A Self-Sensing Shape Memory Alloy Actuator Physical Reservoir Computer

Abstract

A self-sensing shape memory alloy actuator is harnessed as a computational resource by utilizing it as a physical reservoir computer. Physical reservoir computing is a machine learning technique that takes advantage of the dynamics of a physical system for computation. Compared to recurrent neural networks, this architecture can be both fast and efficient with a cheaper training procedure. A shape memory alloy actuator is designed, fabricated, and tested for processing information. Voltage variation along the shape memory alloy wire is used as the reservoir's nodes. The physical reservoir is then used to predict the future trajectory of the actuator's end effector under various driving signals. This self-prediction method is also reconfigurable, as demonstrated by training the reservoir for one waveform but testing it for a different one. A nonlinear autoregressive moving average prediction task was also used to highlight the physical reservoir computer's abilities. Following this methodology, the soft actuator can be used for actuation and computation at the same time without altering its design.