Ann N Y Acad Sci. 2026 Sep;1563(1):e70396. doi: 10.1111/nyas.70396.
ABSTRACT
In robot-assisted minimally invasive surgery, precise grasping force feedback is a key factor for enabling delicate manipulation and reducing tissue damage. However, due to confined workspaces and frequent visual occlusions, achieving force feedback via integrated sensors or image signals remains challenging. To obtain grasping-force feedback for single-port surgical robots, this study proposes a two-stage force estimation framework combining Bidirectional Long Short-Term Memory-improved Transformer (BiLSTM-iTransformer) with time-series features. First, the prediction task is separated into contact state classification and grasping force regression, which effectively reduces misclassification in noncontact states. Then, a dynamic transition window and multiorder differencing are introduced to further improve the prediction performance. Furthermore, the improved Transformer incorporates a dynamic relative position-encoding mechanism, dilated convolution layers, and a learnable scaling factor, while integrating BiLSTM to enhance temporal modeling capacity. The framework was validated on known and unknown material datasets. Results confirm robust performance in both classification and regression tasks. The proposed method achieved classification accuracies of 93.83% and 91.63% on the two datasets, respectively. For regression, it attained an RMSE of 0.64, outperforming comparative models. This study provides a feasible sensorless force feedback solution, offering potential value for enhancing surgical safety and precision.
PMID:42730721 | PMC:PMC13570598 | DOI:10.1111/nyas.70396