Long Short-Term Memory (LSTM) Estimation of Chair Reaction Force during Sit-to-Stand Movements

To assist individuals with an assistive robot who have difficulty in sit-to-stand (STS) movements, many studies have derived an optimal STS movement that minimizes the load, such as joint torques estimated from body model simulations. However, reaction forces (RFs) become undeterminate due to the closed-loop mechanism when the body contacts multiple surfaces, including the ground, chair seat, and robot. Therefore, in previous studies, optimization methods are often performed after the individual's buttocks have left the chair. To address this problem, we propose an approach employing Long Short-Term Memory (LSTM) to estimate the vertical and horizontal RFs at the seat surface based on joint angles. We obtained data from three healthy participants, collecting the hip, knee, and ankle angles and the vertical and horizontal RFs. Three STS trials were employed as training data, while two additional trials were used for testing. The estimation results for the vertical RF proved to be suitable for the joint torque estimation, with the coefficient of determination exceeding 0.9 and a root-mean-squared error (RMSE) below 50 N. On the other hand, the estimation fell short of our expectations in the horizontal direction, possibly due to the considerable variability in the training data across trials.