Presentation of the ToCHI paper at the 33rd Annual ACM Conference on Human Factors in Computing Systems (CHI 2015) in Seoul. We capture non-uniformity of the movement space of the arm in 11 clusters. The clusters are based on muscle co-activations necessary to produce a movement. However, the clusters turn out to be distinct also in 3D space and orientation. Each cluster has specific performance and physical ergonomics properties.
The paper: http://dl.acm.org/citation.cfm?id=2687921
P.S. The slides contain videos and animations which are not displayed by online-viewer, so it is better to watch them offline after downloading.
5. Fitts’ law assumes that the movement space is
uniform
5
MT = 𝑎 + 𝑏 log2(1 +
𝐷
𝑊
)
W
D
Non-uniformity is important for interfaces with large movement space
6. We define neuromechanical equivalence
classes based on muscle co-activation
6
http://www.nature.com/nrn/journal/v5/n7/images/nrn1427-i1.jpg
A
B
C
A B
A C
7. Research objectives
1. Identify equivalence classes of movements for
any given movement space
2. Describe performance and ergonomics
characteristics per movement cluster
3. Inform design of an input methods
7
All movements
equivalent
Every movement
exclusiveDesired clustering:
• few clusters
• easily interpretable
• covers all statistically
significant effects
8. Background
Non-uniformity Performance models Muscle dynamics
• Movement location
• Movement direction
• Movement amplitude
• Speed-accuracy trade-
off models:
• Uniform
• Non-uniform, but
non-physiological
• Movement location
• Movement direction
• Trajectory profile
• Velocity profile
impact
Muscle recruitment
impact
• Movement amplitude
• Movement direction
Muscle activation pattern
impact
[Freund1978, Soechting1995,
Adamovich1999, Gribble2003,
Baud-Bovy1998, Caminity1990,
etc.]
[Fitts1954, MacKenzie1992,
Grossman2004, Cha2013,
Plamondon1997, etc.]
[Gielen1985, Cooke1994,
Koshland1994, Flanders1991,
Wierzbicka1985, etc.]
8
12. 2 and 3: MoCap and biomechanical simulation
yield muscle activations and ergonomics indices
12
13. Muscle activation pattern for a movement is a
multidimensional vector of activation values
13
Time, ms
Activation
level
Acceleration Deceleration
Muscles
0
1
100 200
20. 2 7 10
The clusters exhibit up to 4-fold differences in
total muscle activation
20
All 1 3 4 5 6 8 9 11
Total muscle
activation
Clusters
21. Depending on a cluster different muscle
groups are recruited
21
22. How to apply the clusters to a design task
22
1. Identify movements
involved
2. Map them to 3D space
3. Scope their direction
and length
4. Identify the best-
matching cluster
5. Look at ergonomics and
performance of the
selected cluster
6. Decide whether such
properties suit your
input method
Hi all. My name is Myroslav Bachynskyi. In this presentation you will see what movements are equivalent, and how the equivalence classes can be used to inform the design of new interfaces.
I start with motivation, why we need this method.
Imagine you want to design a menu for public display
It can be placed anywhere on the screen, but how to choose the best placement.
In this particular example designers put it at the top of the screen following patterns from desktop or web interfaces.
As result user has to move his arm at the top to navigate through menu.
We show, that this is bad decision, and lowering the menu by only 35 cm reduces muscle recruitment in up to 2.5 times.
By creating an input method, designer defines which movemnts will user perform during the interaction. It is hard to design interfaces with large movement space as it is complex and non-uniform. We split the space of all possible movements into space segments in which movements are equivalent. Equivalent movement exhibit similar performance and muscle usage properties, and designer can choose which cluster suits needs of input method the best.
Nowadays we have many devices which provide large movement space, however they are analyzed considering Fitts’ law which assumes that the whole movement space is uniform.
Having movement space classes of equal movements allows to use Fitts’ law in each cluster separately and increase the accuracy.
So which movements are equivalent?
For example we have movement A and B. Can they be considered equivalent? And what about A and C? To answer this question we have to look first how our brain controls movements.
Our brain sends a signal through spinal cord and motor neuron to activate muscles, Muscles generate active force and produce skeletal movement. We define two movements as equivalent, if they are produced by the same group of muscles activated similarly.
The CHALLENGE is how to acquire such datasets and IDENTIFY such clusters and whether such cluster exist at all
Our objectives in this work were to identify the clusters, describe them and improve an interface with their help.
The question is how many clusters there should be.
On the one hand, if you have too many clusters, then ..it is hard to apply and deal with them
On the other, … too few clusters will be large and too heterogeneous and hard to interpret.
So we aimed for few well interpretable clusters with significant differences among them.
I\m summarizing here related work under three categories.
The big message is that although movement space is non-uniform, non-uniformity can be characterized by movement location, direction and amplitude.
First, non-uniformity of movement has been found in multiple studies.
Second, performance models assume UNIFORmity> The only exception is kinematic theory of Plamondon, however it does not have physiological base which would allow split space into clusters
Third, movement location and direction predefines muscle group which executes it, direction and amplitude influences individual muscle activations
Here you see the pipeline for muscle coactivation clustering.
First you collect motion capture data for your input method, covering all parts ofhte mvoemetn space
Then you have two standard steps of biomechanical simulation
Then we cluster the muscle coactivation pattern
We describe clusters to be easily used for input method design task
1. Lets firs tlook at the data collection in a motion capture laboratory.
By inverse kinematics we fit musculoskeletal model to the recorded cloud of points, and then with static optimization identify muscle activation patterns
This is an example of computed muscle coactivation for an aimed movement. Such pattern described by single high-dimensional vector is the input data to our clustering.
To go from here to cluster, our hypothesis is that similar muscle activation patterns belong to the same equivalence class.
Now we move to results
Here you see.
This has been achieved by using established in statistical learning techniques… We tried out several …
As goodness metric we use within to between ratio, pearson gamma, dunn index etc.
Here this figure shows …
Here is another view to the clustering
This has been annotated to describe the distinct muscles involved
For example, …
The big picture is that although all muscles contribute to branching, there is still one which dominates the rest.
Here we see the movements in each cluster.
This confirms that clusters are meaningful
Here you can see Fitts’ model for the whole movement space. Modeling Fitts’ law for each cluster separately improved the model fit on average by 2%.
And here we see that Fitts law based throughput differences reach of up to 37%
They also show LARGE differences in TOTAL muscle activation which is correlated with fatiguability of momvement
Here we look on little bit more details of cluster muscle recruitment. As you can see movements within each cluster are produced by distinct muscle group.
The clusters are visually summarized on single page and can be applied in 6 simple steps:
Describe steps
Here is an example.
Lets assume you are designing … The question is whete…
The clusters show that. ..
In this case the movements for this are mapped to cluster 1, 9 and 11, alternatively 35 cm lower is the cluster 2, which provides slightly lower throughput, but requires more than twice less total muscle activation
We have shown a way to identify neuromechanical equivalence classes for a movement space. The clustering is based on co-activation patterns of muscles and it can be used for:
Summarizing differences in movement space
Solving placement problems in design
THE PAPER
Make this a gallery such that each issue (e.g., direction) is its own block?
Start with the big picture
To compute clusters we use R packages and well-established statistical learning techniques
We use Euclidean distance
Ward