AI-Guided Autonomous Robotic Thrombectomy for Acute Ischemic Stroke

1. Problem Statement

Stroke is the fifth leading cause of death and the leading cause of long term disability in the United States. Approximately 800,000 strokes occur in the US each year, with roughly 15 million worldwide. Acute ischemic stroke (AIS) accounts for 87% of all cases, and among AIS patients, an estimated 40 to 50% present with large vessel occlusion amenable to mechanical thrombectomy (MT), producing a pool of 280,000 to 350,000 annual candidates in the US alone.

Mechanical thrombectomy is the standard of care for large vessel occlusion AIS, yet only 10 to 17% of eligible patients currently receive the procedure. ARPA-H states the figure bluntly: only 12% of Americans in need receive a thrombectomy. The bottleneck is human. The procedure requires a fellowship trained neurointerventionalist, and the US has approximately 2,000 practicing physicians qualified to perform cerebral thrombectomy. Geographic disparity compounds the shortage: only 27.7% of rural populations have access to a comprehensive stroke center capable of thrombectomy, compared to 69.5% of urban populations. One in six US patients lacks timely access to any thrombectomy capable facility.

The clinical urgency is extreme. Every 15 minute delay to recanalization reduces the probability of functional independence by approximately 4%. Every 10 minute delay in treatment generates roughly $10,000 in downstream healthcare costs from extended rehabilitation, long term care, and lost productivity. The eligible procedure window spans 6 to 24 hours from symptom onset depending on imaging selection criteria, but outcomes degrade steeply with time. The unmet need is a system that can perform autonomous catheter navigation, clot engagement, and retrieval in facilities that lack a neurointerventionalist on site, converting community hospitals into thrombectomy capable centers without requiring a scarce specialist at the bedside.

2. State of the Art

Three parallel research paradigms have converged to establish technical feasibility for autonomous endovascular thrombectomy.

Autonomous catheter navigation. Artedrone (rebranding to Carvolix), based in Paris, developed the SASHA system and demonstrated end to end autonomous mechanical thrombectomy in porcine cerebral vasculature in April 2025. The system performs catheter navigation through the aortic arch, clot engagement in the target cerebral vessel, and clot retrieval without human manipulation of the catheter. Principal investigator Dr. Frédéric Clarençon (Pitié-Salpêtrière Hospital) leads the clinical program. Artedrone has raised $22.6M in total funding from Elaia Partners, BPI, and Kurma Partners. First human trial is planned for 2027.

Reinforcement learning for vascular navigation. A King's College London team demonstrated TD-MPC2 (Temporal Difference Model Predictive Control 2), a world model based reinforcement learning controller, achieving 65% mean success rate across 10 patient specific cerebral vasculatures (arXiv 2503.24140, accepted MICCAI 2025). The approach uses a multi agent architecture where different RL agents handle different vascular pathway segments, trained on the CathSim simulation platform that enables training on diverse anatomical configurations derived from clinical CT angiography data. Separate RL work in the endovascular navigation literature achieved 100% success rate with reward shaping at 22.6 second average procedure time, demonstrating that RL controllers can learn fast, reliable catheter navigation when the reward structure is well designed.

Magnetic microrobot navigation for vascular therapy. ETH Zurich published results on "clinically ready" magnetic microrobots in Science (2025). Spherical gel capsules containing iron oxide nanoparticles navigated at 4 mm/s in porcine and ovine vasculature under clinical imaging guidance. The University of Twente and Radboudumc demonstrated 3D printed screw shaped magnetic microrobots (approximately 1 mm diameter) performing thrombectomy in sheep iliac artery, immediately restoring blood flow (February 2025). Harbin Institute of Technology showed that tPA conjugated nanorobots achieved a 20x improvement in thrombolysis rate versus pure tPA with near complete recanalization in 40 minutes (Science Advances, 2024).

The convergence is clear. Autonomous catheter systems (Artedrone) validated the clinical concept in a relevant animal model. RL controllers (KCL) proved that navigation through patient specific vasculature can be learned from simulation. Magnetic systems (ETH, Twente) demonstrated alternative actuation modalities for submillimeter scale vascular intervention. What remains absent is the integration of RL driven autonomous navigation with clinical grade catheter hardware, validated in patient specific neurovascular anatomies at the success rates required for human use.

3. Foundational Research

Artedrone/Carvolix (2025). SASHA system preclinical results: autonomous mechanical thrombectomy in porcine cerebral vasculature. Pitié-Salpêtrière Hospital, Paris. Demonstrated end to end autonomous catheter navigation and clot retrieval in large animal cerebral vasculature without human manipulation. The system autonomously navigated from femoral access through the aortic arch to the target cerebral vessel, engaged the clot, and retrieved it. This is the first demonstration of fully autonomous mechanical thrombectomy in a clinically relevant cerebral model. TRL 4 (validated in relevant environment). Funded by $22.6M in private capital from Elaia Partners, BPI, and Kurma Partners.

Rajagopal A, Gudipati M, Wignall A, et al. (2025). "TD-MPC2 World Model for Autonomous Endovascular Thrombectomy Navigation." arXiv:2503.24140, accepted MICCAI 2025. King's College London. Demonstrated a model based RL controller achieving 65% mean success rate across 10 patient specific cerebral vasculature configurations derived from clinical CT angiography. The multi agent architecture assigns different RL policies to different vascular segments, enabling compositional navigation through tortuous anatomy. Trained entirely in the CathSim simulation environment. TRL 3 (proof of concept in simulated patient anatomies). Funded by UK Research and Innovation.

Landers FC, Weissen M, Veciana A, et al. (2025). "Clinically ready magnetic microrobots for gastrointestinal and vascular interventions." Science, 390(6718). DOI: 10.1126/science.adx1708. ETH Zurich. Demonstrated spherical gel capsule microrobots containing iron oxide nanoparticles navigating at 4 mm/s through porcine and ovine vasculature under clinical fluoroscopic and ultrasound imaging. The robots maintained structural integrity through physiological flow conditions and were visualized using standard clinical imaging modalities without contrast agents. TRL 4 (validated in large animal models under clinical imaging). Funded by Swiss National Science Foundation and European Research Council.

Behr J, Rajagopal A, et al. (2025). "CathSim: An Open Source Simulator for Autonomous Endovascular Catheterization." Accepted ICRA 2025. King's College London. Open source simulation platform enabling training and evaluation of autonomous endovascular navigation controllers on patient specific vascular anatomies derived from CT angiography. CathSim provides realistic catheter physics, vessel wall interaction modeling, and fluoroscopic image synthesis. The platform enables reproducible benchmarking of RL controllers across standardized anatomical configurations, addressing a critical infrastructure gap in the field. TRL 3 (simulation validated against physical phantom experiments). Funded by UK Research and Innovation.

Endovascular RL navigation with reward shaping (2024 to 2025). Multiple institutions. A series of studies demonstrated that well structured reward functions enable RL agents to achieve 100% catheter navigation success rates at 22.6 second average procedure times in simulation environments. Key insight: combining sparse goal rewards with dense intermediate guidance (distance to target, vessel wall proximity penalty, catheter curvature penalty) produces policies that are both fast and safe. These results establish the upper bound of what RL navigation can achieve when the reward function adequately captures the clinical objective, and highlight that the remaining challenge is sim to real transfer rather than algorithmic capability. TRL 3. Funded through various national research council grants.

4. Competitive Landscape

Artedrone/Carvolix (Paris). $22.6M raised. Pre-clinical stage with autonomous catheter navigation demonstrated in porcine cerebral vasculature. First human trial planned 2027. The only company with demonstrated autonomous thrombectomy capability in a cerebral model. Not yet commercial.

Remedy Robotics (San Francisco). $35M raised. Completed first human remote endovascular procedures in October 2025. However, the system is teleoperated: a remote physician controls the catheter in real time over a network connection. Remedy addresses the geographic access problem through telepresence but does not eliminate the physician bottleneck, as each procedure still requires a neurointerventionalist's full time and attention.

Microbot Medical LIBERTY (Yokneam, Israel). FDA 510(k) cleared September 2025. Single use disposable endovascular robotic system achieving 100% navigation success rate and 92% reduction in radiation exposure to the operator. Teleoperated, not autonomous. The LIBERTY system validates that robotic endovascular platforms can achieve FDA clearance, establishing a regulatory predicate, but it does not address the specialist bottleneck.

Corindus/Siemens Healthineers CorPath GRX (Waltham, MA). 510(k) cleared for coronary and peripheral vascular procedures. Teleoperated platform with sub millimeter precision. Not cleared for cerebrovascular use. Not autonomous. CorPath GRX serves as a strong predicate device for 510(k) regulatory strategy but does not compete in autonomous neurovascular thrombectomy.

Incumbent device manufacturers (Stryker, Medtronic, Penumbra). Dominate the thrombectomy device market with stent retrievers and aspiration catheters but have no autonomous navigation IP or robotic catheter platforms. Their business models center on consumable devices sold to neurointerventionalists, creating organizational inertia against autonomous systems that would shift value from the device to the robot.

The autonomous neurovascular thrombectomy space has fewer than three companies with any demonstrated capability. The market is pre-commercial with no products available for clinical use, passing the "not yet commoditized" gate decisively.

5. Total Addressable Market

Bottom up. Approximately 800,000 strokes per year in the US. With 87% ischemic, that yields roughly 696,000 AIS cases. An estimated 40 to 50% present with large vessel occlusion eligible for mechanical thrombectomy, producing 280,000 to 350,000 annual candidates. At the current treatment rate of 10 to 17%, only 28,000 to 60,000 procedures are performed per year. At full autonomous access (where any hospital with a CT scanner and the robotic system could offer thrombectomy) the addressable volume rises to 280,000 to 350,000 procedures annually. At an average procedure cost of $30,000 to $50,000, the US mechanical thrombectomy procedure TAM is $8.4B to $17.5B annually at full penetration.

Adjacent market sizing. The endovascular robot market stands at $1.1B (2025) and is projected to reach $3.25B by 2034, growing at 17.1% CAGR. The thrombectomy device market is projected to reach $2.63B by 2030 at 7.1% CAGR. The neurovascular segment constitutes 22.1% of endovascular robotics and is growing at 12.7% CAGR.

Top down cross check. The American Heart Association estimates annual US direct stroke costs at $56.5B. Mechanical thrombectomy is the highest acuity, highest reimbursement intervention in stroke care. A TAM of $8.4B to $17.5B represents 15 to 31% of total direct stroke expenditure, which is consistent with thrombectomy's position as the definitive treatment for the most severe (large vessel occlusion) cases.

Revenue model. Robotic mechanical thrombectomy systems are reimbursable under existing CPT codes: 61645 for intracranial thrombectomy and 37184 to 37186 for peripheral arterial mechanical thrombectomy. Medicare reimburses approximately $24,000 to $40,000 per MT procedure under hospital outpatient prospective payment. The autonomous system creates value by enabling thrombectomy at facilities that currently cannot offer the procedure due to lack of a neurointerventionalist, expanding the number of treatment sites from roughly 1,200 comprehensive stroke centers to potentially any of the 4,500+ primary stroke centers and acute care hospitals with CT capability.

6. Research Gap and Commercial Opportunity

Three gaps separate the current state of the art from a deployable autonomous thrombectomy system.

Simulation to real transfer for neurovascular RL controllers. The best published RL controllers achieve 65% mean success in simulation across patient specific vasculatures (KCL, MICCAI 2025). Clinical deployment requires greater than 95% success with zero vessel perforation. The gap is bridged through domain randomization (varying vessel geometry, friction, and flow parameters during training), curriculum learning (progressive difficulty from simple to tortuous anatomies), and patient specific pretraining where the controller is fine tuned on the patient's own CT angiography derived vasculature before the procedure begins.

Safety constrained policy optimization. An RL controller navigating through tortuous cerebral vasculature must guarantee vessel wall safety at every timestep. Standard RL maximizes cumulative reward without hard safety constraints. The gap requires constrained Markov decision process formulations where vessel wall contact force is bounded as a hard constraint rather than a soft penalty, combined with learned safety envelopes that trigger automatic catheter retraction when predicted contact force exceeds physiological thresholds.

Integration with physical catheter platforms. RL controllers are trained in simulation but must interface with physical electromagnetic or robotic catheter drive systems under real time fluoroscopic guidance. The translation requires sensor fusion (combining fluoroscopic imaging with electromagnetic catheter tip tracking), actuator calibration (mapping RL policy outputs to physical catheter feed and rotation commands), and latency compensation (the control loop must operate at sufficient frequency to maintain safety in a pulsatile flow environment).

Incumbents have not closed this gap because it requires simultaneous expertise in three domains that do not coexist in any single organization. Stryker and Medtronic sell thrombectomy devices (stent retrievers, aspiration catheters) but have no autonomous navigation IP. Siemens/Corindus has teleoperation but not autonomy. Academic RL groups publish navigation controllers but lack catheter hardware, animal lab access, and regulatory strategy. The opportunity belongs to a team that integrates RL autonomy, clinical catheter hardware, and a regulatory pathway built on existing predicates.

7. Comparable Funded Projects

ARPA-H AIR (Autonomous Interventions and Robotics). Explicitly targets autonomous thrombectomy. Program rationale states "only about 12% of Americans in need receive a thrombectomy." Other Transaction Authority (OTA) based solicitation. Proposers Day held December 16, 2025. The program signals federal recognition that the thrombectomy access gap is a national health security issue solvable through autonomous robotics.

DARPA MASH (Medical Autonomy for Surgical Hemorrhage). 36 month program funding autonomous hemorrhage control robots for battlefield and austere environments. Proposers Day September 2025. While focused on hemorrhage rather than ischemia, the program validates the broader thesis that autonomous endovascular intervention is a federal research priority.

NIH NINDS R01 stroke research. Active portfolio of $250K to $500K per year awards funding fundamental stroke neuroscience and device development. NIH Award 1R01EB032301 (NIBIB) specifically funds "AI guided endovascular navigation" research.

Artedrone/Carvolix. $22.6M in private capital (Elaia Partners, BPI, Kurma Partners) raised for autonomous catheter navigation and thrombectomy. The largest private investment in autonomous neurovascular intervention to date.

ETH Zurich magnetic microrobots. Funded by the Swiss National Science Foundation and European Research Council. Multi year program producing the Science 2025 publication on clinically ready magnetic microrobots for vascular intervention.

8. Opportunity Assessment

Technology Readiness Level: 4 (validated in relevant environment). Artedrone demonstrated autonomous thrombectomy in large animal cerebral vasculature. KCL demonstrated RL navigation in physical phantoms with patient specific vasculatures. ETH demonstrated magnetic microrobot navigation in large animal vasculature under clinical imaging.

Top three technical risks and mitigations.

Sim to real transfer gap. The 65% simulation success rate must reach greater than 95% for clinical viability. Mitigation: domain randomization across vessel geometry, friction coefficients, and hemodynamic parameters during training; curriculum learning with progressive anatomical difficulty; patient specific controller adaptation using preoperative CT angiography.

Vessel perforation during autonomous navigation. Cerebral vessels are thin walled and intolerant of mechanical insult. Mitigation: force sensing catheter tips providing real time wall contact measurement; constrained MDP formulations bounding contact force as a hard policy constraint; automatic abort triggers that retract the catheter when predicted forces exceed safety thresholds.

Clot composition variability. Clot mechanical properties vary substantially (fibrin rich vs. red blood cell rich), affecting retrieval strategy and success rate. Mitigation: multi modal imaging feedback combining CT perfusion with real time fluoroscopy to characterize clot composition before and during retrieval; adaptive retrieval strategies that select stent retriever vs. aspiration vs. combined approaches based on clot imaging characteristics.

Regulatory pathway. The system pursues 510(k) clearance via predicate devices: CorPath GRX (endovascular robotic navigation, 510(k) cleared) and Microbot Medical LIBERTY (disposable endovascular robot, 510(k) cleared September 2025). Autonomous navigation features that exceed predicate functionality may require De Novo classification. The critical regulatory distinction is locked vs. adaptive algorithm: the RL controller is trained offline on simulation environments derived from patient CT angiography and deployed as a frozen model (no learning during the procedure). This follows the "locked algorithm" pathway per FDA guidance on AI/ML in Software as a Medical Device, substantially simplifying the regulatory strategy. Post market learning from accumulated procedure data would be incorporated through periodic SaMD updates submitted as new 510(k) clearances, consistent with FDA's Predetermined Change Control Plan framework.

9. Team Requirements

Three capability domains must be represented in any team pursuing autonomous neurovascular thrombectomy.

Biomedical science. Cerebrovascular anatomy and stroke pathophysiology expertise sufficient to define clinically relevant success criteria, design animal and human trial protocols, and navigate the regulatory pathway from 510(k) predicate analysis through IDE application. Includes catheter engineering (materials science, tip force sensing, guide catheter and microcatheter design), clinical trial design for interventional neurology endpoints, and reimbursement pathway analysis for hospital outpatient prospective payment under existing CPT codes.

Computer science. Reinforcement learning algorithm development with specific expertise in model based RL (TD-MPC2, SAC, constrained MDP formulations), sim to real transfer methodology (domain randomization, system identification, curriculum learning), simulation environment design for vascular physics, and safety constrained optimization ensuring hard guarantees on vessel wall contact forces. Experience with real time control systems operating under latency constraints is essential, as the navigation controller must maintain safety in a pulsatile hemodynamic environment.

Manufacturing engineering. Medical device manufacturing capability spanning catheter and endovascular device production, design for manufacturability at clinical volumes, ISO 13485 quality management system compliance, and production scaling from prototype to clinical trial quantities. Includes electromagnetic or robotic drive system assembly, sterilization validation, and supply chain management for single use catheter components.

The combination of cerebrovascular clinical expertise (ensuring the system solves the right clinical problem), RL algorithm development (enabling autonomous navigation at clinical success rates), and manufacturing capability (bridging the prototype to production gap that stalls most academic surgical robotics programs) is what differentiates a fundable autonomous thrombectomy program from an academic publication.