Library
Simulation to robot
26 studies in which a controller or model trained in simulation ran on a physical catheter or guidewire robot, with success in simulation and on the robot as each paper reports it. 19 are linked to a robot record.
| Year | Study | Robot | Simulator | Trained | Real setup | Success in sim | Success on robot | Code |
|---|---|---|---|---|---|---|---|---|
| 2026 | alabay2026_vista arXiv:2608.30220 |
magnetically actuated untethered microrobot (not a catheter robot) | 3D digital twin of vessels (not a physics policy simulator) | none (rule-based milestones) | vascular phantoms with flow + live rat IVC | n/a | navigation time -62%, corrective commands -98%, radiation -57% vs human-in-the-loop | N.R. |
| 2026 | Learning Expert Strategy for Autonomous Robotic Endovascular Intervention via Decoupled Procedural Execution arXiv:2607.00066 (accepted IROS 2026) |
ALLVAS | Sim4EndoR (SOFA) as digital twin of the phantom | policy (SAC high-level) + online nonlinear MPC filter (not learned) | heated silicone phantom with circulation, coronary and transfemoral routes, 20 trials | 98%, 96%, 98% (50 episodes per target); SAC alone 92/94/90% | 18/20 (90.0%), 4.2 +/- 0.8 min per trial | no |
| 2026 | EVEGARS: End-to-End Vision-Based Endovascular Guidewire Autonomous Reinforcement Learning in Large-Scale Simulations from Clinical Dataset PRCV (LNCS) |
guidewire robot under X-ray with "ex-vivo model" and remote station… | EVEGARS: pymunk 2D rigid-body, 120 ARCADE angiogram masks | policy (PPO, ResNet-50 + ViT encoder) from images | 4 delivery tasks in an undescribed "ex-vivo model", trials per task N.R. | 91% (denominator N.R.) | 85%, 80%, 80%, 70% (human 100%, 95%, 100%, 95%; behaviour cloning 0%) | no |
| 2026 | Progressive Experience Fusion for Multi-Task World Model Control in Endovascular Navigation arXiv:2608.18647 |
KCL lead-screw | stEVE / SOFA; 30 training + 10 held-out CTAs | policy (multi-task, progressive experience fusion), fine-tuned 40k steps in a simulation… | unseen printed phantom, 5 runs per sub-task | 90% on 10 unseen vasculatures | 60% without, 67% with fine-tuning (p = 0.623) | no |
| 2026 | Toward Safe Autonomous Robotic Endovascular Interventions using World Models arXiv:2604.20151 (submitted IROS 2026) |
KCL lead-screw | stEVE / SOFA v23.12; 0.0441 in catheter + 0.035 in guidewire; 15 CTAs (5 held out) | policy (TD-MPC2 world model vs SAC) | printed phantom of one training vessel under mobile C-arm; 5 runs per task, n = 25 | hold-out 58% (n = 250) TD-MPC2 vs 36% SAC | 68% vs 60% (p = 0.567); left common carotid 0/5 both | no |
| 2026 | Toward AI Autonomous Navigation for Mechanical Thrombectomy Using Hierarchical Modular Multi-Agent Reinforcement Learning (HM-MARL) IEEE RA-L 11(4) (arXiv:2602.18663) |
KCL lead-screw | stEVE / SOFA; Navien 0.058 in catheter + 0.035 in guidewire (E 47 and 43 MPa); 10 patient… | policy (hierarchical modular multi-agent SAC, HM-MARL) | 3D-printed transparent phantom (Clear V4) in glycerine; 10 bench episodes per task | A1,2 right 100%; A2,3 right 100%; A1,2 left 100%; A2,3 left 94% (50 episodes) | 100%, 70% (80% with 10-anatomy agent), 0%, 0% | no |
| 2026 | Vision-Language Procedural Reasoning for Context-Aware Reward Modeling of Robotic Endovascular Guidewire Navigation arXiv:2606.30698 (accepted IROS 2026) |
ALLVAS | SOFA digital twin of each phantom | policy (SAC) trained with reward weights set by a fine-tuned Qwen2.5-VL-3B stage… | three anthropomorphic phantoms (coronary, carotid, renal) with pump flow, 30 trials per… | N.R. | 100%, 97%, 100% with VLM-shaped reward; SAC alone 70%, 73%, 73%; expert manual 93%, 97%,… | no |
| 2026 | From Patient-Specific Digital Twin to Real-World Phantom: Autonomous Right Heart Catheterization IEEE Robotics & Automation Magazine (Mar 2026) |
custom 2-DoF Swan-Ganz catheter robot: active-passive roller set with… | SOFA + BeamAdapter patient-specific digital twin (phantom E 10 kPa, nu 0.49, density 1.2… | policy by supervised learning (CNN decision model, behaviour cloning of user key-press… | three water-filled patient-specific right-heart phantoms, static and pulsatile heartbeat,… | N.R. (twin used to generate data, no sim success rate) | >= 96% static, >= 94% dynamic; trajectory consistency >= 34.63% better than expert | no |
| 2025 | Learning-based autonomous navigation, benchmark environments and simulation framework for endovascular interventions Comput. Biol. Med. 196 (arXiv:2410.01956) |
Wire manipulator | stEVE / SOFA BeamAdapter; stEVE_bench BasicWireNav, ArchVariety, DualDeviceNav | policy (SAC + LSTM) | 3D-printed rigid aortic arch phantoms; camera bench 100+100 trials, fluoroscopy 10 trials | BasicWireNav 98/100; ArchVariety 90/100; DualDeviceNav 40/100 (sim only) | BasicWireNav camera 97/100 (9/10 on the first 10), fluoroscopy 10/10; ArchVariety camera… | yes: github.com/lkarstensen/stEVE,… |
| 2025 | RL-MG: Reinforcement Learning Framework for Magnetic Guidewire Path Following RCAR 2025 |
Magnet gantry | SOFA + BeamAdapter (Kirchhoff rod), dipole field from COMSOL; SolidWorks vessel models | policy (DDPG, 400 episodes) | femoral-artery model, 4 paths x 3 trials | N.R. (error 2.5/3.1/2.2/2.4 mm, time 7.0/13.0/12.0/9.0 s) | N.R. (error 3.6/3.8/3.1/2.8 mm, time 31.0/64.0/45.0/67.0 s) | |
| 2025 | FreezeSeg2RL: Frozen Segmentation Pretraining for Reinforcement Learning On Vascular Interventional Robot Autonomous Delivering PRICAI 2025 (Springer LNCS, print 2026) |
XMU wire robot | pymunk 2D X-ray simulation platform (same group as EVEGARS); ARCADE tasks + 4… | perception (frozen DeepLabV3+ segmentation encoder pretrained on 5,138 simulated images)… | 4 delivery tasks, 120 s limit, trials per task N.R. | 56.1% ARCADE tasks, 44.8% in-vitro-model tasks (simulated, Table 4) | 85%, 80%, 80%, 70% (human 100/95/100/95; BC 0%); without data augmentation 25%, 15%, 15%,… | no |
| 2025 | Transferring Virtual Surgical Skills to Reality: AI Agents Mastering Surgical Decision-Making in Vascular Interventional Robotics IEEE/ASME T-MECH 30(2) |
XMU wire robot | Pymunk 2D | policy (multiprocess PPO, three training stages) | transparent vessel model, 4 targets, 40 runs per target | 99.30, 99.26 and 98.34 % across three training stages (not the hardware targets) | 82.5 / 87.5 / 82.5 / 72.5 % (surgeon 100 %, replayed surgeon commands 0 %) | |
| 2025 | A simulation environment for robot-assisted endovascular interventions IJCARS |
CathBot | SOFA + BeamAdapter; Signorini contact with Coulomb friction 0.62; vessel FE | none | silicone aorta phantom with fiducials, bead registration | n/a | n/a | no |
| 2025 | Sim2Real Learning With Domain Randomization for Autonomous Guidewire Navigation in Robotic-Assisted Endovascular Procedures IEEE T-ASE 22 |
United Imaging | SOFA 23.06 + BeamAdapter; one 0.9 mm guidewire (density 1.55 g/cm3) | policy (PPO + GAE + domain randomisation; SAC, DDPG baselines) | simplified vascular phantom with physical twin; 4-5 trials per method; 2 human operators… | PPO 7/10 and 8/10; PPO+DR 8/10 and 9/10; PPO+GAE+DR 9/10 and 9/10 (targets A, B) | PPO 60% (5 trials); PPO+GAE+DR 100% (5); SAC and DDPG 75% (4); human 100% (3) in 35-39 s | no (supplement at DOI) |
| 2025 | Sim4EndoR: A Reinforcement Learning Centered Simulation Platform for Task Automation of Endovascular Robotics ICRA |
United Imaging | Sim4EndoR: SOFA v23.06 + BeamAdapter + Blender; phantom CT-scanned for a matching sim… | policy (DDPG, PPO, SAC; path-length reward) | coronary phantom, 2 end points, 5 trials each | Task A 5/5, Task B 4/5 (DDPG) | Task A 5/5, Task B 3/5 | no |
| 2025 | A Robust Deep Reinforcement Learning Framework for Image-Based Autonomous Guidewire Navigation ICRA |
KAIST wire bench | SOFA (new random left-coronary tree each training episode) | policy (PPO with local-map sub-goals along an A* path) | planar phantom, one bench, four paths, 40 runs per test | 86.9% (67 episodes x 3) vs 53.7% without vessel randomisation | 93.3%, 89.2%, 91.7% (three tests); Cho 2022 controller 83.3%, rule-based 19.2% in test I | no (video on YouTube) |
| 2024 | CathSim: An Open-Source Simulator for Endovascular Intervention IEEE T-MRB 6(3) |
CathBot | CathSim (MuJoCo; rigid-link catheter chain, rigid wall) | none for this comparison (RL baselines elsewhere in paper) | silicone Type-I arch phantom, CathBot force experiment | n/a | n/a | yes (github.com/airvlab/cathsim) |
| 2024 | FBG-driven simulation for virtual augmentation of fluoroscopic images during endovascular interventions Healthc. Technol. Lett. 11(6) |
N.R. (user study with FBG-guided catheter on rigid 3D-printed aorta) | SOFA (Timoshenko beam, catheter + wire overlap) | none (FBG shape pulls simulation, gain 0.8) | rigid aorta phantom, 8 locations | n/a | n/a | no |
| 2024 | wen2024_guidewire_seg arXiv:2410.07460 |
none (perception only) | synthetic fluoroscopy from simulation | perception (SAM adapted, style transfer + self-training) | public cardiac dataset + in-house neurovascular dataset | n/a | N.R. here | announced |
| 2022 | Sim-to-Real Transfer of Image-Based Autonomous Guidewire Navigation Trained by Deep Deterministic Policy Gradient with Behavior Cloning for Fast Learning IROS |
KAIST wire bench | SOFA (virtual planar vessel map) | policy (DDPG after behaviour cloning from a rule-based expert) | planar phantom of a different shape, camera 30 Hz, 4 targets x 10 episodes | 100/100, 41.08 s (expert 66.17 s) | 40/40, 35.06 s (expert 63.2 s) | no (video only) |
| 2022 | A Simulation Framework for Magnetic Continuum Robots IEEE Robot. Autom. Lett. 7(3) |
magnetic continuum robot, 1.5 mm (4.5 Fr) OD, three NdFeB ring… | SOFA + BeamAdapter (1D Kirchhoff-rod FEM, Lagrange-multiplier contact, Coulomb friction… | None | planar, 4 mm high, rigid POM and PMMA models (calibration, vasculature-like, free-form);… | n/a | mean tip position error 5.1 to 6.8 mm and orientation error 3.8 to 7.2 deg over the three… | yes… |
| 2022 | Learning-based autonomous vascular guidewire navigation without human demonstration in the venous system of a porcine liver IJCARS 17(11) |
Wire manipulator | SOFA + BeamAdapter (rigid wall, empty lumen; 0.035 in Terumo Radifocus angled stiff… | policy (deep RL, controller with recurrent layers; algorithm not re-checked); perception… | ex vivo porcine liver (one organ; CT mesh from the same liver) with pump flow; 10 targets… | 150/150 (100%) at training end | 9/30 (30%): wrong branch 10, entanglement 11 | no |
| 2022 | Robotic Actuation and Control of a Catheter for Structural Intervention Cardiology IROS 2022 |
MitraClip stabiliser | Unity scene of the CT-reconstructed heart; Cosserat-rod kinematic model of the catheter | policy (BC + GAIL path planner) | patient-specific silicone phantom | N.R. | N.R. (position error 1.12 / 1.09 / 1.66 mm X/Y/Z, max 2.64 mm) | |
| 2020 | Autonomous guidewire navigation in a two dimensional vascular phantom Curr. Dir. Biomed. Eng. 6(1), 20200007 |
Wire manipulator | SOFA + BeamAdapter; phantom walls rigid, lumen empty; wall friction and guidewire… | policy (DDPG + HER) | rigid 2D phantom, stereolithography print, filled with glycerine-water; a loop that… | maximum 96% after 52,500 training episodes (100 test episodes every 500 training… | 'the same results as navigating in the simulation environment' (no number, episodes… | no |
| 2019 | Deep Reinforcement Learning for the Navigation of Neurovascular Catheters Curr. Dir. Biomed. Eng. 5(1) |
Wire manipulator | SOFA + BeamAdapter; rigid vessel walls, wall-guidewire friction 0.1, empty lumen (no… | policy (DDPG, with HER and human demonstrations as options, OpenAI Baselines; DQN with… | idealised PMMA phantom, constant 10 mm vessel diameter, a bifurcation followed by a bi-… | DDPG + HER 89.6% (sparse reward), DQN + Rainbow 86.5% (dense reward), average success… | 7/10 (the trained DDPG model on the test bench) (Sec. 3) | no |
| 2019 | Automatic control of cardiac ablation catheter with deep reinforcement learning method J. Mech. Sci. Technol. 33(11) |
tendon-driven steerable EP-catheter robot, 3 motors (translation,… | Unity (catheter + heart collision model; rendered to angiography-like images) | policy (dueling DQN) | printed rigid right-atrium phantom, 100 episodes per model | 98%, 96%, 97%, 96% (four models) | 73% (no randomisation); 78%, 75% (one noise type); 87% (both) | no |
Data: api/sim2real.json.