228. Harvard’s Soft Robot Gives Movement Back to ALS & Stroke Patients

Engineering & simulationintention detector model94% accuracy within 15-20 minutes

A machine learning model was trained and calibrated for each individual user to learn their specific movement patterns and predict their intent with high accuracy.

“called their intention detector model. And it's a machine learning model that they train uniquely for each user, they have to calibrate this model for each user, and it learns the user's movement patterns. It say, it takes about 15 to 20 minutes, but within that 15 to 20 minute period, it gets up to 94% accuracy…”

Engineering & simulationhistoresis modelreacts in under half a second

A physics-based model was combined with the machine learning model to dynamically calculate the minimum air pressure required to support specific movements, preventing over-assistance.

“the 2nd part of this comes from a physics model that's bolted onto the machine learning model, which they call their historesis model. And this helps to understand how much air pressure is needed to help support and lift and hold the arm depending on the movement that's intended.”