Deep-Learning-Based Control of a Decoupled Two-Segment Continuum Robot for Endoscopic Submucosal Dissection

TL;DR

This work addresses the dexterity gap in endoscopic submucosal dissection by introducing DESectBot, a dual-segment continuum robot with integrated forceps enabling 6-DoF tip control. A data-driven GRU-based inverse-kinematics controller is developed to handle strong nonlinear coupling between segments, and is benchmarked against Jacobian, MPC, FNN, and LSTM controllers. Across nested-rectangle and Lissajous trajectory tracking, as well as orientation tasks and a peg-transfer benchmark, the GRU consistently delivers superior precision, reliability, and speed, achieving a 100% success rate in a standardized task and an ex vivo ESD demonstration confirming practical operability. The results indicate that GRU-based control offers meaningful improvements in precision and usability for ESD training and potential clinical translation, with future work focusing on higher update rates, haptic feedback, self-calibration, and image-guided assistance.

Abstract

Manual endoscopic submucosal dissection (ESD) is technically demanding, and existing single-segment robotic tools offer limited dexterity. These limitations motivate the development of more advanced solutions. To address this, DESectBot, a novel dual segment continuum robot with a decoupled structure and integrated surgical forceps, enabling 6 degrees of freedom (DoFs) tip dexterity for improved lesion targeting in ESD, was developed in this work. Deep learning controllers based on gated recurrent units (GRUs) for simultaneous tip position and orientation control, effectively handling the nonlinear coupling between continuum segments, were proposed. The GRU controller was benchmarked against Jacobian based inverse kinematics, model predictive control (MPC), a feedforward neural network (FNN), and a long short-term memory (LSTM) network. In nested-rectangle and Lissajous trajectory tracking tasks, the GRU achieved the lowest position/orientation RMSEs: 1.11 mm/ 4.62° and 0.81 mm/ 2.59°, respectively. For orientation control at a fixed position (four target poses), the GRU attained a mean RMSE of 0.14 mm and 0.72°, outperforming all alternatives. In a peg transfer task, the GRU achieved a 100% success rate (120 success/120 attempts) with an average transfer time of 11.8s, the STD significantly outperforms novice-controlled systems. Additionally, an ex vivo ESD demonstration grasping, elevating, and resecting tissue as the scalpel completed the cut confirmed that DESectBot provides sufficient stiffness to divide thick gastric mucosa and an operative workspace adequate for large lesions.These results confirm that GRU-based control significantly enhances precision, reliability, and usability in ESD surgical training scenarios.

Figures (10)

• Figure 1: Decoupled Endoscopic Submucosal Dissection Robot (DESectBot): (a) detailed assembly drawing of DESectBot system, (b) detailed parts drawing of double active bending module with surgical gripper in kinematics coordinate system, (c) DESectBot actuation module, (d) demonstration of DESectBot in ESD treatment, (e) part name and corresponding index
• Figure 2: Internal GRU neural network structure for robot control: a 4 stacked layers GRU that is composed of multiple GRU cells. Five time steps are taken as the input epoch, and each time step state $S$ size is 12, which contains the target pose at the moment $t$ and the actual motor rotation counts from the previous moment $t-1$. In this work, each cell consists of 128 neurons, and the overall network is used to predict the motor rotation counts $(t)$.
• Figure 3: Training and validation loss curves of GRU and LSTM models as a function of training epochs.
• Figure 4: Experimental hardware setup for data acquisition, position, and orientation control. $Bottom$: (a) the DESectBot actuation module, (b) passive bending module with four support brackets to maintain its horizontal orientation, (c) double active bending module. $Top$: all parts are local close-up enlargements of the $Bottom$ section. (a) shows detailed structure of actuation module, (b) intercepts a section of the passive bending module and zooms in on the details, (c) shows the detailed structure of DABM and the white square object is the NDI$^\text{\textregistered}$ EM sensor equipment which its tracker tip is affixed to the central hole of the tool connection segment within the DABM.
• Figure 5: Trajectory experiment: (a) Three-dimensional visualization of the five trials for each controller motion trajectory with one ground truth nested rectangle trajectory, (b) Three-dimensional visualization of the five trials for each controller motion trajectory with one ground truth lissajous trajectory. The yellow designed dome-shaped volume indicates the training data collection workspace which has a dimension of $\phi$$60 \times 45$$mm$ (circular area × height), within which all training trajectories were sampled before. The red curve represents the experimental test trajectory, excluded from the training set and used solely for performance evaluation.