Michigan Strength Augmenting Robotic Exoskeleton (M-STARX) Foot Sensor for State Classification

Abstract

Powered exoskeletons are an important growing technology that can be used to augment users' physical abilities or provide gait rehabilitation/locomotion assistance. However, they require robust and responsive control algorithms that allow the exoskeleton to match a user's movement. The Michigan Strength Augmenting Robotic Exoskeleton (M-STARX) team requested that we develop a sensing system to be incorporated into their exoskeleton's foot to classify behavioral states: standing, walking, and running. This requires an original algorithm to process the raw data and output the behavioral state so the M-STARX team will be able to create responsive controls that mimic and assist the user's movement. For this project, we aim to achieve a 95% accuracy in state classification while maintaining a low manufacturing cost and easy integration with M-STARX's exoskeleton. Most of the specifications were determined based on the sponsor's budget and requests. The final design was an inertial measurement unit placed inside a 3D printed housing that transmits a signal of the accelerations and rotational velocities in the x, y, and z directions. The signal is then processed using a fourier decomposition to find the dominant frequency and magnitude. Then the algorithm determines the behavioral state based on thresholds for the signal frequency and magnitude. Verification and validation work was done using data collected by people incrementally going from standing to running then back down as well as moving at randomized speeds. The algorithm's classification was then compared to time stamped video data to verify the accuracy. It was able to achieve an 80% accuracy in state classification, but there is some future work that can be done to improve this. We concluded that increasing the sample rate or changing the high level algorithm into a neural network may be some ways to achieve the desired accuracy. Overall, the sensing system was able to achieve its goal of running in near real time (0.5 - 1.8 seconds delay) and classifying the behavioral state of the user.