1. Clinical Need
Acute ischemic stroke (AIS) accounts for approximately 87% of all strokes, striking roughly 800,000 Americans and 15 million people worldwide each year. Mechanical thrombectomy (MT) is the standard of care for large vessel occlusion AIS, with five landmark randomized controlled trials (MR CLEAN, ESCAPE, EXTEND-IA, SWIFT PRIME, REVASCAT) establishing its superiority over medical management alone. The procedure must be performed within a 6 to 24 hour window from symptom onset, with every 10 minutes of delay associated with approximately $10,000 in additional healthcare costs and measurably worse neurological outcomes.
Despite this evidence, only 10% to 17% of eligible AIS patients actually receive mechanical thrombectomy. ARPA-H data indicates that just 12% of Americans in need undergo the procedure. The bottleneck is access, not efficacy. MT requires a fellowship trained neurointerventionalist, and the United States has approximately 2,000 such specialists, concentrated in urban comprehensive stroke centers. One in six US patients lacks timely access to any thrombectomy capable facility. The disparity is geographic: rural populations reach comprehensive stroke centers at a rate of 27.7%, compared to 69.5% for urban populations. Patients in rural settings, overnight hours, and underserved regions face structural barriers that no amount of additional physician training can resolve within the current decade.
The clinical need is autonomous or semi-autonomous robotic thrombectomy that decouples the procedure from the physical presence of a neurointerventionalist. A system capable of AI guided catheter navigation under remote supervision (or fully autonomous operation in emergency settings) would convert the procedure from a specialist dependent intervention into one deliverable at any hospital with an angiography suite and a robotic platform. This is not a convenience improvement. It is a mortality and morbidity intervention for the 83% to 90% of eligible patients who currently receive no thrombectomy at all.
2. State of the Art
Three distinct paradigms have emerged for autonomous or AI assisted endovascular navigation, each validated in simulation or animal models but none yet deployed clinically for cerebrovascular thrombectomy.
Autonomous endovascular navigation (Artedrone / Carvolix)
The most advanced preclinical demonstration to date comes from Artedrone (now operating as Carvolix), a Paris based startup founded by Dr. Frédéric Clarençon at Pitié-Salpêtrière University Hospital. Their SASHA system achieved autonomous end to end mechanical thrombectomy in porcine cerebral anatomy in April 2025, navigating a catheter from femoral artery access to intracranial clot retrieval without human joystick input. Carvolix has raised $22.6 million and targets first human trials in 2027. This is the first and only published demonstration of fully autonomous MT in a large animal model with clinically relevant cerebrovascular anatomy.
Reinforcement learning for catheter navigation
Researchers at King’s College London published TD-MPC2 based catheter navigation at MICCAI 2025 (arXiv 2503.24140), achieving 65% mean success rate across 10 patient specific vasculature models derived from CT angiography. Their approach uses multi-agent reinforcement learning with temporal difference model predictive control, treating the catheter and guidewire as cooperative agents. Separately, CathSim (ICRA 2025) established a standardized simulation environment for endovascular catheter training, and complementary RL work with reward shaping has demonstrated 100% navigation success with 22.6 second average procedure times in simulation.
Magnetic microrobot approaches
ETH Zurich published magnetic microrobot navigation in Science in 2025 (Landers et al., DOI: 10.1126/science.adx1708), achieving 4 mm/s velocities in sheep and porcine vasculature. The University of Twente demonstrated 3D printed microrobots for sheep iliac artery thrombectomy in February 2025. At the nanoscale, the Harbin Institute of Technology reported tPA loaded nanorobots achieving 20x thrombolysis improvement over free tPA in Science Advances (2024). These approaches operate at a fundamentally different scale than catheter based MT and face distinct regulatory and manufacturing challenges, but they represent convergent evidence that autonomous vascular navigation is physically achievable.
The field is converging on a clear trajectory: autonomous catheter navigation guided by reinforcement learning, trained in simulation on patient specific vascular anatomy, and validated in animal models before human trials. The core technical question is no longer whether autonomous navigation is possible, but how to close the gap between 65% simulation success and the >95% reliability required for clinical deployment.
3. Foundational Research
Artedrone / Carvolix SASHA Preclinical Demonstration (2025). Autonomous end to end mechanical thrombectomy in porcine cerebral anatomy. First published demonstration of fully autonomous catheter navigation from femoral access to intracranial clot retrieval in a large animal model.
Establishes feasibility of autonomous MT at the system level. Validates that AI controlled catheter systems can navigate the tortuous path from femoral artery through the aortic arch and carotid bifurcation into intracranial vessels, the complete procedural trajectory that a human neurointerventionalist performs.
King’s College London, TD-MPC2 Catheter Navigation (2025). arXiv 2503.24140, accepted at MICCAI 2025. Multi-agent reinforcement learning for endovascular catheter and guidewire control across 10 patient specific vascular anatomies.
65% mean navigation success rate across diverse patient anatomies. Demonstrates that temporal difference model predictive control can handle the continuous, high dimensional action space of catheter manipulation. The multi-agent formulation (catheter and guidewire as cooperative agents) reflects clinical reality, where the interventionalist coordinates both devices simultaneously. The 65% success rate identifies the precise performance gap that must be closed for clinical translation.
Landers FC, Hertle L, Pustovalov V, et al. (2025). “Clinically ready magnetic microrobots for targeted therapies.” Science, 390, 710–715. DOI: 10.1126/science.adx1708.
Magnetic microrobot navigation at 4 mm/s in sheep and porcine vasculature under clinical fluoroscopy. While operating at a different scale than catheter based thrombectomy, this work validates that autonomous vascular navigation can be achieved in large animal models under physiological flow conditions and establishes imaging and control paradigms applicable to catheter scale systems.
CathSim: Endovascular Catheter Simulation Environment (ICRA 2025). Standardized simulation platform for training and evaluating AI driven catheter navigation policies.
Provides the simulation infrastructure required for RL policy training at scale. Bridges the gap between simplified tube phantom environments and patient specific anatomical models. Critical enabling technology for any group pursuing sim to real transfer of catheter navigation controllers.
RL with Reward Shaping for Endovascular Navigation (2025). Reinforcement learning catheter control with shaped reward functions achieving 100% navigation success and 22.6 second average procedure time in simulated environments.
Demonstrates that reward engineering can dramatically improve RL navigation performance in simulation. The 100% success rate (compared to 65% from TD-MPC2 on more complex anatomies) suggests that reward shaping, curriculum learning, and simulation fidelity are the primary levers for closing the performance gap, not fundamental algorithmic limitations.
4. Competitive Landscape
Artedrone / Carvolix (Paris, France). $22.6M raised. The only entity with a published preclinical demonstration of fully autonomous mechanical thrombectomy. First human trial targeted for 2027. Preclinical stage; not a commercial product. PI: Dr. Frédéric Clarençon. Key distinction: Carvolix is the technology leader, but remains pre-commercial and geographically concentrated in France. The autonomous navigation algorithms are not publicly available, and the regulatory path to US market access requires independent FDA engagement.
Remedy Robotics (San Francisco, CA). $35M raised. Developing a teleoperated robotic system for endovascular procedures. Critical distinction: Remedy’s architecture is teleoperation (remote physician controls the catheter), not autonomous navigation. This addresses the geographic access problem through remote connectivity but still requires a human neurointerventionalist for every procedure. Does not reduce the specialist bottleneck; redistributes it.
Microbot Medical LIBERTY (acquired by Bioventus). Received 510(k) clearance in September 2025 for a robotic endovascular catheter system. Teleoperated, not autonomous. The LIBERTY system is the first FDA cleared robotic platform specifically designed for neurovascular procedures, establishing a critical regulatory predicate.
Corindus / Siemens Healthineers CorPath GRX. 510(k) cleared for coronary and peripheral vascular interventions. Teleoperated robotic catheter platform with the largest installed base. Not indicated for cerebrovascular procedures. Establishes the broadest regulatory precedent for robotic endovascular systems.
Zero companies are pursuing autonomous cerebrovascular thrombectomy as a commercial product. Carvolix is preclinical. Remedy, LIBERTY, and CorPath are all teleoperated. The autonomous cerebrovascular navigation space is entirely pre-commercial, with published research from academic groups (KCL, CathSim consortium) but no entity combining autonomous RL navigation with clinical catheter hardware for MT.
5. Addressable Scope
US mechanical thrombectomy TAM
- AIS incidence (US): ~800,000/year; large vessel occlusion subset eligible for MT: ~200,000 to 280,000 patients annually
- Current MT penetration: 10% to 17% of eligible patients (~20,000 to 47,600 procedures/year)
- Full penetration scenario: all eligible patients receive MT
- Medicare reimbursement per MT procedure: ~$24,000 to $40,000 (CPT 61645 for intracranial thrombectomy; CPT 37184 to 37186 for peripheral)
- US MT TAM at full penetration: $8.4 billion to $17.5 billion annually
Endovascular robotics market
The endovascular robot market was valued at $1.1 billion in 2025 and is projected to reach $3.25 billion by 2034, growing at 17.1% CAGR. The thrombectomy device market specifically (catheters, aspiration systems, stent retrievers) is projected to reach $2.63 billion by 2030. An autonomous robotic thrombectomy system captures value from both the robotics platform and the per procedure disposables.
Revenue model
Reimbursement flows through established CPT codes: 61645 (intracranial mechanical thrombectomy), 37184 to 37186 (first and second vessel peripheral thrombectomy). Medicare reimburses approximately $24,000 to $40,000 per MT procedure. A robotic platform deployed at 100 hospitals performing 200 procedures per year at $30,000 average reimbursement yields $600 million in annual procedural revenue, with additional revenue from capital equipment sales, software licensing, and per procedure disposable kits.
6. Research Gaps and Opportunity
Three specific gaps separate published simulation results and animal demonstrations from a deployable autonomous thrombectomy system:
Gap 1: Sim to real transfer
The best published RL navigation success rate on patient specific vasculatures is 65% (KCL, TD-MPC2). Clinical deployment requires >95% reliable navigation across the full spectrum of patient anatomy, including tortuous vessels, atherosclerotic calcification, and anatomical variants. The gap is not algorithmic in nature; it is a matter of simulation fidelity, domain randomization, reward engineering, and training corpus diversity. Closing this gap requires high fidelity simulation environments built from large CT angiography datasets, systematic domain randomization to handle anatomical variability, and transfer validation on physical vascular phantoms before any animal work. Whoever demonstrates reliable sim to real transfer on patient specific neurovascular anatomy owns the enabling technology for all downstream clinical applications.
Gap 2: Safety constrained reinforcement learning
Current RL formulations optimize for navigation success without explicit vessel wall protection constraints. Cerebrovascular intervention carries catastrophic failure modes: vessel perforation causes intracranial hemorrhage, a condition with >50% mortality. Safety constrained RL (constrained Markov decision processes, safe exploration, Lagrangian methods for constraint satisfaction) must be integrated into the navigation controller to bound force application on vessel walls. This is not an optional enhancement. Regulatory approval will require demonstrated safety margins, and no published system has addressed this requirement explicitly.
Gap 3: Integration with clinical catheter platforms
Published RL work operates in simulation or on custom laboratory hardware. Clinical translation requires integration with FDA cleared (or clearable) robotic catheter platforms such as CorPath GRX or LIBERTY. This integration involves adapting RL control outputs to the specific actuation interfaces of commercial catheter systems, handling real time imaging feedback from clinical angiography suites, and managing the latency and noise characteristics of physical hardware. No academic group has published results on this integration challenge because it requires access to commercial robotic platforms and engineering resources beyond typical academic lab budgets.
Research thesis: The group that closes all three gaps establishes the platform for autonomous neurovascular intervention. The window is 3 to 5 years, constrained by Carvolix’s human trial timeline (2027) and ARPA-H’s investment in autonomous thrombectomy (AIR program). Academic labs will continue publishing simulation results; they will not build the clinical integration layer or navigate FDA clearance for a complete system.
7. Comparable Funded Projects
| Source | PI / Entity | Amount | Focus |
|---|---|---|---|
| ARPA-H AIR Program | Multiple (Proposers Day Dec 2025) | $5M–$25M/performer (est.) | Autonomous Interventions and Robotics; explicitly targets autonomous thrombectomy |
| DARPA MASH | Multiple performers | $2M–$10M/performer (36 months) | Autonomous hemorrhage control; overlapping autonomous endovascular navigation technology |
| NIH NINDS | Various PIs | $250K–$500K/yr (R01) | Stroke research, AI guided neurovascular intervention, thrombectomy outcomes |
| Private capital | Artedrone / Carvolix | $22.6M | Autonomous mechanical thrombectomy (preclinical, first human trial 2027) |
| SNSF / ETH Zurich | B. Nelson, ETH Zurich | CHF 300K–1M | Magnetic microrobot navigation for vascular intervention (Landers et al. 2025) |
The ARPA-H AIR program represents the highest alignment opportunity. Its Proposers Day in December 2025 explicitly named autonomous thrombectomy as a target application, signaling federal investment priority in precisely this technology space. DARPA MASH funds overlapping autonomous endovascular capabilities for hemorrhage control, creating technology transfer opportunities. NIH NINDS R01 mechanisms fund the foundational science (RL navigation, simulation environments, safety constraints) that feeds into larger system integration efforts.
8. Opportunity Assessment
TRL evidence chain
TRL 4: validated in relevant environment. Carvolix demonstrated fully autonomous MT in porcine cerebral anatomy (2025), the most clinically relevant large animal model for neurovascular intervention. KCL validated RL navigation on patient specific vasculatures in simulation (2025). CathSim established standardized simulation environments accepted at ICRA 2025. Microbot Medical LIBERTY received 510(k) clearance (September 2025) for a robotic neurovascular catheter platform, establishing regulatory precedent. The core technology components (autonomous navigation, robotic catheter control, neurovascular imaging integration) have each been independently validated; the integration of all three into a single deployable system remains at TRL 3 to 4.
Top 3 risks
Sim to real gap in patient specific neurovascular anatomy
Mitigation: Progressive validation pipeline from simulation through vascular phantoms to animal models. Domain randomization across CT angiography datasets to handle anatomical variability. Carvolix’s porcine demonstration proves the gap is closable; the question is cost and timeline, not feasibility.
HighVessel perforation during autonomous navigation
Mitigation: Safety constrained RL with explicit force limits and vessel wall proximity penalties. Real time force sensing at the catheter tip. Mandatory human override capability with <1 second latency. Layered safety architecture: algorithmic constraints, hardware force limits, and physician kill switch.
HighClot variability and composition affecting retrieval success
Mitigation: Multi-modal clot characterization from pre-procedural imaging. Adaptive retrieval strategies based on real time resistance feedback. Training on diverse clot phantoms (fibrin rich, erythrocyte rich, mixed composition). This mirrors the variability that human operators face; the RL controller can be trained on a far larger sample of clot types than any individual physician encounters in practice.
ModerateRegulatory pathway
510(k) clearance via predicate devices. The CorPath GRX (coronary/peripheral robotic catheter) and LIBERTY (neurovascular robotic catheter, cleared September 2025) establish direct predicates for a robotic endovascular catheter system. The autonomous navigation component introduces a software algorithm requiring FDA review under the predetermined change control plan (PCCP) framework for AI/ML enabled devices. A locked algorithm (fixed at deployment, not continuously learning) simplifies the initial regulatory path. Estimated timeline: 12 to 18 months for pre-submission strategy and CDRH engagement, followed by 18 to 24 months for clinical validation studies and 510(k) submission. Total: approximately 3 years to initial clearance for a defined autonomous navigation indication.
9. Team Capabilities
Successful pursuit of autonomous robotic thrombectomy requires three intersecting capabilities: cerebrovascular domain expertise and clinical translation strategy, reinforcement learning and autonomous navigation engineering, and manufacturing scale up for medical device production. HHA’s team provides coverage across all three:
Hass Dhia
MS Biomedical Sciences with medical school background (anatomy, physiology, pharmacology). AI infrastructure architect with production systems at scale. Provides cerebrovascular anatomy expertise required for patient specific vascular modeling, stroke pathophysiology analysis, clinical trial protocol design, and reimbursement pathway strategy. Leads experimental methodology and regulatory framing.
Haedar Hadi
MS Computer Science (Boston University, Information Systems focus). Specializes in reinforcement learning architectures, navigation algorithm design, and evaluation methodology. Provides the RL expertise required for autonomous vascular navigation: state, action, and reward formulation for catheter control, TD-MPC2 and SAC implementation, simulation environment construction from CT angiography data, and safety constrained policy optimization. Leads technical infrastructure and navigation controller development.
Ahmed Dhia
Director of Manufacturing with deep expertise in design for manufacturability (DFM), production scaling, and quality systems. Provides the manufacturing engineering capability that bridges laboratory proof of concept to production ready endovascular devices: the transition from prototype catheter systems to clinical volume production with ISO 13485 compliance, incoming material inspection, and batch record traceability.
This is the precise capability gap where most endovascular robotics research stalls. Academic labs publish navigation results but cannot manufacture clinical grade catheter systems. Ahmed’s background in production scaling and quality systems directly addresses the manufacturing barrier identified in this assessment as the bridge from prototype to deployable product.
10. Recommended Next Steps
Target funding programs
| Program | Mechanism | Range | Fit |
|---|---|---|---|
| ARPA-H AIR | OTA | $5M–$25M | Explicitly targets autonomous thrombectomy; highest alignment of any federal program |
| NIH NINDS | R01 | $250K–$500K/yr | Stroke research, AI guided neurovascular intervention; funds foundational RL navigation science |
| NSF CBET | R01-equivalent | $300K–$500K/yr | Bioengineering and transport systems; robotic catheter systems and vascular navigation modeling |
| DARPA MASH | Performer | $2M–$10M | Autonomous hemorrhage control with overlapping endovascular navigation technology |
| Schmidt Sciences | Open RFP | $1M–$5M+ | Embodied AI convergence; autonomous physical systems for health applications |
Estimated total funding range: $2M to $12M over 24 to 36 months for Phase 1 (autonomous navigation validation, catheter integration, and manufacturing feasibility).
24-month milestone timeline
- M1–3 Literature review completion. CT angiography dataset acquisition for patient specific vascular model generation. RL simulation environment design and initial architecture selection.
- M4–8 RL navigation controller v1 trained on simulated neurovascular anatomies. Manufacturing DFM analysis for catheter integration components. FDA pre-submission meeting preparation and regulatory strategy document.
- M9–14 Sim to real validation on vascular phantoms with catheter hardware. Prototype catheter integration with RL controller. Pre-submission meeting with CDRH for autonomous navigation indication.
- M15–20 In vitro autonomous navigation in patient specific phantoms with safety constrained RL. Manufacturing scale up feasibility assessment. Quality system gap analysis against ISO 13485 requirements.
- M21–24 Navigation controller performance publication. IND-enabling study protocol design. Phase 2 funding applications for large animal validation and manufacturing qualification.