Cub Cadet Time Saver i1042 User manual

Enabling Always-Available Input with
Muscle-Computer Interfaces
T. Scott Saponas1, Desney S. Tan2, Dan Morris2, Ravin Balakrishnan4,Jim Turner3, James A. Landay1
1Compuer Science and Engineering
DUB Group
University of Washington
{ssaponas, landay}@cs.washington.edu
2Microsoft Research
{desney, dan}@microsoft.com
3Microsoft Corporation
[email protected]
4Department of
Computer Science
University of Toronto
[email protected]
ABSTRACT
Previous work has demonstrated the viability of applying
offline analysis to interpret forearm electromyography
(EMG) and classify finger gestures on a physical surface.
We extend those results to bring us closer to using muscle-
computer interfaces for always-available input in real-world
applications. We leverage existing taxonomies of natural
human grips to develop a gesture set covering interaction in
free space even when hands are busy with other objects. We
present a system that classifies these gestures in real-time
and we introduce a bi-manual paradigm that enables use in
interactive systems. We report experimental results demon-
strating four-finger classification accuracies averaging 79%
for pinching, 85% while holding a travel mug, and 88%
when carrying a weighted bag. We further show generali-
zability across different arm postures and explore the tra-
deoffs of providing real-time visual feedback.
ACM Classification: H.1.2 [User/Machine Systems]; H.5.2
[User Interfaces]: Input devices and strategies; B.4.2 [In-
put/Output Devices]: Channels and controllers
General terms: Design, Human Factors
Keywords: Electromyography (EMG), Muscle-Computer
Interface, input, interaction.
INTRODUCTION
Our hands and our ability to control them have evolved
over thousands of years, yielding an amazing ability to pre-
cisely manipulate tools. As such, we have often crafted our
environments and technologies to take advantage of this
ability. For example, many current computer interfaces re-
quire manipulating physical devices such as keyboards,
mice, and joysticks. Even future looking research systems
often focus on physical input devices [5]. However, as
computing environments become more diverse, we often
find ourselves in scenarios where we either cannot, or pre-
fer not to, explicitly interact with a physical device in hand.
Previous work has explored hands-free and implement-free
input techniques based on a variety of sensing modalities.
For example, computer vision enables machines to recog-
nize faces, track movement and gestures, and reconstruct
3D scenes [24]. Similarly, speech recognition allows for
hands-free interaction, enabling a variety of speech-based
desktop and mobile applications [8, 11]. However, these
technologies have several inherent limitations. First, they
require observable interactions that can be inconvenient or
socially awkward. Second, they are relatively sensitive to
environmental factors such as light and noise. Third, in the
case of computer vision, sensors that visually sense the en-
vironment are often susceptible to occlusion.
We assert that computer input systems can leverage the full
bandwidth of finger and hand gestures without requiring the
user to manipulate a physical transducer. In this paper, we
show how forearm electromyography (EMG) can be used to
detect and decode human muscular movement in real time,
thus enabling interactive finger gesture interaction. We en-
vision that such sensing can eventually be achieved with an
unobtrusive wireless forearm EMG band (see Figure 1).
Previous work exploring muscle-sensing for input has pri-
marily focused either on using a single large muscle (rather
than the fingers) [2, 3, 4, 22, 25], which does not provide
the breadth of input signals required for computer input,
and/or on situations where the hand and arm are constrained
to a surface [3, 4, 15, 21, 23, 25], which is not a realistic
usage scenario for always-available input devices. Saponas
et al. [18] demonstrated the feasibility of using offline ma-
chine learning techniques to interpret forearm muscle-
sensing and classify finger gestures on a surface. We extend
their offline classification results to achieve online classifi-
Permission to make digital or hard copies of all or part of this work fo
r
p
ersonal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise,
or republish, to post on servers or to redistribute to lists, requires prio
r
specific permission and/or a fee.
UIST’09, October 4–7, 2009, Victoria, British Columbia, Canada.
Copyright 2009 ACM 978-1-60558-745-5/09/10...$10.00.
Figure 1. Artist rendering of a forearm muscle-sensing band.

cation that enables using muscle-sensing for always-
available input in real-world applications that are not con-
strained to a surface. Note that our contribution is not in the
realm of trying to better understand or measure the physiol-
ogy of human musculature, but rather in simply sensing
muscle activity to enable interaction. Specifically:
1. We leverage existing taxonomies of natural human
grips to develop a gesture set covering interaction in
free space, including when the hands are busy with ob-
jects, and even when hands and muscles are under load.
2. We develop a procedure for rapidly and robustly cali-
brating an activation signal, present a system that clas-
sifies our gestures in real-time, and introduce a bi-
manual “select and activate” paradigm that enables use
in interactive systems.
3. We demonstrate the feasibility of our approach through
a laboratory experiment. Results show average classifi-
cation accuracies of 79% for pinching, 85% while
holding a travel mug, and 88% when carrying a
weighted bag, all for four-finger gesture sets. Results
further suggest generalizability across different arm
postures. Furthermore, we show preliminary evidence
of use within a more ecologically valid example appli-
cation: controlling a simulated portable music player.
We conclude the paper with discussion of our results, the
limitations of our techniques, implications for design, and
proposals for future work.
BACKGROUND AND RELATED WORK
Sensing Muscles with EMG
Humans employ a complex set of skeletal muscles and ad-
joining tendons and bones to create body movement. The
brain initiates movement process by transmitting an elec-
trical signal through the nervous system. This signal stimu-
lates the fibers that make up our muscles, which contract in
response to create forces or body movement.
EMG senses this muscular activity by measuring the elec-
trical potential between pairs of electrodes. This can either
be done invasively (with needles in the muscle) or from the
surface of the skin. While invasive EMG can be very accu-
rate, our work focuses on surface EMG because it is more
practical for HCI applications. For more detailed informa-
tion on electromyography, see Merletti et al. [13].
For either modality (surface or invasive), the EMG signal is
an oscillating electrical wave. When a muscle is contracted,
the amplitude of this wave increases, with most of the pow-
er in the frequency range of 5 to 250 Hz [13].
Applications of EMG Sensing
EMG is frequently used in clinical settings for muscle func-
tion assessment during rehabilitation and for measuring
muscle activation to assess gait [9]. In clinical applications,
a typical statistic computed over the EMG signal is the root
mean squared (RMS) amplitude of the measured potential.
This provides a rough metric for how active a muscle is at a
given point in time. For a review of processing techniques
used in previous work, see [14].
EMG is also used in both research and clinical settings for
controlling prosthetics. This typically involves sensing the
activity in large individual muscles and using it as input to
control the movement of physical devices. For example, the
shoulder muscle might be used to control one of the degrees
of freedom in a lower-arm prosthetic. Other work has ex-
plored similar techniques for sensing activity in large mus-
cles such as the biceps or pectoralis for computer input by
healthy individuals (e.g. [2]). However, learning to perform
fine tasks with muscles that are not normally used for dex-
terous manipulation can be difficult.
Recent work has used surface EMG to sense and decipher
muscle activity that drives fine motor function in our fin-
gers, wrists, and hands. Wheeler et al. [23] explore EMG-
based input systems, but assume that the hands are in a con-
strained, static posture, and do not address calibration issues
associated with real-world use. Ju et al. [6] explored several
machine learning approaches to classifying a finger-pinch
gesture using electrodes placed near participants’ wrists,
and achieved classification accuracies as high as 78% when
differentiating among four gestures. Their work, however,
was focused on machine learning techniques, and does not
address the human-computer interaction issues that impact
the feasibility of real-world EMG applications. In particu-
lar, their work does not address posture-independence (e.g.,
arm rotation), hands-busy scenarios, scenarios in which
hands are not constrained to a surface, the “Midas Touch”
problem (differentiating intended gestures from rest), or
real-time classification. Our work builds on the work of Ju
et al. by addressing each of these issues.
Saponas et al. [18] used 10 EMG sensors worn in a narrow
band around the upper forearm to differentiate position,
pressure, tapping, and lifting gestures across five fingers
placed on a surface. They showed the effectiveness of using
not only RMS amplitude but also frequency energy and
phase coherence features in a linear classifier to attain com-
pelling proof-of-concept results. However, their work was
limited in that participants were constrained to fixed arm
postures while sitting in a chair and working on a physical
surface. Furthermore their data was processed using offline
analysis, which did not allow exploration of real-time inte-
ractions or the potential effects of feedback to the user.
We seek to extend previous muscle-sensing work to explore
real-time classification of finger-level movement for more
naturalistic settings including when people are holding ob-
jects. We also investigate practical concerns including arm
posture independence, “Midas touch,” and visual feedback.
Natural Human Grips
Most of the input devices we use for computing today take
advantage of our ability to precisely operate physical trans-
ducers like buttons, knobs, and sliders. While this is an ex-
cellent approach when a computing device is one’s primary
focus, as in desktop computing, physical devices can be

difficult or impossible to use when a user’s hands or body
are devoted to another activity. For example, a jogger may
strap a music player to her arm or waist. However, even
simple tasks such as changing songs, channels, or volume
can be a struggle, requiring a user to reach across her body,
possibly stop running, find the right button, and manipulate
it. In circumstances such as these, where a user prefers to
keep their hands free or is already holding something other
than an input device, we propose that muscle-sensing offers
an opportunity to take advantage of our manual dexterity
without requiring physical actuation of a device.
To guide the design of muscle-sensing-based interaction
techniques, it is important to consider the space of natural
human grips and hand postures that we might leverage for
gesture design. Over the last century, many grip posture
classifications have been developed for biomechanical
modeling, robotics, and therapy [12]. Schlesinger [20] put
forth most well-known of these taxonomies (see Figure 2),
characterizing six different manual grasps:
•Spherical: for holding spherical tools such as balls
•Cylindrical: for holding cylindrical tools such as cups
•Palmar: for grasping with palm facing the object
•Tip: for holding small tools like a pen
•Lateral: for holding thin, flat objects like paper
•Hook: for supporting a heavy load such as a bag
We explore techniques that will enable people to interact
with computers when their hands are already being used in
one of these grips, or when their hands are unencumbered
but a handheld device is impractical. We divide these grips
into three classes: small or no object in hand (tip and later-
al), tool in hand (cylindrical, spherical, and palmar), and
heavy load in hand (hook). Based on these three classes we
suggest finger gestures, detect and classify these gestures in
real-time using forearm muscle sensing, develop a two-
handed interaction technique that allows for these gestures
to control applications, and experimentally demonstrate the
efficacy of these gestures.
EXPERIMENT
We conducted a laboratory experiment to investigate using
forearm EMG to distinguish finger gestures within the three
classes of grips: (1) small or no object in hand, (2) tool in
hand, and (3) heavy load in hand.
Participants
Twelve individuals (5 female) volunteered to participate in
the experiment. Participants ranged from 18 to 55 years of
age with an average age of 36. All were daily computer
users, and came from a variety of occupations. None re-
ported existing muscular conditions or skin allergies, and
all were right-handed. None were colorblind and all had
20/20 or corrected-to-20/20 vision. The experiment took 1.5
hours and participants were given a small gratuity.
Equipment and Setup
We used a BioSemi Active Two system as our forearm EMG
sensing device (www.biosemi.com). This system samples
eight sensor channels at 2048 Hz. We first had participants
clean their forearms with a soft scrub solution while we
prepared the BioSemi sensors with conductive gel and ad-
hesive. The preparation, gel and adhesive are artifacts of
our EMG setup and could be eliminated if dry electrodes
such as the Dri-Stik (NeuroDyne Medical, Corp.) are used.
This would clearly be more appropriate for real-world use.
To get the best possible signal, EMG sensing is traditionally
conducted with two sensors spread an inch apart on a mus-
cle belly. However, Saponas et al. [18] showed that they
were able to obtain reasonable results even when not pre-
cisely placing sensors. As such, we chose to place six sen-
sors and two ground electrodes in a roughly uniform ring
around each participant’s upper right forearm for sensing
finger gestures. We also placed two sensors on the upper
left forearm for recognizing left-hand squeezes, or activa-
tion intent. This configuration mimics potential use with an
approximately-placed armband EMG device, as illustrated
in Figure 1. Setup took about 15 minutes.
Design and Procedure
We divided the experiment into three parts. Part A ex-
amined gestures when the participant’s hand was free of
objects and explored the sensitivity of our techniques to
arm posture. Part B examined gestures while the hands
were busy holding objects, a travel mug and a weighted bag
that created constant muscular load. In Part C, participants
used the muscle-computer interface (while holding an ob-
ject) to control a simulated portable music player.
Before beginning any of the tasks in each session, we per-
formed a short calibration step. Participants squeezed a ball
for four seconds and then relaxed for another four. This
calibration provided us with approximate maximum and
minimum values across each channel and feature, which we
used for normalizing the signal from each channel. Our
normalization process was to scale the signal from zero to
one based on the observed maximum and minimum value.
Parts A and B of the experiment each contained a training
phase, in which the system prompted the participant to per-
Figure 2. Schlesinger’s natural grip taxonomies [20] as de-
picted in MacKenzie and Iberall [12]. Groupings indicate the
three similarity classes that guide our gesture set.

form finger gestures while it collected training data. This
data was immediately used to train our gesture recognizer
and build a predictive model. The training phase was fol-
lowed by a testing phase in which the system attempted to
classify the participant’s gestures in real-time. Part C used
the training data collected in Part B for real-time control.
In a real-world interactive system, determining when a user
is performing a gesture and when he is not is crucial for
preventing spurious detection of gestures and precisely
labeling gesture onset or offset. This is particularly true if
there is a strong timing component to the application, such
as in games. Even in applications that do not have an intrin-
sic timing component, such as text entry, ambiguities in
timing can yield incorrect results. For example, when
switching from pinching with the index finger to the ring
finger, a user passes through intermediate states, which may
cause spurious or incorrect classifications of user intention.
Our approach to differentiating gesture from rest, and to
simultaneously increasing the precision of gesture timing, is
to introduce an explicit activation gesture. To do this, we
use a second muscle-interface source, making a fist and
squeezing the contra-lateral hand, in this case the non-
dominant hand. Squeezing is a large multi-muscle action
that can be robustly detected with consistent timing, but in
itself is not sufficiently complex for most applications. By
combining rich gestures performed with one hand and ro-
bust but simple gestures performed with the other hand, we
allow reliable and precise muscle-based interactions.
In addition to making the timing of input more predictable,
using the non-dominant hand for gesture activation also
allows the user to rapidly re-execute a single gesture many
times in a row. For example, when scrolling through a list,
the “down” gesture can be held for a second while the non-
dominant hand makes several quick squeezes. This bima-
nual “select and activate” paradigm is the one we used in
the testing phase of our experiment.
Part A: Hands-Free Finger Gestures
The first part of our experiment explored performing finger
gestures when the hands were not holding anything. Each
participant performed pinch gestures with the thumb and
one of the other fingers of their dominant hand. The gestur-
ing arm was held in a comfortable position with a bent el-
bow and the empty hand held at about shoulder height (see
Figure 1 and Figure 3a).
Without the constraint of a surface to rest on, people natu-
rally move and rotate their arms and wrists between ges-
tures. Doing so moves muscles under the skin and relative
to the attached sensors, creating changes to the observed
EMG signals and potentially impacting classification. Most
previous work has carefully constrained arm posture to
avoid this scenario (for example, securing people’s arm to a
surface). However, this is an unreasonable constraint if
muscle-computer interfaces are to be used for real-world
interaction. Hence, we set out to examine whether or not
our decoding techniques generalize to variable postures,
and more importantly, how we can improve our techniques
to better support posture variability.
We chose three different postures to explore: the two ex-
tremes of comfortable rotation of the forearm toward and
away from their shoulder (pronation and supination) as well
as a “natural” midpoint position (see Figure 3a).
Hands-Free Training Phase
Participants sat in a chair facing a desktop display. The sys-
tem prompted participants to pinch each of their fingers to
their thumb by highlighting the appropriate finger on an
outline of a hand (see Figure 4a). We asked participants to
press “a comfortable amount”. If they asked for clarifica-
tion, we told them to “press hard enough to dent a tomato,
but not hard enough to rupture the skin.” They were told to
relax their fingers when nothing was highlighted. Fingers
were highlighted for a second, with a break of three-
quarters of a second in between each stimulus.
We employed a block design, with each block comprising
one trial each of an index, middle, ring, and pinky finger
gesture, presented in random order. We gathered 25 blocks
of training data for each of the three arm postures, the order
of which was counterbalanced across participants.
Hands-Free Testing Phase
In the testing phase, participants performed 25 blocks of
gestures in each of the three arm postures. As in the training
Figure 4. (a) A red highlight indicates that a gesture should be
performed with the given finger; (b) a blue highlight indicates
the currently recognized gesture; (c) a purple highlight indi-
cates that the correct gesture is being performed.
Figure 3. Our finger gesture sets. a) pinch gestures performed
in three different arm postures b) fingers squeezing a travel
mug c) fingers pulling up against the handle of a carried bag

phase, participants received their cues via a highlighted
finger on the display. However, rather than timing their
responses to the timing of the stimuli, participants were
asked to perform the gesture with their right hand and “lock
it in” by clenching their left fist. To aid participants in this,
we provided a small ball that they could squeeze with their
left hand. The gesture could have just as easily been per-
formed without the prop, as we demonstrate in Part B of the
experiment. When the system recognized a squeezing
movement with the left hand, it classified the gesture being
performed with the right hand using the muscle-sensing
data immediately preceding the squeeze.
Locking in a gesture by squeezing made the finger hig-
hlighting disappear for half a second, after which the sys-
tem advanced to the next gesture. Since detecting the acti-
vation gesture is quicker and more robust than that of indi-
vidual finger gestures, the bimanual paradigm allows for
rapid selection of the same gesture multiple times in a row,
as well as a robust way to avoid false positives.
Part B: Hands-Busy Finger Gestures
The second part of our experiment explored performing
finger gestures when the hands are already busy holding an
object. We looked at two different classes of objects. First,
we used a travel mug to represent small tool-sized objects
held in the hand. For this task, participants sat in a chair and
held the mug in the air as one might naturally hold a beve-
rage (see Figure 3b). Second, we tested larger and heavier
objects being carried. Participants stood in front of the desk
and carried a laptop bag in each hand (see Figure 3c). Each
bag held a book weighing approximately one kilogram.
As in Part A, for both object types, we conducted a training
phase and a testing phase. These were done one object type
at a time and the order of the two object types was counter-
balanced across users.
Hands-Busy Training Phase
As before, participants performed 25 blocks of finger ges-
tures in response to stimuli. The same stimuli highlighting
fingers in the outline of a hand were used. Participants were
asked to exert a little more pressure with the highlighted
finger than with the other fingers. With the mug, this meant
pressing on it a little more firmly with the highlighted fin-
ger than with the other fingers. With the bag, this meant
pulling on the handle a little harder with the highlighted
finger than with the other fingers. At the conclusion of the
training phase for each object, the collected data was used
to train the gesture recognition system for use in the subse-
quent phases. Once training data is collected, training the
system requires only a few seconds of computation.
Hands-Busy Testing Phase
In the testing phase of this part of the experiment, partici-
pants used the two-handed technique to perform gestures as
they did in Part A. However, unlike in Part A, participants
completed the stimulus-response task twice: once with vis-
ual feedback about the real-time classification, and once
without visual feedback. The order was counterbalanced
across participants and objects to avoid an ordering effect.
The “no visual feedback” condition was in the same style as
Part A’s testing phase; a finger was highlighted and a par-
ticipant would perform that gesture then squeeze with their
left hand. When holding the travel mug, participants
squeezed an empty left hand with their fingers against the
lower pad of their thumb to “lock in” the current right-hand
gesture. When holding a bag in each hand, participants
squeezed the handle of the left-hand bag to “lock in” the
current right-hand gesture.
The “with visual feedback” condition added a second com-
ponent to the display of the hand. In addition to the red hig-
hlighting of the finger that should be used in the gesture, the
system also continuously highlighted its current gesture
recognition result in a semi-transparent blue (see Figure 4b-
c). We explained to participants that this was the system’s
best guess at their current gesture. Users were asked to per-
form the red gesture and activate their response only when
they were confident it was correctly detected. As a side
effect, visual feedback also allowed participants to under-
stand the system’s recognition behavior and to tailor their
gestures accordingly. The goal of this manipulation was to
explore the importance and tradeoffs of having visual feed-
back while using a muscle-computer interface.
Participants completed 25 blocks of gestures for each object
both with and without visual feedback. The order of the
feedback manipulation was balanced across the order of
participants and objects.
Part C: Controlling a Portable Music Player Application
In addition to testing the accuracy with which our system
was able classify gestures performed by participants, we
also applied these gestures to use in a more ecologically
valid application, a portable music player interface.
Our simulated portable music player (see Figure 5) is con-
trolled through a hierarchical menu interface similar to
those found in many mobile computing devices. Our player
contained eight songs and only the songs menu was popu-
lated. The menu system can be navigated using four direc-
tional arrows where the “up” and “down” arrows move a
selection cursor up and down in the current menu, while the
“left” and “right” arrows navigate backward or forward in
the menu structure. Forward navigation is also used to indi-
cate a final selection at the end of a series of navigations. In
music players, this corresponds to selecting a song.
We asked participants to control the portable music player
menu interface and complete a series of tasks using our
real-time muscle-computer interface. The training data from
Part B was used, since the hands were similarly loaded with
either the mug or the heavy bag. The user’s inputs were
mapped to the directional controller of the portable music
player by assigning the index finger of the right hand to left,
the pinky finger to right, the middle finger to up, and the
ring finger to down. As in the other experiments, the left-
hand grasping gesture was used to activate the gesture be-

ing performed by the right hand. The system continuously
highlighted in red the directional arrow corresponding to
the system’s current finger gesture recognition result. This
visual feedback told a participant what action the system
would take if he squeezed their left hand at that moment.
Participants completed three different tasks with the porta-
ble music player. They (a) navigated from the top of the
menu structure to the list of songs and selected a specified
song, (b) navigated from a random starting point in the
songs list to a particular song, and (c) advanced to the next
song, starting at a random song in the song list. Above the
music player participants were given task instructions such
as “Select Even Flow.” They would then do a series of di-
rection gestures to navigate the menu and select the song.
Participants completed five blocks of these three tasks for
each object (mug and heavy bag), for 30 tasks in total.
Data Processing Technique
To classify gestures from an EMG signal, we used a similar
approach to Saponas et al. [18], performing basic signal
processing, computing a set of features, using those features
to train a support vector machine (SVM) [1], and then using
that SVM to classify finger gestures. While Saponas, et al.
did not test this, we show here that this can be used in a
real-time system. We outline the procedure here, but more
details on the approach can be found in their paper [18].
Basic Signal Processing
Our first step is to convert the raw EMG data into a form
suitable for our machine learning algorithm. We divide the
signal into 32 segments per second (about 31ms per seg-
ment). By dividing the data into segments, we transform it
into a time independent dataset. We can then treat each of
these segments as a single sample of data.
Feature Generation
For each 31ms sample, we generated three classes of fea-
tures, which we use for training and testing the classifier.
The first set of features is the Root Mean Square (RMS)
amplitude in each channel, which correlates with magnitude
of muscle activity. From the six base RMS features gener-
ated by sensors on the right arm, we create another fifteen
features by taking the ratio of the base RMS values between
each pair of channels. These ratios make the feature space
more expressive by representing relationships between
channels, rather than treating each as being independent.
The second set of features is Frequency Energy, indicative
of the temporal patterns of muscle activity. To derive these
features, we compute the fast Fourier transform (FFT) for
each sample and square the FFT amplitude, which gives the
energy at each frequency. We create 13 bins over the 32 Hz
sampling range for each of the six channels on the right
arm. This yields 78 frequency energy features per sample.
The third set of features is Phase Coherence, which loosely
measures the relationships among EMG channels. We
create fifteen such features by taking the ratios of the aver-
age phase between all channel pairs on the right arm.
These calculations result in 114 features per sample for
right-hand gesture classification. The only feature we use
for left-hand “squeeze” recognition is a single RMS feature
computed over the subtractions of the two channels availa-
ble on the left hand.
Classification of Right-Hand Finger Gestures
Support vector machines (SVMs) are a set of supervised
machine learning methods that take a set of labeled training
data and create a function that can be used to predict the
labels of unlabeled data. For our experiment, we used the
Sequential Minimal Optimization version of SVMs [16].
In supervised machine learning, training data inherently
needs to be labeled with a ‘ground truth’. In our case, this is
the gesture being performed by a participant at a given time
when the muscle-sensing data segment was gathered. Be-
cause people respond to a stimulus with varying delay,
there is some amount of mislabeled information early with-
in each stimulus presentation. We combat this issue by dis-
carding all samples from the first half of presentation and
saving only the latter half as training data for our system.
While classification results were generated 32 times a
second, the system determined the currently recognized
gesture at any given time as the last gesture classified three
times in a row. For example, if the previous four samples
were classified as “index, index, index, middle”, the system
would use “index” as the currently recognized gesture. We
chose this approach to reduce sensitivity to momentary
fluctuations in classification. Throughout this paper, our
classifiers were trained and tested independently on data
from each participant during a single participant session.
Classification of Left-Hand Squeeze
Detecting the squeezing gesture performed by the left hand
is much simpler. We take the RMS features from the differ-
ence of the two channels on the left arm. This process re-
moves noise such as a person’s cardiac electrical activity,
giving a good estimate of the total muscle activity in the
upper forearm. The system took any value above 40% of
the maximum value seen during calibration to mean that the
left hand had been squeezed. We empirically selected 40%
from results in pilot studies. The system would then “sleep”
for a quarter-second before attempting to detect another
left-hand squeeze. We enforced this “silent” period to pre-
vent unintentional rapid sequences of selections.
Figure 5. Software mockup of a portable music player

Results
In both parts of our experiment, we collected gesture exam-
ples to train our recognizer and then asked participants to
complete tasks using those gestures in a two-handed tech-
nique. For each part, we examine the average accuracies
our system achieved in classifying finger gestures.
While each part of the experiment was conducted with a set
of four finger gestures, we also present an offline analysis
for Parts A and B of a gesture recognizer that only uses the
first three fingers (index, middle, and ring) to demonstrate
the potential tradeoff of gesture richness against classifica-
tion accuracy. We chose the pinky finger as the finger to
remove in this analysis because participants reported that it
was the most uncomfortable to manipulate.
Part A: Hands-Free Finger Gesture Recognition
As describe above, variability in arm posture (particularly
twisting of the forearm) presents a challenge for accurate
finger gesture classification. To explore this issue, we
trained the gesture recognizer in each of three postures in-
dependently, and performed an offline analysis testing each
recognizer with the test data from the other two postures.
As shown in Table 1, the system performed best when clas-
sifying pinch gestures using training data that was gathered
in the same posture. Furthermore, training transferred more
effectively between postures that were more similar. This
can be seen by grouping these results by distance (in
amount of arm rotation) between training and testing post-
ures. Distance zero represents training and testing on the
same posture. Distance one represents a small rotation
away, that is, either of the extremes to the midpoint or vice
versa. Distance two represents training on one of the ex-
treme positions and testing on the other.
The mean accuracy for distance zero is 77%, while distance
one classifies at 72% and distance two at 63%. A univariate
ANOVA on classification accuracy with rotation distance
as the only factor shows a main effect of distance
(F2,105=5.79, p=0.004). Posthoc tests with Bonferroni cor-
rection for multiple comparisons show this effect driven by
significant differences between distance zero and distance
two (p=0.003) and marginally between distance one and
distance two (p=0.086). Note that a random classifier would
be operating at about 25% for the four-finger gestures.
However, when all of the training data is used (75 blocks)
to train the gesture recognizer, instead of training data from
a single posture, the average accuracy over all of a person’s
test data is 79% with a standard deviation of 13% (see Fig-
ure 6). This demonstrates that training in a variety of post-
ures could lead to relatively robust models that find the
invariants and work well across the range of postures. Ex-
ploring more complex methods of modeling posture inde-
pendence remains future work. Reducing the gesture recog-
nizer to just the first three fingers increased this accuracy to
85% with a standard deviation of 11%.
Part B: Hands-Busy Finger Gesture Recognition
Participants performed finger gestures both sitting down
with a travel mug in their hand and while standing with
laptop bags in their hands. The system attempted to classify
gestures both when the participants did and did not have
visual feedback from the recognizer.
When participants held a travel mug in their hand, the four-
finger recognizer attained an average accuracy of 65%
without visual feedback (see Figure 7). Mean classification
improved dramatically, to 85%, with visual feedback. A
two-way ANOVA (finger × presence/absence of visual
feedback) on classification accuracy revealed that the re-
sults with visual feedback were significantly higher than
without (F1,10=24.86, p=0.001). The system also classified
much more accurately when only classifying among three
fingers instead of four: 77% without feedback and 86%
with feedback.
Participants spent a mean of 1.61 seconds between gestures
without visual feedback. This slowed to a mean of 3.42
seconds when they had visual feedback. An ANOVA re-
vealed a main effect for feedback (F1,10=13.86, p=0.004).
While holding a bag in each hand, the system classified
participants’ four-finger gestures at an accuracy of 86%
without visual feedback and 88% with visual feedback (see
Figure 7). When the classification was reduced to three
fingers, the system’s accuracy was better: 91% without vis-
ual feedback and similarly 90% with feedback.
On average, participants waited 1.69 seconds to squeeze
their left fist when there was no visual feedback. This in-
creased to 2.67 seconds when they had visual feedback of
Train
Test
Left Center Right
Left 78% 72% 57%
Center 70% 79% 74%
Right 68% 73% 74%
Table 1. Classification accuracies among pinch postures,
averaged across all users. Chance classification for this
four-gesture problem is 25%.
Figure 6. Mean classification accuracies for pinch gesture.
Error bars represent standard deviation in all graphs.
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
4Finger 3Finger
Hands‐FreeGestureAccuracy

the system’s current recognition result. A two-way
ANOVA (finger × presence/absence of visual feedback) on
completion time showed that the difference in feedback
conditions was significant (F
1,10
=19.77, p=0.001).
These results suggest that there is a time-accuracy tradeoff
for visual feedback. Participants were probably spending
time inspecting the feedback and making corrections to
increase overall accuracy. In future work, we would like to
explore less intrusive methods of providing feedback.
Part C: Portable Music Player Application Recognition
In the portable music player application, participants com-
pleted five blocks of three tasks with both the mug and
bags. For each of these tasks, we recorded whether they
selected the correct song, how many navigation steps they
used above the minimum steps required to select the correct
song, and how long it took them to complete each task.
In the travel mug scenario, two of the participants found
that the system’s classification of their pinky finger did not
work well enough to effectively complete the portable mu-
sic player tasks. We removed this data from our analysis.
When navigating the three-level hierarchical menu to select
a song, participants on average selected the correct song
85% of the time with bags in their hands and 87% of the
time while holding a travel mug. A failure was selecting
any song besides the one specified. On average participants
spent 45 seconds (median 39 seconds) navigating the menus
through an average of 15 gestures per task with bags, and
58 seconds (median 40 seconds) through an average of 14
gestures with the mug. The goal of this phase of the expe-
riment was to demonstrate that our real-time recognition
system functioned well enough to be used in an interactive
system. Among our participants some found it somewhat
difficult to control the music player, while several stated
that it worked very well for them and were interested when
this might be released as a commercial product.
DISCUSSION
We have explored the feasibility of building forearm mus-
cle-sensing based finger gesture recognizers that are inde-
pendent of posture and shown that these recognizers per-
formed well even when participants’ hands were already
holding objects. In this section, we discuss the implications
of these results for application design.
Posture Independence
The results from Part A suggest that while training data
from one arm posture is most useful in recognizing gestures
in the same posture, it is also possible to use our techniques
to train a single gesture recognizer that works reasonably
well in multiple arm positions. This suggests that electro-
myography based interactions could be deployed without
constraining wrist and hand positions. We feel that this is a
major step toward enabling real-world applications, particu-
larly applications in mobile settings. Users interact with
mobile devices in a variety of body postures (seated, stand-
ing, walking, etc.), and we would therefore expect a similar
variety of postures in the gesturing hand. Requiring a user
to train a separate classifier for multiple hand positions
would be costly, hence we are encouraged by our results
demonstrating the feasibility of cross-posture training.
Hands-Busy Interaction
Traditional input modalities take advantage of our dexterity,
motor ability, and hand-eye coordination. However, in
many scenarios we have to choose between our everyday
behavior and manipulating a physical input device. In these
scenarios, muscle-computer interfaces leveraging gestures
that can be performed while our hands are already gripping
an object provide an opportunity for computing environ-
ments to better support hands-busy activities such as when
using a mobile phone while walking with a briefcase in
hand or operating a music player while jogging. The results
of Part B of our experiment demonstrate the possibility of
classifying gestures involving individual fingers even when
the whole hand is already engaged in a task, and even when
the arm is supporting a heavy load.
Quantity of Training Data and Classification Accuracy
Figure 8 shows that even with limited training data (10
blocks or approximately 70 seconds), average accuracies
exceed 80% for four-finger classification, suggesting that
the required amount of training for a muscle-computer in-
terface would be on par with that typically required to train
a speech recognition system. Future work will explore
building cross-user models that would allow instantaneous
use of our system without per-user training, leveraging per-
user training only to enhance performance.
Cross-User and Cross-Session Models
We trained and tested our classifier for a single participant
in a single session as is common with similar technologies
such as brain-computer interfaces [10, 19]. Future work will
evaluate the degree to which classifiers can be re-used
across sessions, and will focus on automatically configuring
a classification system without careful sensor placement.
Interaction Design Issues
Even if a system can recognize individual gestures with
reasonable accuracy, deployment in real-world scenarios
Figure 7. Mean classification accuracies of hands-busy ges-
tures. Error bars represent the standard deviation.
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Mug Bags
Hands‐BusyGestureAccuracy
4FingernoFeedback
3FingernoFeedback
4Fingerw/Feedback
3Fingerw/Feedback

still requires careful consideration of appropriate interaction
techniques. Here we explore some of the design issues re-
lated to using muscle-computer interfaces for input.
Visual Feedback: Speed and Accuracy
Our experiments demonstrate that the proposed gesture set
can be accurately recognized via muscle-sensing in the ab-
sence of visual feedback, which is critical to many applica-
tions, including nearly all hands-free mobile scenarios.
However, visual feedback makes the system more predicta-
ble and gives users an opportunity to adapt their behavior to
that of the recognition system. For example, participants
could experiment with finger position or exertion to im-
prove recognition. This can be seen in Part B of our expe-
riment where participants held a travel mug in their hands.
The average accuracy of the system was much higher when
participants had visual feedback. However, this came at the
cost of reduced speed. On average, participants spent more
time performing each gesture, as they adjusted their ges-
tures until the system made the correct classification. This
speed-accuracy tradeoff should be considered carefully in
the context of an application. In applications where an error
can easily be undone and the gesture repeated (e.g., in a
mobile music player), the higher speed that comes from
feedback-free gesture input may justify an increased error
rate. In contrast, in applications where an incorrect gesture
might be more costly (e.g., when controlling a mechanical
device or playing a game), the decreased speed that comes
from using visual feedback might be reasonable.
Engagement, Disengagement, & Calibration
A wearable, always-available input system needs a mechan-
ism for engaging and disengaging the system. We do not
want the system to interpret every squeeze or pinch action
as a command. In our experiment, we used the left hand to
support engagement and disengagement, and we feel that
this separation of tasks across the two hands is a reasonable
option for real applications. However, it would be worth-
while to look at how engagement and disengagement might
be supported by sensing only one hand. In particular, is
there a physical action unique enough to be robustly classi-
fied during everyday activity such that it can be used as an
engagement delimiter? One example of such an action
might be squeezing the hand into a fist twice in succession.
In our limited exploration of this topic, a fist clench has
appeared to be easily distinguishable among other typical
movements, so this may be a starting point for future mus-
cle-computer interfaces.
Multi-Finger Interactions
Our experiments focused on recognition of single gestures
performed one at a time. The system’s ability to recognize
these gestures indicates that we could develop interaction
techniques that rely on sequences of gestures. It would also
be interesting to compare such sequenced interaction with
simultaneous performance of several gestures at a time. For
example, how does recognition performance compare when
doing an index finger pinch followed by a middle finger
pinch, vs. a simultaneous index and middle finger pinch.
Apart from recognition performance, users’ perception and
performance of these different styles of multi-finger inte-
ractions must also be considered carefully.
Ongoing and Future Directions
Air-Guitar Hero
Encouraged by the results, we developed an application that
allows a user to use our muscle-computer interface to play
the Guitar Hero game. In Guitar Hero, users hold a guitar-
like controller and press buttons using both hands as the
system presents stimuli timed to popular music. Using our
muscle-computer interface, users can now play with an
“air-guitar”. A user controls four buttons with our pinching
gestures and moves the opposite wrist in a strumming mo-
tion. Informal tests of the system show that users are able to
complete the easy mode of the game. We demonstrate this
system in our video figure.
Wireless Electromyography
Although we extended previous work by not tethering
people’s arms and hands to specific orientations or surfaces,
our experiment was conducted in a lab using a wired elec-
tromyography device, and we have yet to validate our clas-
sification approaches in scenarios with more variable ges-
ture execution. To this end, we have recently created a
small, low-power wireless prototype muscle-sensing unit
(see Figure 9). Each of these units is equipped with four
electrodes (two differential electromyography channels)
sampling at 128 Hz, and multiple units can be used simul-
taneously. We are currently working to put this wireless
unit into an armband form factor with dry electrodes.
Figure 9. Classification accuracy versus blocks of training data
for four finger gestures with bags in hand. Each training block
takes seven seconds for a four finger classifier.
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 5 10 15 20 25
BlocksofTrainingData
QuantityofTrainingDatavs.Accuracy
Figure 8. Our wireless EMG device prototype, weighing
five grams and measuring 26x18x8mm.

CONCLUSION
Our work demonstrates that muscle-sensing can be used to
accurately classify a useful variety of finger gestures, even
when the hands are under load. It also shows that classifica-
tion can be done in real-time, thus making forearm muscle-
sensing viable for human-computer interaction, in contrast
to previous work that relied on off-line analysis. Further-
more, it highlights the tradeoff between speed and accuracy
that results from providing users with immediate visual
feedback. Finally, it introduces a novel bimanual technique
for accurate engagement/disengagement of the recognizer, a
crucial aspect of making muscle sensing usable for interac-
tive tasks. In addition to the formal experimentation and
results, we have demonstrated more holistic interaction via
our portable music player application and a prototype game.
ACKNOWLEDGEMENTS
We thank Jonathan Lester and Carl Ringler for assistance
with the wireless device, Meredith Skeels for fabricating
some armbands, as well as John Platt and Jay Stokes for
machine learning advice and tools.
REFERENCES
1. Burges, C. 1998. A tutorial on support vector machines
for pattern recognition. Data Mining and Knowledge
Discovery, 2, 121-167.
2. Costanza, E., Inverso, S.A., Allen, R., & Maes, P. 2007.
Intimate interfaces in action: Assessing the usability and
subtlety of EMG-based motionless gestures. CHI ’07,
819-828.
3. Englehart, K. & Hudgins, B. 2003. A robust, real time
control scheme for multifunction myoelectric control.
IEEE Trans Biomedical Engineering, 50(7), 848-854.
4. Farry K., Walker I. & Baraniuk R. G. 1996. Myoelectric
teleoperation of a complex robotic hand. Proc IEEE Int
Conf Robot Autom, 775-788.
5. Greenberg, S. & Fitchett, C. 2001. Phidgets: easy devel-
opment of physical interfaces through physical widgets.
UIST '01, 209-218.
6. Ju, P., Kaelbling, L. P. & Singer, Y. 2000. State-based
Classification of Finger Gestures from Electromyo-
graphic Signals. ICML ’08, 439-446.
7. Kiguchi, K., Tanaka, T. & Fukuda, T. 2004. Neuro-
fuzzy control of a robotic exoskeleton with EMG sig-
nals. IEEE Trans. on Fuzzy Systems, 12(4), 481-490.
8. Lakshmipathy, V., Schmandt, C., and Marmasse, N. 2003.
TalkBack: a conversational answering machine. UIST ’03.
9. Lanyi, X. & Adler, A. 2004. An improved method for
muscle activation detection during gait. Canadian Conf.
on Electrical and Computer Engineering, 357-360.
10.Lee, J.C. & Tan, D.S. 2006. Using a low-cost encepha-
lograph for task classification in HCI research.UIST ’06,
81-90.
11.Lyons, K., Skeels, C., Starner, T., Snoeck, C. M., Wong,
B. A. & Ashbrook, D. 2004. Augmenting conversations
using dual-purpose speech. UIST '04, 237-246.
12.MacKenzie, C. L. and Iberall, T. 1994. The Grasping
Hand. Amsterdam: North-Holland, Elsevier Science.
13.Merletti, R., & Parker, P.A. 2004. Electromyography:
Physiology, engineering, and noninvasive applications.
John Wiley & Sons: Hoboken, New Jersey.
14.Naik, G.R., Kumar, D.K., Singh, V.P. & Palaniswami,
M. 2006. Hand gestures for HCI using ICA of EMG.
HCSNet Workshop on the Use of Vision in HCI, 67-72.
15.Peleg, D., Braiman, E., Yom-Tov, E. & Inbar G.F. 2002.
Classification of Finger Activation for Use in a Robotic
Prosthesis Arm. Trans Neural Syst Rehabil Eng, 10(4).
16.Platt, J. 1998. Sequential Minimal Optimization: A fast
algorithm for training support vector machines. Micro-
soft Research Technical Report MSR-TR-98-14.
17.Raez, M.B.I., Hussain, M.S. & Mohd-Yasin, F. 2006.
Techniques of EMG signal analysis: detection,
processing, classification, and applications. Biological
Procedures Online, 8, 11-35.
18.Saponas, T. S., Tan, D. S., Morris, D. & Balakrishnan,
R. Demonstrating the feasibility of using forearm elec-
tromyography for muscle-computer interfaces. CHI ’08,
515-524.
19.Sassaroli, A., Zheng, F., Hirshfield, L.M., Girouard, A.,
Solovey, E.T., Jacob, R.J.K. & Fantini, S. 2008. Dis-
crimination of Mental Workload Levels in Human Sub-
jects with Functional Near-Infrared Spectroscopy. J In-
novative Optical Health Sciences, 1(2), 227-237.
20.Schlesinger, G. Der Adechanische Auflau der kunstli-
chen Glieder. 1919. Ersatzglieder und Arbeitshilfen, 1.
Springer Verlag, Berlin.
21.Tenore, F., Ramos, A., Fahmy, A., Acharya, S., Etienne-
Cummings, R. & Thakor, N. 2007. Towards the Control
of Individual Fingers of a Prosthetic Hand Using Sur-
face EMG Signals. IEEE EMBS.
22.Wang, G., Wang, Z., Chen, W. & Zhuang, J. 2006.
Classification of Surface EMG signals using optimal
wavelet packet method based on Davies-Bouldin crite-
rion. Med Biol Eng Comput 44, 865-872.
23.Wheeler, K.R, Chang M.H. & Knuth K.H. 2006. Ges-
ture-Based Control and EMG Decomposition. IEEE
Trans. on Systems, Man, and Cybernetics, 36(4).
24.Wilson, A. 2005. PlayAnywhere: a compact interactive
tabletop projection-vision system. UIST ‘05, 83-92.
25.Yatsenko, D., McDonnall D. & Guillory, S. 2007. Si-
multaneous, Proportional, Multi-axis Prosthesis Control
using Multichannel Surface EMG. IEEE EMBS.