaks every 10 min) and a 3-min post-training evalu-
ation. During the pre- and post-evaluation periods, all sub-
jects walked on the wide beam (2.5 cm wide) to test for
performance gains and were made aware of this at the
beginning of the experiment. The first two groups walked
with the destabilization device with medium spring stiffness
or high spring stiffness during the training period (Medium
Destabilization or High Destabilization groups, respec-
tively). A third group walked on the narrow beam (1.27 cm
wide) without a device during the training period (Narrow
group). A fourth group walked on the wide beam without a
device during the training period (Wide group). The last
group walked on the wide beam while wearing the kinematic
channel device (Assisted group). Data presented in this paper
from the Wide group were collected and published in a
previous study (Domingo and Ferris 2009) but are used here
to compare to the data from the other three groups.
Treadmill speed was set at 0.22 m/s. This speed was
determined during pilot testing. Subjects were instructed to
walk on the beam for as long as possible without stepping off.
Instructions were given to all subjects by the same experi-
menter. They had to walk heel-to-toe with arms crossed over
their torso. They were also instructed not to lean forward,
twist their trunk, angle their feet away from the longitudinal
direction of the beam, or look down at their feet. View of the
walking surface was obscured by wearing dribble goggles.
Subjects were allowed to move their pelvis and trunk in the
frontal plane to help maintain balance. All subjects wore
standardized orthopedic shoes. Subjects had to wait 5 s after
stepping off the beam before attempting to walk on it again.
Equipment
The equipment for this experiment consisted of a treadmill-
mounted balance beam (beam-mill), a destabilization
device, a kinematic channel device, force plates and a
motion capture system. The beam-mill was made of
interchangeable small wooden blocks attached to the
treadmill belt that lined up to make a continuous balance
beam. One beam was 2.5 cm wide by 2.5 cm tall (Wide)
and the other was 1.27 cm wide by 2.5 cm tall (Narrow).
Smaller wooden blocks were added to either side of the
bases of both beams to make them more stable in the
frontal plane.
Destabilization device
The destabilization device was made up of latex tubing
springs, an over-center linkage and cables that attached to
the subject via a padded hip belt (Fig. 1a). This device
applied forces onto the subject with springs with an
effective negative stiffness. The negative spring stiffness
was achieved by placing an over-center linkage between
the subject and the spring. As the person’s pelvis moved
away from the center of the beam, a proportional force was
applied to the subject in the same direction, but if the
person was centered over the beam, the device applied
approximately zero net force onto the subject.
The linkage on each side of the destabilization device
provided a ‘‘moment balance’’ between the moment pro-
duced by the spring (F
s
) and the moment produced from
cable tension that applied a force to the person (F
p
) (Fig. 1b).
Therefore, the force applied to the person by the left or right
side of the device is described by the following equation:
Fp¼Fs
rs
rp
where F=force, r=moment arm, p=person, and
s=spring. The springs were pre-tensioned so that the
spring force on each side was non-zero at each side when
the person was centered over the beam, but the net force
(F
p, right
?F
p, left
) summed to approximately zero and the
person felt no pull to either side. When the subject moved
their pelvis toward the left, the linkage on the left side of
the device rotated away from the person. The spring on the
same side then shortened and the spring force, F
s, left
,
decreased. However, the moment arms also changed,
where the spring moment arm (r
s, left
) length increased and
the moment arm length of the person (r
p, left
) decreased.
The relatively larger changes in moment arm, and rela-
tively smaller decrease in F
s, left,
resulted in F
p, left
increasing overall, and pulled the person with greater force
in same direction as they originally had moved. On the
right half of the device, movement of the subject’s pelvis to
the left caused the linkage on the right side to rotate toward
the person. On the right side, F
p, right
then decreased
(F
s, right
was greater, but r
s, right
was shorter and r
p, right
was
longer). The net force (F
p, right
?F
p, left
) on the subject
resulted in the device pulling the subject to the left when
the subject’s pelvis moved to the left.
The device made it difficult to stay on the beam if the
hips moved away from the center of the beam. The device
also gave subjects feedback about their position relative to
the beam. The subjects were made aware of the function of
the device and were encouraged not to translate anteriorly
or posteriorly on the treadmill. When the subject’s pelvis
was centered over the beam, there was approximately zero
net force applied to the subject. We had 8 pairs of springs
of different stiffnesses. For each subject, we chose the
spring that would provide the stiffness closest to the non-
dimensionalized spring stiffness of 0.2978 for the Medium
Destabilization group and 0.4404 for the High
Exp Brain Res
123
Destabilization group. To determine the desired spring
stiffness, we used the following equation:
k¼
kl
mg
where k=dimensionalized stiffness,
k¼non-dimension-
alized stiffness, l=leg length and mg =bodyweight. The
non-dimensionalized spring stiffnesses of 0.2978 and
0.4404 were based on springs used during pilot testing. The
average total stiffness of the device was 192.5 N/m for the
Medium Destabilization group and 298.4 N/m for the High
Destabilization group.
Kinematic channel device
For the second part of the experiment, we tested a group of
subjects that walked on a beam using a device that allowed
normal frontal plane movement variability but helped to
minimize stepping off of the beam. This training device
was made up of lightweight cables and adjustable straps
that attached to the subject via a padded hip belt (Fig. 1c).
The straps were set so that each subject would have max-
imal space to move their pelvis in the frontal plane but not
so much space that the subjects would be unable to right
themselves as they were beam walking. We placed single-
axis tension/compression load cells (1,200 Hz; Omega
Engineering, Stamford, CT, USA) in series with the cables
on both sides of the subject to measure the tension in the
cables produced by the subjects during walking. Subjects
were instructed not to depend on the device because they
would not be able to use it during the post-training period.
Recording procedures
The treadmill was placed above two force plates (sampling
rate 1,200 Hz; Advanced Mechanical Technology Inc.,
Watertown, MA, USA) so that we could calculate center of
pressure from the forces and moments produced by the
subject while walking. The center of pressure helped us
determine when the subject was on or off the beam.
We used an 8-camera video system (frame rate 120 Hz;
Motion Analysis Corporation, Santa Rosa, CA, USA) to
record the positions of 4 reflective markers placed on the
subject’s pelvis, neck and shoulders during walking. We
calculated the standard deviation of the medio-lateral
movement of the marker placed at the sacrum and neck to
determine movement variability.
Performance measures
We recorded the number of times the subject stepped off the
beam per minute. We then divided this quantity by the
fraction of time the subject was on the beam (not touching the
treadmill surface with either foot). This quotient, failures per
minute, was our primary performance metric because it took
into account the number of errors with respect to the amount
of time the subject successfully walked on the beam. We also
calculated the standard deviation (SD) of the medio-lateral
movement of markers placed at the sacrum and the neck
(Motion Analysis Corporation, Santa Rosa, CA; 120 Hz).
We calculated percent change of the performance variables
by subtracting the pre-training value from the post-training
value and dividing by the pre-training value for each subject
to normalize to pre-training performance.
For the pre- and post-training periods, we recorded data
for the duration of the 3-min trial. For the 30-min training
period, we collected only 20 s of data per each minute of
training. We used a 4th order low-pass zero-lag Butter-
worth filter with a cutoff frequency of 6 Hz to smooth raw
marker data. Values for SD of markers were calculated
only using the data from when subjects were on the beam.
We used a 4th order low-pass zero-lag Butterworth filter
with a cutoff frequency of 25 Hz to smooth raw force data,
then a 4th order low-pass zero-lag Butterworth filter with a
cutoff frequency of 6 Hz to smooth center of pressure data.
Data were processed using custom software written in
MATLAB (The MathWorks, Inc., Natick, MA).
Statistical analysis
We first performed an analysis of variance (ANOVA) to
determine whether groups evaluated on the same beams
had similar failures per minute during pre-training (JMP IN
software, SAS Institute, Inc., Cary, NC).
We then performed an ANOVA to test for differences
between the groups for each of the following dependent
variables: percent change for failures per minute, failures
per minute during training and sacral marker SD. For post
hoc analysis, we performed Tukey’s honestly significant
difference (THSD) test to compare results between groups
as needed to delineate the differences between groups.
For the error augmentation groups and the Wide group,
we calculated the correlation coefficient between sacral
marker standard deviation and percent change in failures
per minute. This would help to determine whether there
was a relationship between movement variability while
walking on the beam and the performance gains.
To analyze the load cell data from the assist device, we
calculated net force and then normalized the data to each
subject’s body weight. We calculated the root mean square
(RMS) of the data only from when the subject was on the
beam. The force RMS data for the training period was then
averaged into six 5-min blocks. We then performed a
repeated measures ANOVA as an omnibus test to find
differences in force RMS between the 5-min blocks. We
performed a paired ttest to find statistical difference
between the first and last 5-min blocks of force RMS data.
Exp Brain Res
123
We also compared the sacral marker SD data during
training between groups to further examine whether
movement variability was similar between groups
throughout the training period. We performed ttests to
compare data between groups during the first 5-min of
training and the last 5-min of training.
Results
All groups had the similar pre-training values for failures
per minute (P=0.087) and sacral marker standard devi-
ation (P=0.364).
Error augmentation groups versus Wide group
The groups that practiced with error augmentation experi-
enced more failures per minute (Medium Destabilization:
27.3 ±2.0, High Destabilization: 29.6 ±1.4, Narrow:
26.5 ±2.8, mean ±SEM) during the training period than
the Wide group (12.6 ±1.3) (Fig. 2a) (ANOVA,
P\0.0001, power =0.99, THSD, P\0.05). All three
error augmentation groups had similar amounts of failures
per minute during training (THSD, P[0.05).
Although more error was experienced during practice in
the error augmentation groups than in the Wide group, the
Wide group had significantly greater performance gains
than all other groups (-61.2 ±6.0%) (ANOVA,
P\0.0001, power =0.99, THSD, P\0.05) (Fig. 2b).
The High Destabilization group had a smaller percent
change in failures per minute (-8.1 ±5.3%) than the
Medium Destabilization group (-23.6 ±6.2%), but the
difference was not significant (THSD, P[0.05). The
performance gains were significantly higher in the Narrow
group than in the High Destabilization group (THSD,
P\0.05) (Fig. 2b). The Narrow group had a
-34.6 ±7.9% change in failures per minute.
Sacral marker movement variability
versus performance gains
The relative trend in performance gains (the additive
inverse of percent change in failures per minute) for all
groups was similar to that in the movement variability of
the sacral marker (Fig. 3a, b). The correlation coefficient,
q, between these variables was 0.4281 (P=0.0059) and
R
2
=0.1833. The relative trend in failures per minute was
opposite that of the performance gains (Fig. 3a, c).
Assisted group versus Wide group
Subjects in the Assisted group decreased use of the device
as the training period progressed. Root-mean-square
(RMS) of the net force (normalized to bodyweight) per
minute was calculated for each subject. Force data were
only included in calculations from when the subject was on
the beam. Figure 4a shows the averaged force RMS for
each minute of data across subjects that used assistance
during the training period. We averaged the RMS across 5-
min intervals and then performed a repeated measures
ANOVA to find if there were differences in force RMS
across the 30-min training period. The analysis showed that
there was a statistically significant difference between the
different 5-min blocks (ANOVA, P\0.0001). Post hoc
analysis showed that there the force RMS for the first
5 min (1.2 ±0.24% bodyweight) was significantly greater
than for the last 5 min of the training period
(0.72 ±0.15% bodyweight) (paired ttest, P=0.0265).
Using the assist device greatly reduced the number of
failures during training, but during post-training, the
number of errors returned to pre-training values. Figure 4b
shows the averaged failures per minute for both groups
during pre- and post-training and during each minute of
training. Figure 4c shows the averaged sacral marker SD
0
10
20
30
Failures per
minute
during training
NarrowMD HD
Wide
***
-60
-40
-20
0
Failures per
minute
(% change)
** **
a
b
Fig. 2 Averaged failures per minute and percent change. a. Averaged
failures per minute during training across subjects for each group.
Error bars are ±1 Standard Error of the Mean (SEM). The Medium
Destabilization (MD), High Destabilization (HD) and Narrow groups
had significantly greater failures per minute during training than the
Wide group (ANOVA, P\0.0001, THSD*, P\005). There was no
statistical difference between the error augmentation groups (MD,
HD, Narrow) (THSD*, P[0.05). bAveraged percent change ((post-
training-pre-training)/pre-training values) for failures per minute
across subjects for each group. Error bars are ±1 SEM. The Wide
group had greater performance gains after training than both
Destabilization groups and Narrow group (ANOVA, P\0.0001,
THSD*, P\0.05). The Narrow group had significantly greater
performance gains than the High Destabilization group (THSD*,
P[0.05)
Exp Brain Res
123
for both groups during pre-and post-training and during
each minute of training.
We wanted to verify that both groups had similar
amounts of movement variability over the whole training
period. Sacral marker movement variability was slightly
greater in the Wide group (39.0 ±2.7 mm, mean ±SEM)
than the Assisted group (32.7 ±4.7 mm), but the differ-
ence was not significant (ANOVA, P=0.2626) (Fig. 5a).
When comparing 5-min blocks of data during the training
period, we found that there were no differences in sacral
marker SD (ttest, P=0.8760) between groups during the
first 5 min of training. During the last 5 min of training,
movement variability was greater in the Wide group (ttest,
P=0.0143) by 30%.
We also wanted to ensure that the kinematic channel
device was effective at preventing subjects from stepping
off the beam. We compared the number of failures per
minute during training for both groups and found that they
were significantly different (ANOVA, P\0.0001)
(Fig. 5b). The Assisted group had an average of 1.7 ±0.44
failures per minute during the training period, while the
Wide group had an average of 12.6 ±1.3 failures per
minute during training.
Practicing with the assist device clearly hindered
learning (Fig. 5c). The Assisted group had -1.7 ±11.7%
change in failures per minute, while the Wide group had
-61.2 ±6.0% change in failures per minute. There were
much greater performance gains in the Wide group
(ANOVA, P=0.0003, power =0.99).
0
20
40
60
Performance
gains
0
10
20
30
40
Sacral marker
SD during
training (mm)
Narrow
MD
HD
Wide
0
10
20
30
Failures per
minute
during training
a
b
c
Fig. 3 Performance gains versus sacral marker movement variability
and failures per minute during training. Error bars are ±1 SEM.
aPerformance gains are the additive inverse of the percent change in
failures per minute for each group. The relative performance gains
between groups were similar to the brelative sacral marker
movement variability during training between groups and had an
inverse relationship with cfailures per minute during training
0 10 20 30
0
0.5
1
1.5
Minute of training
Force RMS
(% bodyweight)
0 10 20 30
0
5
10
15
20
25
Failures
per
minute
Minute of training
Pre-
training
Post-
training
100 20 30
0
20
40
60
Minute of training
Sacral
marker
SD (mm)
Wide
Assisted
Pre-
trainin
g
Post-
trainin
g
a
b
c
Fig. 4 Averaged time series data from the training period. aAveraged
root-mean-square (RMS) of net force from the assist device as a
percent of bodyweight. Data are taken only from when subjects were
walking on the beam. bAveraged number of failures per minute for
each minute across subjects for each group. The Assisted group had
very few failures per minute after 10 min of the training period.
cAveraged standard deviations (SD) for the sacral marker in the
frontal plane as a measure of movement variability across subjects for
each group. Data included in the calculation was only from when
subjects were on the beam. Averaged data from the first 5 min of
training showed that there were no differences in movement
variability (SD) between groups. Averaged data from the last 5 min
of training showed that movement variability was higher in the Wide
group (ANOVA, P=0.0143)
Exp Brain Res
123
Discussion
Error augmentation
The main result of this study showed that augmented error
training with either the destabilizing device (Medium
Destabilization and High Destabilization) or with a nar-
rower balance beam was actually worse for short-term
learning of walking balance than unaltered practice. This
was contrary to our hypothesis that subjects using error
augmentation during practice would have greater perfor-
mance gains than subjects that did not use error augmen-
tation. We also found that when the error augmentation has
more similar task dynamics to the desired task (narrow
beam training), it led to greater performance gains com-
pared to error augmentation with less similar task dynamics
compared to the desired task (destabilization device
training).
One explanation for why practicing with the destabili-
zation device led to poorer performance gains compared to
unaltered practice is the role of internal models in motor
learning. Considerable research supports the theory that the
nervous system forms internal models of movement
dynamics during motor learning (Kawato 1999; Wolpert
et al. 2001). Recent studies have provided specific evidence
that humans use internal models during walking (Emken
and Reinkensmeyer 2005; Lam et al. 2006) and stationary
balance (Ahmed and Ashton-Miller 2007). When using the
destabilization device, the dynamics of the task were
changed. As a result, the learner may have formed an
internal model for walking on the beam that included the
device dynamics. Once the device was removed, the sub-
jects had not developed the internal model for beam
walking without the device and exhibited minimal learning
during the post-training period. Detecting and correcting
errors are important for motor learning, but the errors must
be specific to the dynamics of the desired task.
The importance of task dynamics on internal models
could also explain why subjects in the Narrow group had
greater performance gains than the High Destabilization
group (Fig. 2b). Walking on the narrow beam during
practice likely has more similar task dynamics than walk-
ing with the destabilization device because using the
destabilization device applies additional external forces to
the pelvis and walking on the narrow beam does not. As a
result, the internal model formed during narrow beam
walking was more transferable to wide beam walking than
the internal model formed during walking with the desta-
bilizing device.
Another possible reason why the Wide beam group may
have had the greatest performance gains is that practicing on
the wide beam unassisted may have provided optimal level
of error experience (i.e., stepping off the beam) during
practice [i.e., at the ‘‘optimal challenge point’’ (Guadagnoli
and Lee 2004)]. Too many errors experienced during prac-
tice may not allow for an appropriate example of the task
(Sanger 2004) and may lead to decreased motivation because
of frustration. In contrast, too few errors experienced during
practice may not provide enough feedback to refine the
internal model of task dynamics (Scheidt et al. 2000; Patton
et al. 2006). The error augmentation groups in our study had
experienced more errors during practice than the Wide
group. The increased task difficulty may have been too high
to stimulate motor learning.
Our findings were different than previous studies that
found error augmentation to be beneficial for motor
learning. There are several reasons why this may be the
Wide Assisted
0
10
20
30
40
50
Sacral marker
SD during
training (mm)
Wide Assisted
0
2
4
6
8
10
12
14
Failures per
minute during
training
*
Wide Assisted
-80
-60
-40
-20
0
20
Failures per
minute
(% change)
*
a
b
c
Fig. 5 Averaged sacral marker SD during training, failures per
minute during training, and percent change in failures per minute
across subjects for each group. aSacral marker SD calculated from
marker position in the medio-lateral direction when subjects were on
the beam. Data is averaged over the entire training period across all
subjects. The difference in movement variability between groups was
not significant (ANOVA, P=0.2626). bAveraged failures per
minute over the entire training period across all subjects. The Assisted
group had significantly less failures per minute during training
(ANOVA, P\0.0001). cAveraged percent change in failures per
minute from pre- to post-training. Using the assist device during
practice clearly hindered learning, as there were significantly greater
performance gains in the Unassisted group than the Assisted group
(ANOVA, P=0.0003)
Exp Brain Res
123
case. The present study is different than previous studies
involving error augmentation because of (1) the type of
error that was amplified and (2) the type of task goal.
Previous studies involved augmenting continuous trajec-
tory errors to adapt to a novel task [e.g., with a force field
(Patton et al. 2006; Emken and Reinkensmeyer 2005)or
visual feedback (Wei et al. 2005)]. In our study, we aug-
mented the discrete error of stepping off of the beam with
the destabilization device or with the narrower beam in
order to enhance learning of the unaltered task.
Stepping off the beam is a discrete and qualitatively
catastrophic error (i.e., losing balance so that beam walking
is no longer possible). We chose to augment this type of
error because this error was directly related to the subjects’
task goal: walking on the balance beam for as long as
possible without stepping off. Other parameters of walking
dynamics, such as center of mass amplitude or movement
variability, were not a primary measurement of error in this
study because these parameters were not part of the stated
task goal. Subjects could have learned better control of
their center of mass, allowing for greater center of mass
movement without stepping off of the beam with extended
practice and therefore this parameter would not be an
accurate measure of error.
The second reason our results may have been different
than those of other studies is because previous studies
utilized error augmentation to enhance motor adaptation to
a novel environment. That is, subjects learned a new sen-
sorimotor calibration in order to adapt to a different
dynamic or visual environment, resulting in the desired
trajectory. The motor learning paradigm in the present
study is different than that of motor adaptation. In our
study, subjects practiced with a variation of the goal task
(one that was similar, but with increased catastrophic
errors) to test if this variation would enhance motor
learning of the unaltered task goal. Although there are
similarities between motor adaptation and motor learning,
these are distinct processes that likely have different neural
and behavioral mechanisms (Huang and Krakauer 2009).
Our findings may also have differed from previous
studies because we specifically tested learning of walking
balance, while others examined learning of discrete arm
movements in a plane (Patton and Mussa-Ivaldi 2004;
Patton et al. 2006) or learning to step through a viscous
force field (Emken and Reinkensmeyer 2005). These types
of movements may be less complex than the task of
maintaining walking balance, which involves multiple
sensory inputs (visual, vestibular, and proprioceptive) and a
high degree of coordination among multiple body segments
in the upper and lower body. Perhaps, the complexity and
higher degree of difficulty of our task would not be aided
by error augmentation, especially in the earlier stages of
learning for our naı
¨ve subjects.
There may be some instances when error augmenta-
tion for learning walking balance may be useful. A
common issue in rehabilitation is preparing patients for
the ‘‘real world.’’ Walking does not always occur in a
straight line and over smooth surfaces. Practicing with
error augmentation may help patients respond to pertur-
bations or changes in the environment. If the unaltered
task can be performed proficiently, augmenting error
with different task dynamics may be beneficial. By
having diverse practice conditions, individuals can gen-
eralize learning so that learning of a new task happens at
a faster rate (Seidler 2004). Similarly, a recent study
showed that humans use an internal model for the
dynamics of an environment and task to create appro-
priate feedback responses for unanticipated errors, even
if they had never previously experienced the error
(Wagner and Smith 2008).
Sacral marker movement variability in the frontal plane
correlated well with performance gains (the additive
inverse of the percent change in failures per minute)
(Fig. 3a, b). The destabilization device in this study
increased catastrophic error (i.e., stepping off the beam)
based on the subject’s movements (Fig. 3c), but it also
limited the amount of movement variability that the subject
was able to experience while walking on the beam
(Fig. 3b). Movement variability at the pelvis may reflect
the number of smaller errors in control that are made,
allowing for updates to the internal model. This may be an
alternative indicator of learning compared to catastrophic
errors experienced during practice. The destabilization
device may have increased catastrophic errors, but it also
decreased the smaller errors experienced while walking on
the beam that are evidenced by movement variability.
There was a significant correlation between movement
variability and performance gains (q=0.4281, P=
0.0059), but only 18% of the variance was explained by
this relationship due to high inter-subject variability.
Greater movement variability may be indicative of
greater learning for this walking balance task. This is
supported by the observation that humans seem to detect a
loss of balance with a ‘‘control error signal anomaly’’
(CEA) during standing balance (Ahmed and Ashton-Miller
2004,2007). To determine motor output for a desired
movement, the central nervous system creates an internal
model of limb dynamics based on previous sensorimotor
experiences. The expected sensory feedback is then com-
pared to the actual sensory feedback. If a sufficiently large
difference between the two is detected, or CEA, then a
compensatory response will occur. Subjects that success-
fully learned to walk on the balance beam had experienced
greater movement variability, better explored the move-
ment space, and used the movement errors to update the
internal model.
Exp Brain Res
123
Movement variability versus stepping off the beam
Based on these results, we wanted to more specifically
examine the role of movement variability and smaller
control errors to delineate its effects on learning relative to
larger error experience. For example, learning to ride a bike
with training wheels that do not touch the ground while the
bicycle is vertical should provide a means for riders to
explore the task space of balancing without falling over.
We built a similar type of stabilization device for walking
on the beam-mill. It provided a channel of very low forces
on the torso during task space exploration while providing
high forces if the torso moves too far to one side or the
other to prevent failure. This could be seen as similar to the
type of kinematic channel in hindlimb movement used
during robotic locomotor training in spinalized mice (Cai
et al. 2006).
The main result of our follow-up experiment showed
that experiencing catastrophic errors (stepping off the
beam) during practice may be important for learning this
beam walking task, contrary to our hypothesis. Subjects
that used the kinematic channel device experienced
movement variability similar to unassisted subjects
(Fig. 5a) and had a reduced number of failures during
practice (Fig. 5b), but had very small performance gains
(Fig. 5c). This suggests that giving assistance that reduced
qualitatively catastrophic errors hindered motor learning of
a walking balance task.
It is likely that catastrophic balance errors involve
higher level cortical processes. This has been demonstrated
in a number of EEG studies on humans performing balance
tasks (reviewed in Maki and McIlroy 2007). Results from
follow-up studies in our laboratory have found that the
anterior cingulate displays an error-related negativity
(ERN) event related potential when subjects step off of the
balance beam. Our working hypothesis is that this ERN
detected in the anterior cingulate is important for motor
learning during walking balance. Future studies will con-
tinue to explore this relationship.
There are several reasons why ‘‘kinematic channel’’
assistance may have hindered learning of narrow beam
walking. A learner’s ability to recognize and correct their
errors increases as movement skill improves (Liu and
Wrisberg 1997). Although using the assist device allowed
for a similar amount of movement variability as the Wide
group, it also greatly reduced opportunities for detection
and correction of the larger errors (stepping off the beam).
After about 10 min of training, the Assisted group rarely
stepped off the beam, while the Wide group continued to
step off the beam throughout the training period. It is
possible that stepping off the beam actually provides higher
level information about performance compared to the
smaller control errors, contributing to the learning process
(Wei and Kording 2009). Learning occurs in response to
both small and large errors and with distinct neural path-
ways (Criscimagna-Hemminger et al. 2010).
The assist device also changed task dynamics by
applying forces to stop lateral translation of the pelvis once
the subject reached a predetermined distance away from
center. The presence of these forces could affect how
subjects learn to maintain balance on the beam-mill.
Strategies formed to balance while using the assist device
are likely very different than those used without the device.
Subjects in the Assisted group had about the same
amount of movement variability at the pelvis as the Wide
group throughout most of the training period (Fig. 5a). This
shows that the assist device did allow enough space for
normal movement variability. However, the subjects in the
Assisted group were less variable with their movements by
the end of training. These subjects were told at the
beginning of the experiment not to become dependent on
the device because they would be evaluated on Wide beam
walking. They may have concentrated too much on
avoiding using the device and as a result, ended up with
reduced movement variability. Alternatively, subjects may
have been able to use very low forces from the device
toward the end of training for feedback to limit their
movement variability.
Although the Assisted group used the device minimally
during training, especially toward the end of training
(Fig. 4a), even very small forces may have helped to
maintain balance. Several studies have shown that light
touch (less than 1 Newton of force) at the fingertip can
greatly reduce postural sway during standing with eyes
closed due to the augmented sensory feedback rather than
physical stabilization (Holden et al. 1994; Jeka and Lack-
ner 1994; Kouzaki and Masani 2008). In our study, most
subjects in the Assisted group reported that they felt they
had greatly decreased the use of the device at the end of the
training period. However, it is possible that subjects
unknowingly became dependent on the very low forces
during practice. These forces may have been able to give
some cues to their position in space. Perhaps, these low
forces from the device within the kinematic channel could
be eliminated by placing physical blocks a small distance
away from each side of the pelvis. Even so, the restriction
in movement provided by these blocks would likely change
task dynamics enough to hinder learning.
There is another possible reason why the Assisted group
had lower performance gains. There are two separate parts
of this task that require different dynamics: getting on the
beam initially and then taking steps to stay on the beam.
Because this group spent most of their time walking on the
beam, they had fewer opportunities to learn the act of
successfully getting back on the beam after a failure.
Without this skill, subjects were more likely to step off the
Exp Brain Res
123
beam soon after stepping on, greatly increasing the number
of failures per minute. Using the kinematic channel device
made practice less similar to the walking on the beam
without any assistance.
A recent study comparing the effectiveness of locomotor
training using the Lokomat (a robotic exoskeleton used for
automated treadmill stepping) versus conventional gait
training in patients with subacute stroke (Hidler et al. 2009)
supported the results of our study. They found that subjects
that received conventional gait training had greater
improvements in gait speed and walking distance than
those that trained in the Lokomat. They attributed these
results in part to how the Lokomat provides guidance of the
lower extremities and greatly restricts motion at the trunk
and pelvis. If motion is limited at the trunk and pelvis, the
patients are unable to sense and correct for movement
errors during walking and would greatly limit learning of
balance.
Conclusions
This study showed that (1) error augmentation achieved
with destabilization or a narrower balance beam is not
better than practicing the unaltered beam-walking task, (2)
task-specific dynamics are important considerations for
practice, (3) movement variability of the pelvis correlates
well with performance gains for beam walking, and (4)
making qualitatively catastrophic errors may be important
for short-term learning of walking on a narrow beam. This
suggests that rehabilitation strategies should be devised so
that assistance allows patients to make catastrophic errors
(so that the goal movement is no longer possible) during
practice but still maintain safety and prevent falls. Future
studies should be conducted to further examine the rela-
tionship between catastrophic errors and motor learning.
Acknowledgments The authors would like to thank Daniela Weiss,
Sarah Weiss, Evelyn Anaka and other members of the Human
Neuromechanics Lab for help with data collection and processing. We
would also like to thank Shawn O’Connor and Peter Adamczyk for
help with data analysis and Steve Collins for help with the negative-
stiffness spring design. This work was supported by the Rackham
Graduate Student Research Grant, the Foundation for Physical
Therapy PODS II Scholarship, and National Institutes of Health F31
HD056588-01.
Conflict of interest The authors declare that they have no conflict
of interest.
References
Ahmed AA, Ashton-Miller JA (2004) Is a ‘‘loss of balance’’ a control
error signal anomaly? Evidence for three-sigma failure detection
in young adults. Gait Posture 19:252–262
Ahmed AA, Ashton-Miller JA (2007) On use of a nominal internal
model to detect a loss of balance in a maximal forward reach.
J Neurophysiol 97:2439–2447
Armstrong TR (1970) Training for the production of memorized
movement patterns. Dissertation, University of Michigan
Cai LL, Fong AJ, Otoshi CK, Liang Y, Burdick JW, Roy RR,
Edgerton VR (2006) Implications of assist-as-needed robotic
step training after a complete spinal cord injury on intrinsic
strategies of motor learning. J Neurosci 26:10564–10568
Criscimagna-Hemminger SE, Bastian AJ, Shadmehr R (2010) Size of
error affects cerebellar contributions to motor learning. J Neuro-
physiol 103:2275–2284
Dancause N, Ptito A, Levin MF (2002) Error correction strategies for
motor behavior after unilateral brain damage: short-term motor
learning processes. Neuropsychologia 40:1313–1323
Domingo A, Ferris DP (2009) Effects of physical guidance on short-
term learning of walking on a narrow beam. Gait Posture
30:464–468
Donelan JM, Shipman DW, Kram R, Kuo AD (2004) Mechanical and
metabolic requirements for active lateral stabilization in human
walking. J Biomech 37:827–835
Emken JL, Reinkensmeyer DJ (2005) Robot-enhanced motor learn-
ing: accelerating internal model formation during locomotion by
transient dynamic amplification. IEEE Trans Neural Syst
Rehabil Eng 13:33–39
Guadagnoli MA, Lee TD (2004) Challenge point: a framework for
conceptualizing the effects of various practice conditions in
motor learning. J Mot Behav 36:212–224
Hidler J, Nichols D, Pelliccio M, Brady K, Campbell DD, Kahn JH,
Hornby TG (2009) Multicenter randomized clinical trial eval-
uating the effectiveness of the Lokomat in subacute stroke.
Neurorehabil Neural Repair 23:5–13
Holden M, Ventura J, Lackner JR (1994) Stabilization of posture
by precision contact of the index finger. J Vestib Res 4:
285–301
Huang VS, Krakauer JW (2009) Robotic neurorehabilitation: a compu-
tational motor learning perspective. J Neuroeng Rehabil 6:5
Jeka JJ, Lackner JR (1994) Fingertip contact influences human
postural control. Exp Brain Res 100:495–502
Kawato M (1999) Internal models for motor control and trajectory
planning. Curr Opin Neurobiol 9:718–727
Kouzaki M, Masani K (2008) Reduced postural sway during quiet
standing by light touch is due to finger tactile feedback but not
mechanical support. Exp Brain Res 188:153–158
Lam T, Anderschitz M, Dietz V (2006) Contribution of feedback and
feedforward strategies to locomotor adaptations. J Neurophysiol
95:766–773
Lisberger SG (1988) The neural basis for learning of simple motor
skills. Science 242:728–735
Liu J, Wrisberg CA (1997) The effect of knowledge of results delay
and the subjective estimation of movement form on the
acquisition and retention of a motor skill. Res Q Exerc Sport
68:145–151
Maki BE, McIlroy WE (2007) Cognitive demands and cortical control
of human balance-recovery reactions. J Neural Transm
114:1279–1296
Marchal-Crespo L, Reinkensmeyer DJ (2009) Review of control
strategies for robotic movement training after neurologic injury.
J Neuroeng Rehabil 6:20
Patton JL, Mussa-Ivaldi FA (2004) Robot-assisted adaptive training:
custom force fields for teaching movement patterns. IEEE Trans
Biomed Eng 51:636–646
Patton JL, Stoykov ME, Kovic M, Mussa-Ivaldi FA (2006) Evalu-
ation of robotic training forces that either enhance or reduce
error in chronic hemiparetic stroke survivors. Exp Brain Res
168:368–383
Exp Brain Res
123
Reinkensmeyer DJ, Patton JL (2009) Can robots help the learning of
skilled actions? Exerc Sport Sci Rev 37:43–51
Reinkensmeyer DJ, Emken JL, Cramer SC (2004) Robotics, motor
learning, and neurologic recovery. Annu Rev Biomed Eng
6:497–525
Rumelhart DE, Hinton GE, Williams RJ (1986) Learning represen-
tations by back-propagating errors. Nature 323:533–536
Sanger TD (2004) Failure of motor learning for large initial errors.
Neural Comput 16:1873–1886
Scheidt RA, Reinkensmeyer DJ, Conditt MA, Rymer WZ, Mussa-Ivaldi
FA (2000) Persistence of motor adaptation during constrained,
multi-joint, arm movements. J Neurophysiol 84:853–862
Scheidt RA, Dingwell JB, Mussa-Ivaldi FA (2001) Learning to move
amid uncertainty. J Neurophysiol 86:971–985
Schrager MA, Kelly VE, Price R, Ferrucci L, Shumway-Cook A
(2008) The effects of age on medio-lateral stability during
normal and narrow base walking. Gait Posture 28:466–471
Seidler RD (2004) Multiple motor learning experiences enhance
motor adaptability. J Cogn Neurosci 16:65–73
Thoroughman KA, Shadmehr R (2000) Learning of action through
adaptive combination of motor primitives. Nature 407:742–747
Wagner MJ, Smith MA (2008) Shared internal models for feedfor-
ward and feedback control. J Neurosci 28:10663–10673
Wei K, Kording K (2009) Relevance of error: what drives motor
adaptation? J Neurophysiol 101:655–664
Wei Y, Bajaj P, Scheidt R, Patton JL (2005) Visual error augmen-
tation for enhancing motor learning and rehabilitative relearning.
In: International conference on rehabilitation robotics. IEEE,
Chicago, IL, pp 505–510
Winstein CJ, Pohl PS, Lewthwaite R (1994) Effects of physical
guidance and knowledge of results on motor learning: support
for the guidance hypothesis. Res Q Exerc Sport 65:316–323
Wolpert DM, Ghahramani Z (2000) Computational principles of
movement neuroscience. Nat Neurosci 3:1212–1217
Wolpert DM, Ghahramani Z, Flanagan JR (2001) Perspectives and
problems in motor learning. Trends Cogn Sci 5:487–494
Exp Brain Res
123