Deep-LSTM-Based Human-Robot Collaboration Control With Static-Dynamic Calibration

Abstract

Using electromyography (EMG) signals to estimate human 3-D upper limb motion intention is a promising technique in human–robot collaboration systems. This technique is challenging due to the nonlinear components within the 3-D upper limb neuromusculoskeletal system and the intricate dynamic online environment in the human–robot coupling systems. In this article, a deep long short-term memory network was proposed to establish the nonlinear EMG-to-3-D-force model for an upper limb cable-driven robot (UL-CDR). To enhance the model's adaptability to dynamic environments, we proposed a static-dynamic calibration method that initially trains the network using static measurement data and then fine-tunes it using dynamic data. Subjects were recruited to track reference trajectories for 160 trials with the assistance of the UL-CDR. Results showed that the proposed approach got better tracking accuracy than the method with only static calibration, demonstrating the effectiveness of the proposed method in transitioning between static and dynamic environments and potentially enhancing the performance of human–robot collaboration.