Progressive Experience Fusion for Multi-Task World Model Control in Endovascular Navigation

Harry Robertshaw; Maxence Boels; Nikola Fischer; Sebastien Ourselin; Christos Bergeles; Alejandro Granados; Thomas C Booth

cs.ROAug 19, 2026

Progressive Experience Fusion for Multi-Task World Model Control in Endovascular Navigation

Authors: Harry Robertshaw, Maxence Boels, Nikola Fischer, Sebastien Ourselin, Christos Bergeles, Alejandro Granados, Thomas C Booth

Organizations: Surgical & Interventional Engineering, School of Biomedical Engineering & Imaging Sciences, King’s College London, UK · Department of Neuroradiology, King’s College Hospital, UK

Abstract

Autonomous endovascular navigation could support the delivery of mechanical thrombectomy to underserved areas, but controllers must navigate long, multi-stage paths across varying vascular anatomies. This study investigates Progressive Experience Fusion (PEF) to train a multi-task TD-MPC2 controller. We additionally evaluate a heuristic that changes the Model Predictive Path Integral planning horizon using residual action-sequence dispersion, and fine-tuning in a patient-specific simulation. Across five subtasks in ten known training anatomies with held-out targets, PEF achieved a mean success rate of 74%, compared with 37% for Soft Actor-Critic (p < 0.001) and 65% for base TD-MPC2 (p = 0.053). A PEF controller with adaptive-horizon planning trained on 30 vasculatures achieved a mean success rate of 90% in ten held-out vasculatures. The PEF agent successfully transferred to an unseen in vitro stroke patient vasculature under fluoroscopy, achieving a mean path ratio improvement from 63% to 80% with fine-tuning (p < 0.001), following 40x103 fine-tuning steps (corresponding to approximately 107 min of clinical inter-hospital transfer time). This work represents a proof of concept for multi-vasculature training and patient-specific adaptation, while further validation is required before clinical deployment.

Figures & tables

Fig. 1 : Study overview. Progressive Experience Fusion (PEF) organizes replay data in two stages: experience is first grouped across vascular geometries within each navigation task and is then grouped across tasks to train one multi-task TD-MPC2 controller. At run time, TD-MPC2 evaluates candidate action sequences in its latent dynamics model using Model Predictive Path Integral (MPPI) planning. The adaptive horizon heuristic increases the planning horizon when the final MPPI action-sequence proposal remains dispersed. The generalist agent is fine-tuned in silico using time-limited patient-specific interaction prior to intervention, enabling adaptation to previously unseen vascular anatomies without modifying hyperparameters.

Fig. 2 : Navigation tasks investigated. A target was randomly sampled from the set of points within the task finish area, while a guide catheter and guidewire navigated to it from a randomly sampled point in the task start area. DA: descending aorta, LCCA: left common carotid artery, LICA: left internal carotid artery, RCCA: right common carotid artery, RICA: right internal carotid artery.

Fig. 3 : In vitro 3D vascular phantom with labeled anatomy.

Fig. 4 : In vitro testbed showing the 3D vascular phantom, mobile C-arm, and robotic manipulator.

Fig. 5 : Illustration of the Progressive Experience Fusion (PEF) process. Each agent at layer $l>0$ is initialized from the fused post-training replay buffers of its assigned agents at layer $l-1$ and is subsequently trained across the combined configuration set. Its updated post-training replay buffer is then passed to the next fusion layer, culminating in the generalist policy $\pi^{(L)}$ .

Fig. 6 : Phantom vessel model under fluoroscopy, showing task start area and target locations for $A_{2L}$ - lower branch and $A_{2L}$ - upper branch . Expert neurointerventionalists typically use an access catheter with an extreme shape (Simmons catheter) for this maneuver – this task can be considered a superhuman experiment [ 33 ] .

Task	SAC			Base TD-MPC2			PEF
Task	SR (%)	PT (s)	PR (%)	SR (%)	PT (s)	PR (%)	SR (%)	PT (s)	PR (%)
$A_{1}$	96 $\pm$ 20	8.9 $\pm$ 1.6	89 $\pm$ 16	96 $\pm$ 20	16.7 $\pm$ 4.1	90 $\pm$ 3	100 $\pm$ 0	12.3 $\pm$ 2.1	$-$
$A_{2L}$	$10\pm 30$	14.5 $\pm$ 7.1	$27\pm 30$	22 $\pm$ 41	$17.0\pm 6.1$	45 $\pm$ 33	8 $\pm$ 27	14.7 $\pm$ 6.0	75 $\pm$ 8
$A_{2R}$	$8\pm 27$	12.8 $\pm$ 9.9	$29\pm 27$	66 $\pm$ 47	$16.5\pm 6.2$	77 $\pm$ 32	72 $\pm$ 45	28.6 $\pm$ 15.8	79 $\pm$ 29
$A_{3L}$	$16\pm 37$	7.2 $\pm$ 7.0	$26\pm 34$	50 $\pm$ 50	$13.7\pm 6.3$	66 $\pm$ 38	94 $\pm$ 24	13.6 $\pm$ 8.6	92 $\pm$ 18
$A_{3R}$	$56\pm 50$	5.6 $\pm$ 4.5	$63\pm 43$	90 $\pm$ 30	$12.9\pm 5.8$	90 $\pm$ 21	98 $\pm$ 14	42.7 $\pm$ 10.4	94 $\pm$ 2

TABLE I : In silico results per navigation task for a dataset of 10 real patient vasculatures, with mean results across all tasks and vasculatures. SAC and base TD-MPC2 results taken from [ 33 ] . Bold indicates the best value for each task (highest SR and PR; lowest PT).

Task	SAC			Base TD-MPC2			PEF
Task	SR (%)	PT (s)	PR (%)	SR (%)	PT (s)	PR (%)	SR (%)	PT (s)	PR (%)
$A_{1}$	100 $\pm$ 0	46.2 $\pm$ 3.6	$-$	100 $\pm$ 0	$80.4\pm 19.7$	$-$	100 $\pm$ 0	80.6 $\pm$ 10.1	$-$
$A_{2L}$ (lower)	0 $\pm$ 0	$-$	$70\pm 20$	40 $\pm$ 55	175.5 $\pm$ 0.7	82 $\pm$ 7	60 $\pm$ 55	173.7 $\pm$ 81.1	82 $\pm$ 6
$A_{2L}$ (upper)	0 $\pm$ 0	$-$	$47\pm 11$	0 $\pm$ 0	$-$	71 $\pm$ 14	0 $\pm$ 0	$-$	72 $\pm$ 5
$A_{2R}$	$40\pm 55$	97.0 $\pm$ 1.4	$42\pm 4$	60 $\pm$ 55	$125.0\pm 21.0$	49 $\pm$ 7	80 $\pm$ 45	127 $\pm$ 51	66 $\pm$ 0
$A_{3L}$	100 $\pm$ 0	54.4 $\pm$ 0	$-$	100 $\pm$ 0	$126.4\pm 57.7$	$-$	100 $\pm$ 0	$109.0\pm 42.4$	$-$

TABLE II : In vitro results per navigation task for training on a dataset of 10 real patient vasculatures and testing on a single known vasculature, with mean results across tasks. SAC and base TD-MPC2 results taken from previous work, but updated for the splitting of the

A_{2L}

task [ 45 ] . Bold indicates the best value for each task (highest SR and PR; lowest PT), ‘

-

’ denotes cell is not applicable (PT is not recorded when SR is 0%, and PR is not recorded when SR is 100%).

Task	PEF			PEF + Adaptive Horizon
Task	SR (%)	PT (s)	PR (%)	SR (%)	PT (s)	PR (%)
$A_{1}$	100 $\pm$ 0	8.0 $\pm$ 0.6	$-$	100 $\pm$ 0	$10.0\pm 1.8$	$-$
$A_{2L}$	$52\pm 50$	$35.6\pm 16.0$	$69\pm 14$	60 $\pm$ 49	28.5 $\pm$ 17.5	79 $\pm$ 13
$A_{2R}$	96 $\pm$ 20	$26.5\pm 13.4$	77 $\pm$ 26	$92\pm 27$	24.6 $\pm$ 14.7	$72\pm 3$
$A_{3L}$	100 $\pm$ 0	19.5 $\pm$ 14.2	$-$	$98\pm 14$	$20.2\pm 14.0$	76 $\pm$ 0
$A_{3R}$	$96\pm 20$	$20.7\pm 16.6$	$62\pm 19$	100 $\pm$ 0	18.4 $\pm$ 15.8	$-$

TABLE III : In silico results per navigation task for 30 training and 10 unseen patient vasculatures. Bold indicates the best value for each task (highest SR and PR; lowest PT), ‘

-

’ denotes cell is not applicable (PT is not recorded when SR is 0%, and PR is not recorded when SR is 100%).

Fig. 7 : Plot of mean task success rate vs number of fine-tuning steps for a single in silico vasculature using PEF and an adaptive horizon. Training times in relation to key times are highlighted: $90$ $\mathrm{min}$ = $34\times 10^{3}$ steps, $190$ $\mathrm{min}$ = $70\times 10^{3}$ steps

Fig. 8 : Heatmap of individual task success rate vs number of fine-tuning steps for a single in silico vasculature using PEF and an adaptive horizon. $20\times 10^{3}$ fine-tuning steps $\approx 54$ $\mathrm{min}$ .

Task	PEF + Adaptive horizon			PEF + Adaptive horizon + Fine-tuning
Task	SR (%)	PT (s)	PR (%)	SR (%)	PT (s)	PR (%)
$A_{1}$	100 $\pm$ 0	53.8 $\pm$ 4.8	$-$	100 $\pm$ 0	$61.0\pm 18.2$	$-$
$A_{2L}$ ( lower )	$20\pm 45$	$157\pm 0$	$77\pm 12$	40 $\pm$ 55	136.5 $\pm$ 40.3	91 $\pm$ 2
$A_{2L}$ ( upper )	$0\pm 0$	$-$	$67\pm 10$	$0\pm 0$	$-$	75 $\pm$ 3
$A_{2R}$	60 $\pm$ 55	167.7 $\pm$ 64.4	$73\pm 3$	60 $\pm$ 55	$175.7\pm 92.6$	74 $\pm$ 4
$A_{3L}$	$80\pm 45$	$81.3\pm 20.0$	$37\pm 0$	100 $\pm$ 0	57.0 $\pm$ 9.2	$-$

TABLE IV : In vitro results per navigation task for 30 training and one unseen patient vasculature. Bold indicates the best value for each task (highest SR and PR; lowest PT), ‘

-

’ denotes cell is not applicable (PT is not recorded when SR is 0%, and PR is not recorded when SR is 100%).

Fig. 9 : Vessel model under fluoroscopy, showing (a) successful navigation of $A_{2L}$ - lower branch , and (b) unsuccessful navigation of $A_{2L}$ - lower branch due to catheterization of the BCA (Brachiocephalic Artery).

Fig. 10 : Vessel model under fluoroscopy, showing (a) successful navigation of $A_{2R}$ , and (b) unsuccessful navigation of $A_{2R}$ due to catheterization of the RSA (Right Subclavian Artery).

Fig. 11 : Patient timeline from stroke onset to reperfusion, showing potential benefit of patient-specific fine-tuning for local admission at an MT-capable center, and outside hospital transfer to an MT-capable center. Data from [ 34 ] .

Robotic Endovascular Intervention

Robotic endovascular intervention uses robotic systems to navigate vascular instruments and perform procedures with greater precision, control, or autonomy.

See all 24 papers

Toward Safe Autonomous Robotic Endovascular Interventions using World Models
Apr 22, 2026Harry Robertshaw, Nikola Fischer, Han-Ru Wu +6World Models for Robotics Robotic Endovascular Intervention
Surgical & Interventional Engineering, School of Biomedical Engineering & Imaging Sciences, King’s College London, UK · Department of Radiology, National Taiwan University Hospital, Taiwan · Department of Neuroradiology, King’s College Hospital, UK
Toward AI Autonomous Navigation for Mechanical Thrombectomy using Hierarchical Modular Multi-agent Reinforcement Learning (HM-MARL)
Feb 20, 2026Harry Robertshaw, Nikola Fischer, Lennart Karstensen +6Robotic Endovascular Intervention Multi-Agent Reinforcement Learning
School of Biomedical Engineering & Imaging Sciences, Kings College London, London, UK · School of Engineering & Materials Science, Queen Mary University of London, London, UK
Vascular Geometry Characterization for AI-Based Endovascular Navigation
Jul 10, 2026Han-Ru Wu, Harry Robertshaw, Lisa Dwyer-Joyce +2Reinforcement Learning Medical Image Analysis
Department of Medical Imaging Science National Taiwan University Hospital Taiwan · Surgical & Interventional Engineering School of Biomedical Engineering & Imaging Sciences Kings College London UK · Chelsea & Westminster Hospital UK

Apr 22, 2026J/K move · Enter open · S save

Toward Safe Autonomous Robotic Endovascular Interventions using World Models

Harry Robertshaw, Nikola Fischer, Han-Ru Wu, Andrea Walker Perez+5

Surgical & Interventional Engineering, School of Biomedical Engineering & Imaging Sciences, King’s College London, UK · Department of Radiology, National Taiwan University Hospital, Taiwan · Department of Neuroradiology, King’s College Hospital, UK

World Models for Robotics Robotic Endovascular Intervention Surgical Robotics Model-Based RL

Open paper AskPDF arXiv

Autonomous mechanical thrombectomy (MT) presents substantial challenges due to highly variable vascular geometries and the requirements for accurate, real-time control. While reinforcement learning (RL) has emerged as a promising paradigm for the automation of endovascular navigation, existing approaches often show limited robustness when faced with diverse patient anatomies or extended navigation horizons. In this work, we investigate a world-model-based framework for autonomous endovascular navigation built on TD-MPC2, a model-based RL method that integrates planning and learned dynamics. We evaluate a TD-MPC2 agent trained on multiple navigation tasks across hold out patient-specific vasculatures and benchmark its performance against the state-of-the-art Soft Actor-Critic (SAC) algorithm agent. Both approaches are further validated in vitro using patient-specific vascular phantoms under fluoroscopic guidance. In simulation, TD-MPC2 demonstrates a significantly higher mean success rate than SAC (58% vs. 36%, p < 0.001), and mean tip contact forces of 0.15 N, well below the proposed 1.5 N vessel rupture threshold. Mean success rates for TD-MPC2 (68%) were comparable to SAC (60%) in vitro, but TD-MPC2 achieved superior path ratios (p = 0.017) at the cost of longer procedure times (p < 0.001). Together, these results provide the first demonstration of autonomous MT navigation validated across both hold out in silico data and fluoroscopy-guided in vitro experiments, highlighting the promise of world models for safe and generalizable AI-assisted endovascular interventions.

Abstract

Robotic Endovascular Intervention

Explore similar work

Toward Safe Autonomous Robotic Endovascular Interventions using World Models

Toward AI Autonomous Navigation for Mechanical Thrombectomy using Hierarchical Modular Multi-agent Reinforcement Learning (HM-MARL)

Vascular Geometry Characterization for AI-Based Endovascular Navigation