A Residual Spatial Attention Graph Network for Sensorless Force Estimation in Teleoperated Robots

Precise force feedback is essential for teleoperated robots to perceive unknown environments and realize high-precision, safe motion control. However, harsh conditions and space limitations in demanding environments impede force sensor deployment, making sensorless force estimation imperative. This study proposes a residual spatial attention graph-temporal convolutional network (RSAG-TCN) that fuses spatial and temporal features for accurate sensorless force estimation in teleoperated robots. Comparative and ablation experiments were conducted in contact environments with varying stiffness. The comparative models include gated recurrent unit (GRU), long short-term memory (LSTM), PSTA-TCN, Densenet-Transformer, ASGRNN, and our prior MDELM. Experimental results showed that RSAG-TCN significantly outperformed all comparative models. Specifically, compared to the best-performing baseline (MDELM), RSAG-TCN reduced the root mean square error by 0.0236 N in soft contact environments and by 0.1040 N in hard contact environments. Ablation experiments further verified the necessity of the GCN and CAM-SAM modules for spatial feature extraction and key feature enhancement. Beyond offline accuracy evaluation, a human-in-the-loop experiment with six volunteers was conducted to demonstrate the proposed method's effectiveness in enhancing teleoperation transparency and stability in real-time operation. This work provides an efficient sensorless force estimation solution for teleoperation systems, offering technical support for high-precision robotic force control in extreme scenarios.