Using data to continually improve a physical product

While the Myo armband is a physical device, much of its power comes from the software inside it. Since we’re able to continually update and change the Myo software, we use usage data to continually improve performance and make the product even better over time. Over the past year, our team has been collecting data from many different users trying the Myo armband, to build the algorithms we use to recognize gestures.

In a nutshell, performing data collection with the Myo armband consists of allowing users to wear the device while they follow a series of gestures.

myo-connect-screenshot

The reason we need to collect this data is that the structure and physiology of our muscles can vary from person to person and even day to day. The way one individual performs a fist may look the same as the individual next to them, but under the surface of the skin, there’s actually a lot more to it than meets the eye. This is precisely why collecting as much data, from as many individuals as possible, is so important.

signals
Varying EMG patterns for two people performing the same gesture.

In regards to the Myo armband, there are a variety of factors that can affect why the EMG (Electromyography) signals are different for every user. While most of the difference comes from variations in muscle anatomy, factors such as arm circumference, forearm strength, arm hair, and even body fat, have an influence on the signal measured at the surface of the skin. Reducing the impact of these variations has been a top priority for the design team working on the Myo armband.

In the earliest days of the Myo armband, data collection was limited to a small set of people. This typically resulted in functional and accurate gesture recognition from our test subjects, but we didn’t have a broad enough data set to cover the population. Although this limited testing was sufficient for that stage of development, we knew that we would need to extend our reach in order to really nail down the gesture recognition for consumer launch. The end goal was ultimately to have every gesture performed by the user be classified as the intended gesture. To do this, we needed to collect data for every possible method of doing each of our selected gestures.

wave-out-gesture
Two individuals performing the same gesture in different ways.

In order to work towards these goals, we set up camp at a variety of locations in order to gather a wider sample of the population’s EMG data. It’s incredibly important to have users that have never used the Myo armband try it out in order to record how they naturally perform the gesture. This allows us to gather organic data – the way a user thinks a gesture should be done – rather than the data from someone influenced into performing the gesture through instruction. Our algorithms analyze the organic data and make sense of the different methods of performing gestures, allowing us to classify all the different variations in the ways people perform each gesture.

Watching individuals learn how the Myo armband works for the first time is both rewarding and entertaining. Whether they’ve heard of the device before or are just being introduced to it for the first time, their reactions are priceless. Once they connect the dots and realize that their arm movement is actually controlling what’s happening on screen, their interest and curiosity about how it works only intensifies. We quickly realized that observing individuals using the Myo armband for the first time was also a form of data collection in itself. We were able to witness how people naturally performed gestures, what position their arm was in when performing a gesture, or how hard they assumed they had to clench their fist, etc.  These factors had been, for the most part, ignored when collecting data in-house or on individuals familiar with our gesture set. While this visual data collection won’t come into play when creating the algorithms, it does help us with the overall user experience and allows us to better inform new users on how to use the device when they first open the box.

While the data collection process has been a long and experimental journey thus far, we’re not quite done yet. Over the past 6 months Myo Alpha units have been going out to developers across the world and their set-up experience has allowed us to gather insight into how the first gestures are typically performed. We’ve also integrated an option into Myo Connect with the latest SDK release that allows developers to perform gestures purely for the purpose of allowing us to collect their data – and with over 10,000 developers and a Myo on their arms, this could possibly prove to be our largest data collection event yet! And last but not least, all final Myo units will have gesture data collection built-in, so we can continue to build an even bigger set of gesture data. Needless to say, in terms of data collection the more data the better, and as both the software and firmware will be upgraded consistently in the future, the performance of the gesture recognition will only improve from here.

Newsletter

Enter your email address and get all latest content delivered to your inbox every now and then.