Reconstruction of robot teleoperation force information in simulated time-delay network based on LSTM-Transformer

Accurate force feedback is a prerequisite for high-transparency teleoperation, yet it is highly susceptible to distortion under time-varying network delays. To address this issue, this paper proposes a time-varying delay force feedback reconstruction method that integrates force sensor-based compensation with deep learning techniques. In this study, the least squares method is applied to estimate end-effector force sensor parameters for zero-offset removal and gravity compensation, followed by Kalman filtering to reduce noise in the six-axis sensor data. To simulate network disturbances, a bidirectional communication channel with stochastic time-varying delay (TVD) and packet loss is constructed using a time-delay network simulation algorithm. A cascaded LSTM-Transformer prediction architecture is proposed for accurate prediction of both time delay and contact force. Experimental validation on a teleoperation platform comprising an Omega7 haptic device and a Barrett WAM robotic arm demonstrates that the proposed method outperforms other comparative approaches in terms of Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Error Percentage (MAEP) for both single-feature and multi-feature prediction tasks. In the final stage, the proposed architecture is utilized for network delay prediction. When the predicted delay exceeds a predefined threshold, the same model is further used to perform predictive compensation for the force feedback information. This strategy not only improves the precision of force feedback but also effectively mitigates packet loss caused by network fluctuations, providing a delay-resilient force feedback solution for high-transparency teleoperation systems.