Uncategorized

How does MYO wearable gesture control work?

Myo literally means muscle and the gesture control from Myo is not a leap in innovation in this area but rather (what it seems) a lot of incremental improvements in a range of different areas. First and foremost, the first affordable medical prosthesis technology available for consumers.

There are three main parts to building a myo-electric device (prosthesis, the Myo or any other type of muscle controlled device) 1. Signal Acquisition 2. Signal Amplification 3. Signal Analysis and interpretation

1. To acquire a muscle signal most devices rely on the small electrical signal that muscle cell produces. More commonly known as Electromyography (EMG for short). An EMG electrode is attached to your skin which detects the subtle differences in the electric signal, or the electric potential difference over time.

Now these EMG electrodes usually looks like small stickers, and why these are always stickers i’ve no good answer to, i think it is because any other form factor haven’t been necessary in the present applications. But it is of utter importance that there are excellent contact between the skin and the electrode during the signal acquisition to ensure a correct reading. Some medical EMG readings are actually done intrusively, that is a needle (used as an electrode) is stuck into the muscle itself to get a even better reading. MYO or Thalmic have come up with a electrode which is a non stick type but still gets appropriate readings regardless, it might also be that their electrodes or readings are somewhat redundant, which would almost eliminate the problem of bad contacts (and get two or more independent readings, which will prove important in the next steps). Medically sometimes a conductive gel is also used on the skin to improve electrode contact, which is certainly not feasible in a consumer product.

Moreover, traditionally the placing of the electrodes have an impact on the readings and the results, which makes the MYO approach even more challenging since they’re limited to one confined area of the arm. This is the first incremental step.

2. Amplifying the signal is crucial, not to say necessary to be able to read it accurately, I will not be as elaborate on this part, although important, amplifying electrical signal is something that we are quite good at in the scheme of things. I would use some kind of OP-amp for this. Further before the processing is done, one might incorporate some kind of high band and/or low band filtering (hardware, Myoelectric Controlled Prosthetic Hand). Although i’m speculating but if i had the power to this this in software or in a FPGA of some sort i think i’d leave the signal as is and do all processing in “software” (someone out there might have a better suggestion though?). I don’t think Thalmic is focusing on this part, of course they make sure they have a good quality signal but not putting their energy here.

3. Processing the signal , now this is an interesting and tricky part. To make up for some of the drawbacks of the design i’ve talked about in the earlier parts, this is where you have to make up for it to make the device a great experience! MYO have developed a Machine learning algorithm to detect a number of gestures (~20 something it is said). Further you’re specific muscle/gesture activation signal will be used as training data for this model to increase its accuracy. I’m no expert in machine learning but i will take a stab at it. I would do it one of two ways (a current Kaggle competition that one could draw inspiration from is the The Marinexplore and Cornell University Whale Detection Challenge, which is detecting the presence of whale given a sound signal, one could easily see this problem as a similiar classification problem, given a muscle signal, classify which of 20 gestures this is).

– Random forest decision with 20 “classes”. Train a random forest for the input signal and using some feature scaling to get a better result.

-Deep belief network with 20 outputs, one for each gesture. Train a DBN with Restricted Boltzmanns.

Of course there is a number of other possibilities. Further for both these approaches appropriate features might have to be extracted (MFCC for instance).

One particular challenging thing in this part is of course the noise, in EMG “Crosstalk” is something that is a reoccuring problem, which relates to the noise signal from a co-contract muscle or from a muscle which is not contracted.

Hopefully this answer give you something, some of this might of course be completely wrong since i have no direct insight into the making of the Myo, but if the MYO works as well as some have said (first hand on HN for instance) and  what the concept video promises we are in for a treat and Thalmic have actually taken quite an exciting and big leap in EMG and “prosthesis” technology, at least in the consumer space. I’m filled with admiration either way and i have of course preordered one 🙂

Order Myo for $199: Myo Gesture Control Armband (Black)

 

Leave a Reply

Your email address will not be published. Required fields are marked *