SEMG-based continuous estimation of grasp movements by long-short term memory network

Abstract

Controlling a robotic hand or arm by surface Electromyographic signals (sEMG) is an important research direction. The present pattern-recognition-based control strategy can realize some myoelectric control but it is not as smooth as human hand. In this paper, we proposed a continuous estimation method for 6 daily grasp movements by Long-Short Term Memory Network (LSTM). In addition, we compared Sparse Gaussian Processes using Pseudo-inputs (SPGP) and Radial Basis Function Neural Network (RBF) with LSTM for continuous estimation of 6 grasp movements. These methods are tested on the NinaPro dataset and evaluated by Pearson Correlation Coefficient (CC) as well as Root Mean Square Error (RMSE) and Normalized Root Mean Square Error(NRMSE) between real and estimated joint angles. They can estimate 20 joint angles in hand synchronously. The average CC of LSTM (0.8402 ± 0.0009) is significantly higher than RBF (0.774 ± 0.0018, p = 0.0124) and SPGP (0.7883 ± 0.0005, p = 0.0075). The RMSE of LSTM ((5.89 ± 0.95)°) is less than RBF ((8.08 ± 2.911)°, p = 0.0213) and SPGP ((7.16 ± 1.178)°, p = 0.0495) for the 6 movements. In addition, the NRMSE of LSTM (15.27% ± 0.02%) is less than RBF (19.84% ± 0.06%, p = 0.0033) and SPGP (18.11% ± 0.02%, p = 0.0117).