Simultaneous Estimation of Shape and Force along Highly Deformable Surgical Manipulators Using Sparse FBG Measurement

Yiang Lu, Bin Li, Wei Chen, Junyan Yan, Shing Shin Cheng, Jiangliu Wang, Jianshu Zhou, Qi Dou, Yun-Hui Liu

Abstract

This paper presents a machine learning framework for the simultaneous estimation of shape and external contact forces in highly flexible surgical robots using a single-core Fiber Bragg Grating (FBG) sensor. By helically wrapping the sensor around a soft sensing tube and employing a Conv1D neural network to process sparse strain data, the system achieves high accuracy in predicting deformations and contact parameters. This approach eliminates the need for expensive multi-core fibers and complex physical calibrations, offering a robust solution for real-time feedback in minimally invasive surgery.

Introduction

Flexible surgical robots, such as ureteroscopes, must navigate tortuous paths inside the human body, often bending up to 270 degrees. For safe and effective control, surgeons require real-time feedback on both the robot's exact shape and any external contact forces occurring along its length. Traditional Fiber Bragg Grating (FBG) sensing methods rely on model-based approaches that require ultra-precise physical alignment and exact mathematical models, both of which degrade under the large deformations typical of soft robotics. Furthermore, determining the specific location of a force along the manipulator—rather than just at the tip—is a persistent challenge. While multi-core FBG fibers can address some of these issues, their cost and thickness make them difficult to integrate into the narrow working channels of surgical endoscopes.

Methods and Architecture

The researchers propose a data-driven approach that shifts the burden from mechanical precision to machine learning. The hardware consists of a thin-walled, highly deformable silicone "soft sensing tube" featuring a helical groove. A standard, low-cost single-core FBG fiber containing 14 sensing nodes is embedded into this groove, allowing the ureteroscope to slide inside.

Hardware design showing the soft sensing tube and helical FBG fiber integration.

Instead of using rigid kinematic equations, the system treats shape and force estimation as a supervised learning problem. The core of the framework is a neural network that processes normalized strain measurements from the FBG nodes. The team evaluated three spatial encoders: Fully Connected (FC), LSTM (processing the spatial sequence of nodes), and Conv1D. The Conv1D encoder proved most effective, as it utilizes a sliding window to capture local spatial relationships and continuity between neighboring FBG nodes.

The machine learning architecture featuring Conv1D, LSTM, and MLP encoders feeding into specialized decoder heads.

The encoded features are passed to separate decoder "heads": a Shape Head (predicting curvature and twist) and a Force Head. To improve learning stability, discrete point-contact forces were converted into a continuous Gaussian-like spatial distribution along the robot's length, making the contact location easier for the network to identify accurately.

Key Results

The proposed Conv1D model demonstrated excellent performance, significantly outperforming traditional model-based baselines and other network architectures:

  • Static Accuracy: The model achieved a tip position error of just 0.93 mm (0.78% of the total length) and a force location error of 2.35 mm.
  • Dynamic Robustness: During dynamic tracking involving 270-degree bends, the Conv1D model maintained a 1.05 mm tip error and a 3.79 mm force location error.
  • Baseline Comparison: Under the same dynamic conditions, traditional model-based approaches suffered a massive 5.00 mm tip position error, highlighting the superiority of the data-driven method for large deformations.

By bypassing the need for perfect FBG placement, this learning-based paradigm simplifies the manufacturing process. The network inherently learns to map strain patterns to the correct shape and force even if the fiber alignment is not mathematically perfect.

Conclusion and Future Work

This work demonstrates that sparse, single-core FBG measurements combined with deep learning can provide high-fidelity shape and force sensing for soft surgical tools. While currently validated for 2D in-plane bending, future research will expand the system to handle full 3D spatial deformations. The authors also plan to integrate FBG strain data with robotic actuation inputs and computer vision to further enhance robustness against environmental variables like temperature and material hysteresis.

Summary updated

BibTeX
@inproceedings{lu2024simultaneous,
  author = {Yiang Lu and Bin Li and Wei Chen and Junyan Yan and Shing Shin Cheng and Jiangliu Wang and Jianshu Zhou and Qi Dou and Yun-Hui Liu},
  title = {{Simultaneous Estimation of Shape and Force along Highly Deformable Surgical Manipulators Using Sparse FBG Measurement}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2024},
  doi = {10.1109/icra57147.2024.10611403},
}
Figure