Data-Driven Tracking of Wrist Motions for a Soft Exoskeleton Based on Twisted String Actuators

This paper presents experimental testing of a Twisted String Actuator (TSA)-based wrist exoskeleton integrating \MD{a data-driven feedforward command estimator within a position-tracking control architecture, for rehabilitation applications.} The device has three degrees of freedom (DOF), allowing for flexion/extension, abduction/adduction, and pronation/supination motions through the coordinated actuation of four motors. The exoskeleton structure is lightweight and compliant. Given the nonlinear transmission ratio of TSA and the user's specific tendon routing characteristics, \MD{a data driven feedforward command estimator is implemented.} Three regression models are trained to map desired joint angle trajectories to motor commands: a Multilayer Perceptron (MLP), a Feedforward Neural Network (FFNN), and a Gradient Boosted Tree ensemble (XGBoost). All models share a ten-dimensional input vector combining kinematic features and autoregressive motor state, achieving $R^2 > 0.996$ across all motors on the held-out test set. Experimental validation across 120 trials covering all six wrist movement directions confirms autonomous trajectory replay with mean RMSE below $2.6^\circ$, \MD{without requiring continuous IMU feedback during the replay phase.} The results demonstrate that integrating AI with TSA actuation is a viable strategy for \MD{accurate, continuous-sensor-free replay for wrist rehabilitation exoskeletons.