Enhancing the Online Estimation of Finger Kinematics from Semg Using LSTM with Attention Mechanisms

Simultaneous and proportional estimation of human finger kinematics using muscle interface has gained significant attention for human-robot interaction. Most existing researches focused on proposing novel estimation methods to achieve high estimation accuracy and validated them offline. However, the gap between the online and offline estimation has not been fully considered which may lead to unpredictable deterioration or failure in practical application. This paper quantifies the gap between online and offline estimation performance and the underlying factors for such gap. The temporal variability of muscle activities represented by surface electromyography (sEMG) is found to be more significant in online estimation which challenges the estimation model in computation complexity and temporal feature extraction. To improve the online estimation accuracy, this paper proposes to combine an attention module with the long short-term memory (LSTM) network which enables not only the global but also the local key information of sEMG signals being focused. The finger kinematics of five primary degrees of freedom (DoFs) is estimated from the sEMG recordings during some grasping tasks. Compared with the sole-LSTM model (CC=0.85,RMSE=9.41), the LSTM with attention mechanism (CC=0.92,RMSE=7.96) demonstrates an improvement in online estimation accuracy, presenting an effective way of kinematics decoding for human-robot interaction.