publications

2025

1. RoboSoft

   Geometric-Aware Attention Networks with Adaptive Constraint Adjustment for Static Modeling of Tendon-driven Continuum Robots

   Xin Xu, Yang Song, Bowen Su, Ke Wu, Di Wu, Wendi Liang, and 2 more authors

   In Proceedings of the 8th IEEE-RAS International Conference on Soft Robotics, 2025

   Spotlight Paper Abs PDF

   This paper is selected as spotlight paper

   We propose a geometry-aware attention network (GAANet) for end-to-end static modeling of tendon-driven continuum robots with hyper-redundant DOFs, integrating domain-specific geometric constraints and demonstrably outperforming a multilayer perceptron (MLP) and a pseudo rigid body model (PRBM) in terms of tip position accuracy and real-time feasibility.

2. CRAS

   Cycle-Consistent Neural Networks for High-Precision Force-Position Mapping in Tendon-driven Surgical Soft Robot

   Yang Song, Bowen Su, Wendi Liang, Di Wu, and Xin Xu

   In Proceedings of the 14th Conference on New Technologies for Computer/Robot Assisted Surgery, 2025

   Abs PDF

   Accurate mapping of force and position is crucial for the management of tendon-driven surgical soft robot. We introduce a neural network framework that maintains cycle consistency, which incorporates an encoder-decoder based on convolutional neural networks (CNN) for translating force to position (measured in Newtons and meters) and a model based on Kolmogorov-Arnold networks (KAN) for the inverse kinetostatic mapping from position to force. A combined training approach, which includes distinct training phases followed by joint fine-tuning with the incorporation of cycle consistency loss, guarantees that closed-loop consistency is preserved. When tested in a tendon-driven soft robotic data set, our model demonstrates exceptional accuracy without load (forward mean square error: 0.000148 m²; inverse mean squared error: 0.0098 N²) and maintains strong performance under load conditions (forward mean squared error: 0.000337 m²; inverse mean squared error: 0.0376 N²). The cycle consistency errors are minimal (0.000217 m² without load, 0.000347 m² with load), confirming the validity of the physical consistency. With prediction durations between 0.15 and 0.25 ms per sample, this framework supports real-time control, validated through closed-loop experimental results.