A VR Training System for Robot-Assisted Vascular Intervention with Dual-Mode Control and Model-Based Interaction Response

Abstract

Virtual reality training for robot-assisted vascular intervention requires accurate robotic control and effective feedback on guidewire–vessel interactions. This paper presents a VR training system combining asynchronous dual-mode command mapping with sensorless interaction estimation. The system supports velocity-controlled navigation and position-driven incremental manipulation. Guidewire deformation and a Hunt–Crossley-type contact response are combined and temporally filtered to estimate interaction intensity, while trajectory-based visualization supports post-procedure review. Tests confirmed distinct command-dispatch patterns across three velocity gears and the position mode. During hybrid-scene navigation, the estimated response remained low during straight insertion and increased as the guidewire bent within the vascular model. Trajectory analysis yielded a model-derived cumulative trauma work (CTW) value of 15.166 in normalized units and localized regions of elevated interaction. In a preliminary study with 15 novices, physical-phantom task success increased from 53.3% to 93.3% after VR training, while mean completion time decreased from 132.4 s to 96.7 s. These results demonstrate the system’s dual-mode operation, deformation-sensitive feedback, and potential to improve subsequent physical-phantom task performance.

Index Terms:

Vascular interventional surgery, virtual reality training, guidewire simulation, virtual interaction estimation, dual-mode control.

I Introduction

Vascular interventional surgery (VIS) enables minimally invasive treatment of cardiovascular and cerebrovascular diseases. Robot-assisted VIS (RA-VIS) can improve manipulation precision and reduce occupational radiation exposure [1, 2, 3]. However, navigating flexible guidewires through tortuous vessels remains demanding because operators must coordinate instrument motion while interpreting guidewire–vessel interactions. This creates a need for training systems that reproduce the control characteristics of robotic platforms and provide interpretable interaction feedback.

Previous studies have investigated robotic delivery [4], force- and torque-feedback interfaces [5], virtual-to-real autonomous navigation [6], curvature-aware delivery control [7], and catheter interaction simulation [8]. Simulation and extended-reality environments have also been investigated for endovascular training and performance assessment [9, 10, 11, 12]. Although these studies have advanced individual aspects of robot-assisted intervention, few have established an integrated physical-leader–virtual-follower framework that connects robotic manipulation, virtual vascular simulation, interaction feedback, and post-procedure spatial review within one VR training workflow. In particular, reproducing the visual workflow alone does not ensure that inputs from a robotic leader-side device are consistently mapped to virtual guidewire motion or that trainees can interpret the effects of their manipulation.

Two technical gaps remain. First, operator input, command dispatch, physics simulation, and visual rendering may operate at different update rates. Feedback-control and haptically enhanced interfaces have been studied for intervention teleoperation [13, 14], but direct event-by-event forwarding in a hardware-connected simulator may produce uneven actuation under asynchronous execution. This issue becomes more complex when continuous velocity commands and discrete position increments coexist within the same procedure.

Second, interaction feedback in vascular intervention training is difficult to obtain without dedicated force sensors. Guidewire deformation and tip–vessel proximity provide complementary cues, but both are affected by numerical noise, sliding motion, and intermittent contact. Existing deformable-body simulation, catheter interaction modeling, and surgical haptic-feedback methods [15, 16, 8, 17] provide useful foundations, but they do not directly provide a unified sensorless interaction representation for training. Moreover, instantaneous feedback alone does not record where interaction occurred along the trajectory or quantify its accumulated burden for post-training review. Mixed-reality visualization studies [18] motivate such spatially linked feedback.

To address these gaps, this paper presents an integrated physical-leader–virtual-follower VR training framework for robot-assisted vascular intervention. The framework reproduces the manipulation workflow of the robotic platform, provides sensorless interaction feedback during guidewire navigation, and supports post-procedure spatial review. Within this framework, a multi-cue interaction representation method combines distributed guidewire deformation with local guidewire–vessel proximity to estimate relative interaction intensity without a force sensor. The main contributions of this paper are:

• An integrated physical-leader–virtual-follower VR training framework with asynchronous dual-mode command mapping. The framework connects physical manipulation, virtual instrument simulation, velocity-controlled navigation, position-driven incremental manipulation, and real-time interaction feedback within a unified training workflow.
• A sensorless multi-cue interaction representation method for vascular intervention VR training. The method combines distributed guidewire deformation with local guidewire–vessel proximity and a nonlinear virtual contact response to provide relative interaction-intensity feedback without requiring a force sensor.
• A trajectory-linked post-procedure analysis function that records estimated interaction intensity together with guidewire-tip positions and visualizes its spatial distribution along the vascular geometry.

II Teleoperation Platform

The physical robotic platform adopts a leader–follower teleoperation architecture. The leader-side console contains two horizontally arranged joystick interfaces for catheter and guidewire manipulation, respectively. These interfaces reproduce the relative arrangement of the instruments on the operating table and allow the operator to generate advancement, withdrawal, rotation, and combined translation–rotation commands. The console can be connected either to the physical follower robot for robotic intervention or to the computer running the proposed VR training system. The physical follower robot is therefore not required during VR training.

The physical follower robot uses four drive units to manipulate as many as three catheters and one guidewire. Axial instrument motion is produced by paired friction and pinch rollers. Catheter rotation is transmitted through a geared Y-valve mechanism, whereas guidewire translation and rotation are generated by dedicated roller assemblies. This arrangement supports coaxial catheter–guidewire manipulation and rapid instrument exchange. Although the simulation platform implements control switching between the catheter and guidewire, the interaction-estimation method and quantitative evaluation presented in this study focus primarily on guidewire–vessel interaction.

The leader-side control interfaces support velocity and position modes. In velocity mode, the instrument advances or withdraws at one of five predefined speed levels: 2, 5, 10, 30, and 60 mm/s for Gears 1–5, respectively. In position mode, the displacement of the control interface is reproduced as a corresponding displacement of the virtual instrument, enabling fine adjustment in tortuous branches, near bifurcations, and within delicate anatomical structures. These control characteristics form the basis of the velocity and position modes implemented in the VR training system.

The virtual follower side is implemented using the Simulation Open Framework Architecture (SOFA), an open-source framework for real-time interactive mechanical simulation with an emphasis on biomechanics and medical simulation, rather than the physical follower robot. The catheter and guidewire are represented as deformable virtual instruments within a surface-based vascular model. Commands acquired from the leader-side console are mapped to virtual advancement, withdrawal, and rotation. During simulation, SOFA updates the instrument configurations under vascular constraints and provides the guidewire centerline nodes and instrument-tip positions required for interaction estimation. In the present implementation, guidewire deformation and tip-to-wall proximity are used to estimate guidewire–vessel interaction, while catheter motion is retained as part of the virtual procedural workflow.

Operators must learn to coordinate control modes while interpreting guidewire bending, intermittent contact, and sliding.

An effective VR system should reproduce these interactions and the teleoperation control characteristics while maintaining stable command execution across asynchronous input, simulation, and rendering updates. The following framework addresses these requirements through dual-mode command mapping and model-based interaction estimation.

III Control and Virtual Interaction Estimation Framework

The proposed framework integrates dual-mode command mapping, SOFA-based virtual instrument execution, state acquisition, interaction estimation, and post-procedure analysis for VR-based vascular intervention training. As illustrated in Fig. 2, the operator manipulates the catheter or guidewire interface on the leader-side console. The acquired CAN messages specify the selected instrument, motion direction, control mode, and gear level. The asynchronous command-mapping module converts these inputs into executable virtual advancement, withdrawal, or rotation commands. SOFA executes these commands to update catheter and guidewire motion under vascular constraints and contact.

During virtual guidewire navigation, SOFA provides the guidewire centerline nodes and tip position at each simulation step. The centerline nodes are used to calculate the curvature-derived deformation component, while the tip position is used to calculate the proximity-based virtual contact component. The two components are subsequently fused and temporally filtered. The resulting interaction estimate is presented to the operator as a numerical value and a color-coded visual cue on the VR interface and is recorded together with the guidewire-tip position for post-procedure analysis.

III-A Asynchronous Dual-Mode Command Mapping

The leader-side console supports velocity and position control modes for reproducing the manipulation characteristics of the robotic platform. Velocity mode is used for continuous advancement and withdrawal, with five gear levels determining the command magnitude and dispatch interval. Position mode is used for fine manipulation in tortuous vessels and near bifurcations, where the physical displacement of the leader-side interface is converted into incremental motion of the virtual instrument.

Because leader-side input events and SOFA simulation updates are asynchronous, the command-mapping module processes the latest operator input before dispatching a virtual motion command. In velocity mode, commands are issued according to the selected direction, gear-dependent velocity, and dispatch interval. In position mode, small input fluctuations are suppressed by a displacement deadband, while the remaining displacement is accumulated and converted into bounded position increments. This module serves as the control interface between the physical leader console and the SOFA-based virtual follower.

III-B Curvature-Based Deformation Estimation

The guidewire deformation state is characterized using directional changes along its discretized centerline. Local bending angles provide a geometric cue for deformation-related interaction and can be computed from the node positions available at each simulation step.

III-B1 Guidewire Kinematic Representation

The guidewire centerline consists of ordered nodes \{\mathbf{p}_{i}\}_{i=1}^{N}, where \mathbf{p}_{i}\in\mathbb{R}^{3}. Nodes are ordered from the proximal end toward the distal tip. The segment vector and length are defined as

\displaystyle\mathbf{s}_{i} \\ \displaystyle= \\ \displaystyle\mathbf{p}_{i+1}-\mathbf{p}_{i}, \\ \displaystyle\ell_{i} \\ \displaystyle= \\ \displaystyle\|\mathbf{s}_{i}\|_{2},

where i=1,\ldots,N-1. A segment is valid when \ell_{i}>\varepsilon, with \varepsilon>0 denoting a numerical tolerance. In the implementation, \varepsilon=10^{-4} is used to exclude degenerate centerline segments. Its normalized direction is

\mathbf{v}_{i}=\frac{\mathbf{s}_{i}}{\ell_{i}}.

Only pairs of adjacent segments that are both valid are used to calculate bending angles. The same tolerance is used in all vascular environments because the guidewire discretization and coordinate scale remain consistent.

III-B2 Robust Curvature Estimation

For a valid adjacent pair, the bending angle at the shared node \mathbf{p}_{i+1} is

\theta_{i}=\arccos\!\left(\operatorname{clip}(\mathbf{v}_{i}^{\top}\mathbf{v}_{i+1},-1,1)\right).

Clipping prevents floating-point errors from placing the dot product outside the domain of the inverse cosine. An angular dead zone suppresses negligible bending and numerical fluctuations:

\tilde{\theta}_{i}=\max(0,\theta_{i}-\theta_{\mathrm{th}}),

where \theta_{\mathrm{th}}=0.08 rad suppresses discretization and simulation noise. This threshold is fixed across vascular environments.

Let M denote the number of valid bending angles, and reindex the retained angles as \tilde{\theta}_{j}, j=1,\ldots,M, in proximal-to-distal order. For M\geq 2, the normalized index and weight are

\displaystyle r_{j} \\ \displaystyle= \\ \displaystyle\frac{j-1}{M-1}, \\ \displaystyle w_{j} \\ \displaystyle= \\ \displaystyle r_{j}^{2}.

The weighting emphasizes distal bending in the interaction cue. The resulting deformation descriptor is

C_{\mathrm{def}}=\sum_{j=1}^{M}w_{j}\tilde{\theta}_{j}.

When M<2, the descriptor is set to zero because the normalized weighting is undefined or lacks a valid range.

The scalar C_{\mathrm{def}} summarizes weighted bending along the guidewire. It is a curvature-related geometric descriptor based on turning angles, rather than an arc-length-normalized estimate of local curvature or a direct measurement of contact force.

III-C Nonlinear Contact Response and Interaction Fusion

The deformation descriptor captures guidewire bending, while tip-to-wall proximity provides a complementary cue for local virtual contact. The framework combines these cues through a nonlinear contact-response model and applies scalar Kalman filtering to reduce temporal fluctuations associated with numerical noise, sliding motion, and intermittent contact. The curvature, proximity-response, and temporal-filtering stages serve distinct roles in the sequential estimation pipeline and are evaluated here as an integrated method.

III-C1 KD-Tree-Based Distance Query

Let the vascular surface mesh contain V vertices. These vertices are organized in a preconstructed KD-tree T_{\mathrm{vessel}} to support repeated nearest-neighbor queries. At simulation step k, the guidewire-tip position \mathbf{p}_{\mathrm{tip},k} is used to obtain

d_{k}=\mathcal{Q}_{\mathrm{KD}}\left(T_{\mathrm{vessel}},\mathbf{p}_{\mathrm{tip},k}\right).

Each vascular mesh uses the SOFA scene’s scale and orientation before KD-tree construction. The nearest-vertex distance is divided by 100 to obtain d_{k}, with d_{\mathrm{safe}}=0.5 on the same normalized scale.

For a different vascular environment, the tree is rebuilt from its mesh while normalization and threshold settings remain fixed. Anatomical differences enter through the queried distance sequence, which approximates tip-to-surface proximity using mesh vertices.

III-C2 Hunt-Crossley-Type Repulsive Response

A nonlinear virtual contact response incorporates both proximity and approach rate. Let d_{\mathrm{safe}} denote the virtual contact threshold. The effective virtual penetration is

\delta_{k}=\max(0,d_{\mathrm{safe}}-d_{k}).

This quantity represents intrusion into a prescribed proximity zone. When contact is active and a previous distance sample is available, the approach-rate term is approximated as

\dot{\delta}_{k}\approx\frac{d_{k-1}-d_{k}}{\Delta t},

where \Delta t>0 is the elapsed time between consecutive simulation samples. The rate term is set to zero when no previous sample is available.

The Hunt-Crossley-type response is

F_{\mathrm{hc},k}=\begin{cases}\delta_{k}^{n}(K+\lambda\dot{\delta}_{k}),&d_{k}<d_{\mathrm{safe}},\\
0,&\text{otherwise},\end{cases}

where the stiffness K=35, damping \lambda=2, and exponent n=1.5 control response magnitude, approach-rate dependence, and nonlinearity, respectively. These settings are fixed across vascular environments under the common distance normalization. The repulsive component is constrained to be nonnegative:

F_{\mathrm{rep},k}=\max(0,F_{\mathrm{hc},k}).

This constraint prevents an attractive response during separation. For a bounded approach rate, both response terms tend to zero as \delta_{k} approaches zero, supporting a gradual transition at the virtual contact boundary.

III-C3 Interaction Combination and Kalman Temporal Filtering

The descriptor in Section III-B is scaled to form the curvature-derived component:

F_{\mathrm{curv},k}=k_{c}C_{\mathrm{def},k},

where k_{c}\geq 0 is the deformation scaling coefficient. In the implementation, k_{c}=10 is used to scale the weighted curvature descriptor before it is combined with the virtual contact component. The two components are expressed on a common interaction-signal scale and summed as

z_{k}=F_{\mathrm{curv},k}+F_{\mathrm{rep},k}.

The resulting signal is a model-derived interaction estimate constructed from guidewire deformation and tip-to-wall proximity. It is subsequently filtered and reported in model-derived interaction units for relative feedback and comparison during VR training.

A scalar random-walk state model and observation model are used for temporal filtering:

\displaystyle x_{k} \\ \displaystyle= \\ \displaystyle x_{k-1}+\eta_{k}, \\ \displaystyle z_{k} \\ \displaystyle= \\ \displaystyle x_{k}+\nu_{k},

where x_{k} is the latent interaction state. The process noise \eta_{k}\sim\mathcal{N}(0,Q) and observation noise \nu_{k}\sim\mathcal{N}(0,R) are modeled as independent, with Q\geq 0 and R>0.

The recursive prediction and correction equations are

\displaystyle\hat{x}_{k|k-1} \\ \displaystyle= \\ \displaystyle\hat{x}_{k-1}, \\ \displaystyle P_{k|k-1} \\ \displaystyle= \\ \displaystyle P_{k-1}+Q, \\ \displaystyle K_{k} \\ \displaystyle= \\ \displaystyle\frac{P_{k|k-1}}{P_{k|k-1}+R}, \\ \displaystyle\hat{x}_{k} \\ \displaystyle= \\ \displaystyle\hat{x}_{k|k-1}+K_{k}\bigl(z_{k}-\hat{x}_{k|k-1}\bigr), \\ \displaystyle P_{k} \\ \displaystyle= \\ \displaystyle(1-K_{k})P_{k|k-1}.

Here, \hat{x}_{k|k-1} and P_{k|k-1} are the predicted state and covariance, K_{k} is the Kalman gain, and \hat{x}_{k} and P_{k} are the posterior estimate and covariance. The contact stiffness K and the time-indexed gain K_{k} denote different quantities.

The filter parameters are set to Q=0.05 and R=2.0, with the initial covariance set to P_{0}=1.0. These settings smooth frame-to-frame variations in the combined signal while preserving sustained changes caused by guidewire bending and wall approach.

The filter is initialized with \hat{x}_{0}=0. An initialization interval of T_{\mathrm{init}}=3.0 s is used. Whenever the simulation is paused or the accumulated simulation time T_{k} is below this interval, the filtered output is set to zero and the covariance is reset to P_{0}. After initialization, the filtered estimate \hat{x}_{k} is displayed to the operator as a numerical value and a color-coded visual cue and is recorded for post-procedure analysis. Algorithm 1 gives the complete estimation and filtering procedure, including these reset conditions.

III-D Spatiotemporal Interaction Analysis

Filtered estimates and guidewire-tip positions are recorded for two complementary analyses: spatial mapping of local interaction intensity and displacement-weighted accumulation as cumulative trauma work (CTW).

III-D1 Cumulative Trauma Work Formulation

Let \Delta s_{k} denote the incremental guidewire-tip displacement between consecutive simulation samples:

\Delta s_{k}=\left\|\mathbf{p}_{\mathrm{tip},k}-\mathbf{p}_{\mathrm{tip},k-1}\right\|_{2}.

Let \mathcal{I} contain the recorded step indices satisfying the minimum-displacement criterion. Cumulative trauma work is defined as

W_{\mathrm{CTW}}=\sum_{k\in\mathcal{I}}\hat{x}_{k}\Delta s_{k}.

This formulation incorporates interaction magnitude and traveled distance, providing a cumulative measure of interaction burden. Unlike an instantaneous peak or a collision count, it accounts for interaction along the recorded path. Because the metric is displacement-weighted, stationary contact does not contribute to the sum.

In this study, CTW is used as a model-derived surrogate metric for vascular injury risk by combining estimated interaction intensity with guidewire displacement. By characterizing the interaction accumulated along the guidewire trajectory, CTW provides a quantitative indicator for assessing the potential severity and risk of vascular injury. Under consistent signal-scaling and sampling conditions, higher CTW values indicate greater accumulated interaction burden and a correspondingly higher potential for vascular injury.

III-D2 Spatiotemporal Heatmap Visualization

The recorded tuples \{(\mathbf{p}_{\mathrm{tip},k},\hat{x}_{k})\}_{k\in\mathcal{I}} are displayed as a color-coded overlay on the vascular geometry. Each tip position is assigned a color according to its associated interaction estimate. Cooler colors indicate lower estimated interaction, and warmer colors indicate higher estimated interaction. A consistent color scale is used when comparing trials.

The overlay relates elevated interaction to bends, bifurcations, and narrow segments, complementing CTW for post-procedure review.

IV Experimental Evaluation

The experimental evaluation examined four aspects of the proposed training system: qualitative agreement between physical and virtual guidewire configurations, dual-mode command dispatch, the response of the curvature-derived interaction component to guidewire bending, and post-procedure visualization of recorded interactions. A preliminary user study additionally examined whether practice in the VR environment was associated with improved subsequent performance in the physical vascular phantom. Together, these experiments characterize both system behavior and the transfer of training to a physical task.

IV-A Experimental Setup

The experimental evaluation used virtual and physical configurations (Fig. 3). In the virtual configuration, the joystick serving as the leader-side control device was connected to a computer running the SOFA-based VR environment. Commands from the joystick controlled virtual guidewire and catheter advancement, withdrawal, and rotation. In the physical configuration, the same joystick was connected to the physical follower robot, which manipulated the guidewire and catheter within the transparent vascular phantom. The physical follower robot was used in the phantom-based experiments but was not part of the VR training loop.

The vascular model in the VR environment was paired with a corresponding transparent physical vascular phantom, providing matched vessel paths for qualitative comparison of guidewire configurations. The physical–virtual comparison used both configurations, whereas command-dispatch validation, deformation-related interaction evaluation, and trajectory-linked visualization were conducted in the VR environment. The preliminary user study used the physical robotic configuration for the pre-training and post-training assessments and the virtual configuration for the intermediate VR training session. The control pipeline and initialization conditions were kept consistent across repeated trials of each virtual navigation task.

IV-B Physical–Virtual Guidewire Configuration Comparison

The matched vascular geometries were used to examine whether the virtual guidewire reproduced the principal deformation pattern visible in the physical phantom. Fig. 4 presents representative configurations in the corresponding vessel segment.

In both images, the guidewire follows the lower curved vessel segment and bends as it approaches the branch region. The similar large-scale paths provide qualitative support for the guidewire deformation represented in the VR environment. This comparison concerns visible guidewire configuration; the interaction estimates evaluated below are model-derived signals calculated from virtual deformation and proximity cues.

IV-C Dual-mode Control Validation

The dual-mode controller was evaluated through an equal-distance traversal test using command-dispatch records from a representative session. The test examined the stepwise output characteristics of the velocity gears and position mode. Gear 1, Gear 3, Gear 5, and position mode each moved the guidewire over the same prescribed distance, and their activation and dispatch intervals were analyzed.

Fig. 5 shows distinct dispatch patterns and traversal durations for equal-distance movements. Lower gears were used early in the session, followed by higher gears and position increments, demonstrating mode changes within one workflow.

The record provides a functional demonstration that the different gear settings and position mode can be used to complete equal-distance guidewire movements within one training session.

IV-D Curvature-Based Interaction Response Evaluation

A hybrid simulation scene containing a straight glass tube and an anatomical vascular model was used to examine the response of the curvature-derived interaction component. The glass tube provided a low-bending reference segment, while the vascular model introduced curved paths that constrained guidewire deformation. This configuration allowed the guidewire to pass from a nearly straight configuration into progressively bent configurations within the same navigation sequence.

To examine the repeatability of the curvature-derived interaction response, the navigation experiment was repeated 20 times under the same simulation settings. Each recorded trajectory was divided into three representative stages: straight-tube insertion, entry into the curved vascular segment, and navigation within the deeper curved segment. For each trial and stage, the curvature-derived response F_{\mathrm{curv},k} was averaged over all simulation samples within that stage. The reported values are model-derived responses in arbitrary units, rather than measured physical forces or maximum response values.

During advancement through the straight glass tube, the guidewire centerline remained nearly collinear and the curvature-derived response stayed close to zero. After entering the vascular model, the guidewire bent to follow the lumen, and the response increased with the bending level (Fig. 6). Table I reports the mean and sample standard deviation of the stage-averaged responses across the 20 trials, demonstrating an increase from straight-tube insertion to curved-segment entry and deeper curved navigation.

IV-E Spatiotemporal Interaction Visualization

After each training session, recorded guidewire-tip positions and filtered interaction estimates were used to generate a spatial overlay on the vascular geometry. Samples were color-coded by interaction magnitude, while cumulative trauma work (CTW) summarized displacement-weighted interaction burden. The representative heatmap showed localized regions of elevated interaction separated by lower-interaction segments (Fig. 7), providing complementary spatial and cumulative information for post-training review.

The preceding experiments evaluate the functional behavior of the control, interaction-estimation, and post-procedure analysis modules. To examine the training objective of the complete system, a paired pre-training and post-training study was additionally designed to assess whether novice users transferred the practiced manipulation skills from the VR environment to the physical vascular phantom.

IV-F Preliminary User Training Evaluation

Fifteen participants without prior experience in robot-assisted vascular intervention completed a physical-phantom test before and after a 30-min VR training session. Before the initial test, participants received a standardized introduction to the leader-side console and task rules without a practice trial. During the physical tests, participants operated the leader-side joystick, and the follower robot manipulated the guidewire within the transparent vascular phantom. The initial guidewire position, target branch, and completion criterion were identical in the pre-training and post-training tests.

A trial was successful when the guidewire reached the target branch within 180 s without entering a non-target terminal branch. Completion time was measured from the first guidewire movement to target acquisition; unsuccessful trials were assigned the 180-s time limit. Incorrect branch entries and excessive wall-contact events, defined as visible guidewire buckling or sustained tip impingement, were independently assessed from recorded videos by two reviewers, with disagreements resolved by consensus. Completion time was analyzed using a paired t-test after confirming approximate normality of the within-participant differences. Paired count and binary outcomes were analyzed using Wilcoxon signed-rank and McNemar’s tests, respectively, with p<0.05 indicating statistical significance. Table II summarizes the results.

Task success increased from 53.3% to 93.3%, accompanied by a reduction in mean completion time. The reductions in navigation errors and excessive wall contacts further indicate improved post-training performance. Because the study used a single-group pre–post design, these findings should be interpreted as preliminary evidence rather than as a controlled comparison with conventional training.

V Conclusion

This paper presented a VR training framework for robot-assisted vascular intervention that integrates asynchronous dual-mode control, model-based interaction estimation, and post-procedure analysis. The experimental evaluation verified the operation of the dual-mode controller, qualitative agreement between physical and virtual guidewire configurations, the deformation-sensitive interaction response, and the trajectory-linked review functions.

In the preliminary study with 15 novice participants, physical-phantom task success increased from 53.3% before VR training to 93.3% after training, accompanied by a reduction in mean completion time. More importantly, the improved post-training performance in the physical robotic task suggests that the proposed system can support the transfer of navigation and manipulation skills practiced in VR to a corresponding physical intervention environment. The interaction feedback and trajectory-based review may further help trainees recognize potentially unsafe manipulation patterns rather than focusing solely on task completion.

These findings provide preliminary support for the proposed framework as a training and performance-review tool. Because the participant evaluation used a single-group pre–post design, however, the observed improvements cannot be attributed exclusively to VR training. Future work will validate the interaction estimates against measured forces and evaluate training effectiveness through randomized controlled studies with independent training and assessment tasks.

References

1. [1] W. Duan, T. Akinyemi, W. Du, J. Ma, X. Chen, F. Wang, O. Omisore, J. Luo, H. Wang, and L. Wang (2023) Technical and clinical progress on robot-assisted endovascular interventions: a review. Micromachines 14 (1), pp. 197.
2. [2] Y. Kim, E. Genevriere, P. Harker, J. Choe, M. Balicki, R. W. Regenhardt, J. E. Vranic, A. A. Dmytriw, A. B. Patel, and X. Zhao (2022) Telerobotic neurovascular interventions with magnetic manipulation. Science Robotics 7 (65), pp. eabg9907.
3. [3] G. Dagnino, D. Kundrat, T. M. Y. Kwok, M. E. M. K. Abdelaziz, W. Chi, A. Nguyen, C. Riga, and G. Yang (2023) In-vivo validation of a novel robotic platform for endovascular intervention. IEEE Transactions on Biomedical Engineering 70 (6), pp. 1786–1794.
4. [4] S. Cao, S. Guo, J. Guo, J. Wang, Y. Zhang, Y. Zhang, P. Yang, and J. Liu (2024) A reciprocating delivery device-based endovascular intervention robot with multimanipulators collaboration. IEEE Transactions on Instrumentation and Measurement 73, pp. 1–12. External Links: Document
5. [5] X. Bao, S. Guo, C. Yang, and L. Zheng (2024) Haptic interface with force and torque feedback for robot-assisted endovascular catheterization. IEEE/ASME Transactions on Mechatronics 29 (2), pp. 1111–1125. External Links: Document
6. [6] Z. Mei, J. Wei, S. Pan, H. Wang, D. Wu, Y. Zhao, G. Liu, and S. Guo (2025) Transferring virtual surgical skills to reality: AI agents mastering surgical decision-making in vascular interventional robotics. IEEE/ASME Transactions on Mechatronics 30 (2), pp. 1261–1272. External Links: Document
7. [7] S. Cao, S. Guo, J. Guo, J. Wang, C. Li, H. Xu, B. Wang, and M. Ding (2025) Design of an autonomous delivery system for vascular intervention robots based on curvature variation. In 2025 IEEE International Conference on Mechatronics and Automation (ICMA), pp. 104–109. External Links: Document
8. [8] P. Shi, S. Guo, X. Jin, H. Hirata, T. Tamiya, and M. Kawanishi (2023) A novel catheter interaction simulating method for virtual reality interventional training systems. Medical & Biological Engineering & Computing 61 (3), pp. 685–697.
9. [9] M. S. Schneider, K. O. Sandve, K. D. Kurz, I. Dalen, J. Ospel, M. Goyal, M. W. Kurz, and L. Fjetland (2023) Metric based virtual simulation training for endovascular thrombectomy improves interventional neuroradiologists’ simulator performance. Interventional Neuroradiology 29 (5), pp. 577–582.
10. [10] O. Sovik, A. Tveiten, H. Oygarden, P. J. Stokkeland, H. B. Hetland, M. S. Schneider, K. O. Sandve, M. Altmann, D. L. Hykkerud, J. Ospel, M. Goyal, H. L. Ersdal, M. W. Kurz, and P. K. Hyldmo (2026) Virtual reality simulation training in stroke thrombectomy centers with limited patient volume - simulator performance and patient outcome. Interventional Neuroradiology 32 (1), pp. 85–92.
11. [11] S. A. Patel, M. M. Covell, S. Patel, S. Kandregula, S. K. Palepu, A. A. Gajjar, O. Shekhtman, G. S. Sioutas, A. Dhanaliwala, T. Gade, J. Burkhardt, and V. M. Srinivasan (2024) Advancing endovascular neurosurgery training with extended reality: opportunities and obstacles for the next decade. Frontiers in Surgery 11 (), pp. 1440228.
12. [12] M. Pescio, C. Li, D. Kundrat, M. Casadio, and G. Dagnino (2025) A simulation environment for robot-assisted endovascular interventions. International Journal of Computer Assisted Radiology and Surgery 20 (11), pp. 2259–2267.
13. [13] J. Cheng, F. Abi-Farraj, F. Farshidian, and M. Hutter (2022) Haptic teleoperation of high-dimensional robotic systems using a feedback MPC framework. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Vol. , pp. 6197–6204.
14. [14] W. Duan, Z. Li, O. M. Omisore, W. Du, T. O. Akinyemi, X. Chen, X. Gao, H. Wang, and L. Wang (2025) Development of an intuitive interface with haptic enhancement for robot-assisted endovascular intervention. IEEE Transactions on Haptics 18 (1), pp. 80–92.
15. [15] W. Baumgartner and G. Kayar-Ceylan (2026) SPHaptics: a real-time bidirectional haptic interaction framework for coupled rigid-soft body and lagrangian fluid simulation in virtual environments. The Visual Computer 42 (7), pp. .
16. [16] G. Westergaard, M. Ellis, J. Barker, S. Garces Palacios, A. Desir, G. Sankaranarayanan, S. De, and D. Demirel (2026) Real-time haptic-based soft body suturing in virtual open surgery simulations. Computers & Graphics 134 (), pp. 104507.
17. [17] M. Bergholz, M. Ferle, and B. M. Weber (2023) The benefits of haptic feedback in robot assisted surgery and their moderators: a meta-analysis. Scientific Reports 13 (1), pp. 19215.
18. [18] V. Ruozzi, G. B. Regazzo, M. C. Palumbo, W. Beckers, M. Ourak, X. Zhang, F. Perico, A. Caimi, E. Vander Poorten, and E. Votta (2026) Catheter monitoring in intelligent endovascular navigation systems: interactive simulations and mixed reality for enhanced navigational awareness. International Journal of Computer Assisted Radiology and Surgery (), pp. .