Guest Post by Victoria Preston
Current prosthetic technology requires patients to undergo extensive training in order to use a replacement arm or leg, so that they can adapt to the technology. Achieving a basic level of competency with a new limb can take up to 2 months. Rejection of a prosthesis is not only a physical decision (discomfort, aesthetic), but can also be psychological (feelings about the prosthetic) or monetary (cost proportional with intuition).
Prosthetics have been dated back as far as 3,000 years, and the main technology – structure, bearing, actuation, suspension – has remained relatively constant to modern day prosthetics. The major changes today lie in the materials, actuation techniques, and smarts of the limb. The most advanced prosthetics today can ‘guess’ what a user wants to do, and act upon it, relying on expensive sensors to pull it off. The need for high-quality sensors could be subverted, however, by adding a layer of machine learning to lower-quality sensors. Check out this group of Illinois undergrads making new strides to develop prosthetics with real, adaptable, smarts. By using the same simple tech and advances in rapid prototyping, and marrying it with machine learning, real good can be done right now.
One of the most celebrated advancements in prosthetics include the availability of myoelectric sensors, which can read in and translate electrical impulses from the user to some particular motion in the prosthetic. To implement a myoelectric prosthetic, the patient needs to have precise measurements for sensor location and special muscle training. The differences between a myoelectric and a body-powered limb come down to smarts. The myoelectric is a ‘smart system’ which can parse the muscular desires of the user, and offer fine-tuned response for a cost of $30,000-$40,000. A body-powered prosthetic has only mechanical actuation controlled by a user via shoulder or body motion to pull strings and levers for a cost of several hundred.
The desirable characteristic of a prosthetic limb is smarts, with cost, weight, comfort, and intuition as a set of sliders that the user wants to optimize. There immediately comes to mind this idea of marrying simple, low-cost sensors with machine learning algorithms and rapid fabrication to create a system that adapts to a user in fewer training sessions, could physically grow with a user, and personalize to each individual’s nuances and desires.
There is a growing network of DIY prosthetic limbs out there, The Open Hand Project and E-nabling the Future are active examples. All focus on this idea that the simple technology exists to create affordable prosthetics that can satisfy an individual user without breaking the bank. Uniting these projects under the umbrella of machine learning could strengthen the power of the rapid-prototyping and personalization aspect of this solution. Training via reinforcement learning techniques such as apprenticeship learning or learning from demonstration connects nicely with the idea of personalization. Given a series of motions with particular objectives, a user can train their new limb throughout time to respond to their particular nuances.
The whole idea of making hardware “more personal” lies at some of the draw of machine learning. Machine learning is not only good for parsing large datasets, but can be used in a powerful way to do immediate real-world development and change. With self-examination and evaluation of engineering strategies, a mass of technology may be able to adapt to the human, rather than require the human to adapt to the technology.