Authors: Chen Qian, Sicong Gao, Laurence Xian, Zongyuan Ge, Song Yang and Leo Wu

Abstract: Autonomous navigation is crucial for medical endoscopy and healthcare robots operating in confined tubular environments. For follow-the-leader (FTL) continuum robots, this problem is especially important because the trajectory traced by the robot tip is propagated along the body. Existing depth-guided navigation strategies commonly use the deepest visible region as the navigation target, but this target may not coincide with the lumen center in curved or irregular environments. This paper summarizes a depth-estimation-based navigation strategy for FTL continuum robots. The method selects the nearest reliable depth contour, fits the local lumen cross-section, and uses the estimated center as a steering target before FTL propagation. We also adapt Depth Anything V2 for tubular/endoscopic scenes to improve depth consistency under weak texture and specular reflection. Simulations in S-shaped tube and colon models demonstrate improved centerline alignment compared with deepest-region navigation. Physical validation on the JammingSnake FTL continuum robot further shows reduced wall-contact forces in tube and colon-like environments. 

Authors: Laurence Xian, Yang Song, Shuai He, Wei Huang, Maurice Pagnucco and Liao Wu

Abstract: Knee Arthroscopy is a minimally invasive procedure that is vital in treating knee joint pathologies, but carries the risk of accidental contact with healthy tissue that can cause nerve damage or bacterial infection. Deep learning-based monocular depth estimation (MDE) has shown remarkable results for enhanced spatial understanding from single images. However, metric depth estimation is required for safe implementation, and this requires substantial labeled and relevant data for supervised training. To this end, we present the first synthetic dataset for knee arthroscopy that includes depth, segmentation, surface normal, and pose ground truth. To bridge the sim-to-real gap, this dataset was created using 3D models derived from MRI scans and then realistic textures and surgical tools were are created in Blender to represent all aspects of a knee arthroscopy. We fully finetuned the foundation MDE DepthAnythingV2 and performed qualitative evaluation on real arthroscopic footage. Future work will include quantitative evaluation using sparse 3D pointclouds obtained from phantom and porcine knees. 

Authors: Olatunji Omisore, Wenke Duan, Wenjing Du, Toluwanimi Akinyemi and Lei Wang

Abstract: We investigate the use of learning-based surgical data modeling for assessing the outcomes of recorded robot-assisted endovascular catheterization trials. First, a multimodal data fusion method was developed for offline feature extraction, followed by stacking-based deep neural network (SDNN) proposed for classifying trial outcomes been highly successful, slightly successful, or unsuccessful. Multimodal features were extracted from data recorded at both the operator’s and robot’s sides, then pre-processed to train the SDNN model. For model evaluation, sixty robot-assisted catheterization trials performed by nine operators in rabbit and pig animals were used for model training. 10-fold cross-validation was performed and the proposed SDNN’s performances were compared with those of existing machine and deep learning models, using mean accuracies from both data sets. Furthermore, the results of ablation studies performed showed the SDNN model could generalize well across data from both animals. We also showed that the model offers a higher performance when trained with features from multimodal data sources than using features from unimodal and bimodal sources. This study is intended for surgical training assistance than replacing experts. 

Authors: Yunong Li, Guoheng Ma, Bo Feng, Jiewen Lai and Hongliang Ren

Abstract: Haptic interfaces are important for teleoperation and for collecting multi-modal human demonstration data in surgical robotics, yet existing haptic interfaces are typically designed from a general free-space motion perspective. In many surgical tasks, however, operator motions are organized around more structured task motions and require specific degree of freedom (DOF), such as remote-center-of-motion (RCM) orientation, local plane translation, and tool-frame axial insertion. To explore a more task-oriented interaction paradigm, we present Ori-Interface, a compact origami-inspired haptic interface for task-oriented data collection. Instead of using a conventional 3 translational (T) DOF + 3 rotational (R) DOF architecture, Ori-Interface decomposes manipulation into three physical channels: planar 2T motion, spatial 3R RCM motion, and local 1T axial interaction. The core mechanism is a redundantly actuated origami parallel structure with asymmetric branch chains, designed to avoid the initial singularity while maintaining compactness and required DOF. A customized moving platform integrates an IMU, a micro encoder, and a linear resonant actuator to support orientation sensing and axial input. A preliminary prototype with a compact mechanical size of 60 mm $\times$ 55 mm $\times$ 45 mm was fabricated, and an initial control framework was developed. The initial test demonstrates basic planar translation and three-dimensional RCM motion and data collection. Ori-Interface will simplify the teleoperation and haptic data collection process in surgical robotics in the future. 

Authors: Kai Pan, Shoudong Huang, Shuai Zhang and Liang Zhao

Abstract: Reconstructing patient-specific three-dimensional (3D) knee anatomy from biplanar radiographs is highly under-constrained, since two projections provide only sparse geometric evidence of the underlying volume. Although additional views could reduce this ambiguity, acquiring extra radiographs is often undesirable due to radiation exposure and clinical workflow constraints. This motivates radiographic novel-view synthesis as a way to provide supplementary projection information without additional image acquisition. However, when synthesised views are used for downstream reconstruction, visual fidelity alone is insufficient: pose-inconsistent or anatomically unreliable views may introduce incorrect geometric constraints and degrade the reconstructed volume. In this work, we investigate whether pose-consistent generated radiographs can improve limited-view knee reconstruction. We propose a two-stage framework that first uses Pose-consistent Diffusion (PDiff) to synthesise target-view DRRs from biplanar inputs, and then incorporates the original and synthesised views into Template-assisted Neural Attenuation Fields (TNAF) for voxel reconstruction. Given two fixed biplanar DRRs acquired in the sagittal and coronal plane, the proposed method synthesises target-view DRRs at specified poses and incorporates them as additional observations for template-assisted volumetric reconstruction. Experiments show that the proposed pose-consistent synthesis model improves novel-view quality over Zero-1-to-3 methods and provides more reliable constraints for downstream voxel reconstruction. These results suggest that pose consistency is critical when using generative views for quantitative 3D reconstruction. 

Authors: Haitao Gao, Yang Song and Liao Wu

Abstract: Autonomous minimally invasive surgery (MIS) with concentric tube robots (CTRs) requires fast, reliable inverse kinematics (IK), yet IK for CTRs is challenging due to nonlinear task to joint space mappings, multimodality from trigonometric parameterizations, and kinematic redundancy. We investigate how different task space representations affect learning-based IK for three-tube CTRs and propose a sparse SE(3) shape feature-enhanced diffusion model that characterizes the full distribution of IK solutions. We show that deterministic regression with recorded joint configurations is fundamentally incompatible with the multimodal and redundant nature of the IK problem. Our diffusion-based approach instead approximates the multimodal joint space distribution, enabling sampling of diverse, valid solutions. We further introduce distribution aware metrics quantifying task space accuracy, joint solutions diversity, physical plausibility, and best-of-K quality. On the benchmark dataset, our method gives state-of-the-art, a tip position error of 0.083 ± 0.141 mm, tip orientation error of 0.218◦ ± 0.285◦, backbone shape error of 0.122±0.147mm and orientation error of 0.173◦ ± 0.210◦. 

Authors: Xiyue Xiao, Chi Kit Ng, Xinxin Liu, Jian Zhu, Yang Yi, Long Bai, Guankun Wang, Huxin Gao, Ruizhou Zhao, Zhiwei Fang, Jinsong Lin, Zhiqing Tang, Hongliang Ren

Abstract: Healthcare robots increasingly rely on learned visual policies, language interfaces, and shared-autonomy loops, but a policy also needs to know when the current perceptual state is reliable enough for action. In endoscopic robotics, target loss, occlusion, blur, and tissue deformation can make a visually plausible frame unsafe for continued autonomy. We present Endo-DyLat++, a lightweight semantic readiness monitor for endoscopic vision-language-action robots. The monitor encourages a split between slowly varying anatomical/camera context and dynamic visual-risk evidence, predicts weak-proxy near-future target loss, degradation, and occlusion, and selects a readiness operating point under an unsafe-ready false-positive constraint. On a Kvasir-SEG real-image pseudo-video benchmark, Endo-DyLat++ is comparable to a strong frozen-DINOv2 single-latent baseline within seed variance, reduces variance on several aggregate metrics, and exposes a clear safety/coverage tradeoff: higher readiness recall and coverage come with more unsafe-ready acceptances. These results should not be interpreted as clinical validation or evidence of a deployed surgical robotic system; rather, they provide a compact testbed for converting endoscopic visual risk into an auditable pre-action readiness signal. 

Authors: Chun Fong Wong, Tang You Liu, Shing Wai Wong and Liao Wu

Abstract: Stereo endoscopy can improve depth perception and three-dimensional (3D) scene understanding in minimally invasive surgery (MIS), but conventional systems are constrained by trocar size, limiting stereo baselines to 5--12 mm. This restriction reduces depth estimation accuracy, while increasing baseline introduces occlusion, creating a fundamental design trade-off. This study presents a deployable stereo endoscopic system with an expandable baseline that passes through a standard 12 mm trocar and mechanically increases baseline after insertion. For 3D scene reconstruction, we integrate a pretrained FoundationStereo model for zero-shot disparity estimation without task-specific retraining. The system is evaluated on a heart phantom across baselines from 5--30 mm and working distances of 60--180 mm using root mean square error (RMSE) and an occlusion metric. Results show that increasing baseline improves reconstruction accuracy, reducing RMSE from 1.02 mm to 0.61 mm, while occlusion increases from 1.6\% to 8.9\%. A trade-off analysis identifies an optimal region around a 20 mm baseline, balancing accuracy and occlusion. These findings provide design guidelines for stereo endoscopic systems and demonstrate the feasibility of combining deployable stereo hardware with foundation-model-based depth estimation for enhanced 3D Surgical Visualization. 

Authors: Dominik Slomma, Shoudong Huang and Liang Zhao

Abstract: Robot-assisted minimally invasive surgery requires surface reconstruction methods that remain accurate under non-rigid tissue motion. A reconstructed surface alone, however, does not reveal how strongly local tissue regions deform relative to a reference state. This paper investigates triangle-wise affine deformation maps as a vision-based representation of local tissue state. Tissue-aware reconstruction refers here to deformation-aware reconstruction and does not imply direct force sensing, material identification, or tissue damage prediction. Unlike reconstruction-first pipelines, the proposed method estimates local deformation maps before global surface integration. These maps drive a closed-form linear global ray-constrained discrete differential geometry reconstruction and also provide scalar Green-Lagrange (GL) strain-magnitude indicators that summarise template-relative metric change. Controlled synthetic experiments with exact meshes and deformation maps separate global integration accuracy from map-estimation error. Simulated endoscopic, phantom, and in vivo experiments evaluate progressively more realistic conditions. The results show distinct template-relative deformation responses while maintaining competitive surface accuracy. The formulation therefore provides local deformation information together with the reconstructed geometry. 

Authors: Toluwanimi Akinyemi, Olatunji Omisore, Wenke Duan, Wenjing Du, Peiwu Qin, Wang Lei and Minxin Wei 

Abstract: In this study, we examine the viability of a geometry-aware representation for continuous control of guidewire navigation. While existing studies have mainly addressed discrete control with online reinforcement learning, we present that continuous control with offline reinforcement learning for endovascular navigation is feasible. The proposed representation integrates a path-distance guidance map as a spatial cue, a tangent-normal local action representation that reduces global action ambiguity, and route-aware reward shaping that limits deviation from the intended navigation route. Sixty navigation trials were collected across two distinct targets in a vascular phantom using the leader-follower endovascular robot. The pre-processed data were used to train two state-of-the-art offline reinforcement learning (ORL) models, and the in silico results show that the proposed representation improved the ORL models’ ability to learn continuous control policies for autonomous guidewire navigation from a static dataset. Across methods, success rates of 90.00±2.89 for navigation to the right external iliac artery and 96.00±0.02 for the right internal iliac artery were observed. Furthermore, ablation results indicate that the ORL model used the geometric representation during inference, while the performance of these models degrades extensively without the geometry-aware representation. 

Authors: Guoheng Sun, Shuhua Peng and Liao Wu

Abstract: Robotic palpation often converts each indentation into an apparent-stiffness map, reducing a force–depth response to a single slope. This convenient scalarization hides loading history and makes the representation depend on the selected indentation depths. We ask whether complete displacement and Fz trajectories provide a more faithful input for single-probe deep-lesion localization. Using a controlled soft-body simulator, we generated soft phantoms with paired curve scans, stiffness maps, and projected inclusion masks, and compared matched U-Net models under fixed and variable indentation depths. Preserving the curve increased held-out Dice from 0.792 to 0.819 under fixed depths and from 0.573 to 0.712 when terminal depth varied across scan locations. Together, these results identify scalar stiffness as both an information bottleneck and a depth-pair-dependent measurement. These findings support process-level force sensing for simulated single-probe localization, while physical validation remains future work.