Shape-Based Force Prediction of a Soft SMA Planar Actuator Using Machine Learning

Soft actuators driven by shape memory alloy (SMA) wires are capable of undergoing a large deformation and offering a fast response, which allows them to assemble soft robots for diverse operations. However, their further applications, especially in scenarios of force control and compliant operation, are limited due to the lack of force-sensing capability. The main challenges in force sensing of SMA soft actuators include two factors: 1) their deformable structures and relatively small volumes limit the use of rigid force sensors and 2) it is difficult to derive the loaded external force through mechanical modeling. In this work, the magnitude and the direction angle of a 2-D external force were predicted according to the shape information of a soft SMA planar actuator (SPA) using a force-magnitude prediction model and a force-direction prediction model, respectively. Both the two prediction models were trained by a shallow neural network based on the shape information of the SPA collected automatically through experiments. The effectiveness of the two trained models was proved through validation analysis and prediction experiments. In the magnitude range 0–400 mN [nearly four times the SPA’s weight (10.5 g)] and the direction angle range 0°–60° of external force, mean absolute prediction errors of 16.56 mN and 2.06° were achieved, respectively. Finally, two application experiments employing the prediction capabilities in the magnitude and the direction angle of external forces were presented. These results proved that SPAs can utilize their deformable structures to achieve both actuation and 2-D force perception. This work is expected to improve the sensing capability and extend the force-related application range for SMA soft actuators.