Autonomous Endovascular Robot Clears 97% of Navigation Runs With Standard 2D Imaging Only

A new robotic navigation system can steer a guidewire through branching blood vessels autonomously — on a physical robotic platform — using only the standard 2D X-ray imaging already present in every hospital catheterization lab in the world. The result, published as a preprint on arXiv on October 17, eliminates what has been the defining infrastructure obstacle separating laboratory autonomous endovascular robots from real hospital deployment: the requirement for expensive 3D imaging reconstruction hardware that most facilities do not have during catheter procedures.

Across 65 closed-loop navigation runs on a physical dual-device robotic platform, all runs reached at least 95% of the planned guidewire path to the target vessel segment. Two runs — 3.1% of the total — triggered the prespecified failure criterion of three or more autonomous stop events within a single run. The remaining 63 runs, 96.9%, completed without hitting that threshold. These results are author-reported from a phantom bench study and have not yet been independently reproduced.

A Navigation Problem Every Catheterization Lab Can Pose

Endovascular procedures — inserting catheters and guidewires through blood vessels to treat blocked coronary arteries, intracranial aneurysms, or peripheral vascular disease — are performed millions of times per year worldwide. Robotic systems like the FDA-cleared CorPath GRX (Corindus, now Siemens Healthineers) reduce operator radiation burden by allowing remote control, but they remain semi-autonomous: a surgeon still makes every navigation decision.

Full autonomy requires the robot to perceive vessel anatomy, plan a path, and execute that path while detecting and recovering from errors. The central perception problem is that endovascular procedures normally use 2D digital subtraction angiography (DSA) — the imaging standard in every catheterization lab globally. Previous fully autonomous navigation systems sidestepped 2D's inherent ambiguity by importing a 3D model of the patient's anatomy from a preoperative CT or MR scan and registering that 3D volume to the live 2D view in real time. That 3D-2D registration requires additional imaging hardware and a preoperative scan not always part of the procedure workflow.

The paper from Sheng Cao, Shuxiang Guo, and colleagues takes a different position: accept the 2D image as the complete information environment and build a navigation framework that works within it.

Centerline Graphs and Five Programmed Behaviors

The system segments vessel anatomy from 2D DSA images and converts the resulting vessel masks into a labeled centerline graph — a topological data structure where branch points become nodes, vessel segments become edges, and each segment carries anatomical labels identifying which vessel it represents. A planned path from the access point to the target vessel is then parameterized by arc length, assigning a continuous distance coordinate to every point along the intended route.

This arc-length coordinate becomes the scalar progress metric the controller uses throughout the procedure. A tip-tracking algorithm monitors the guidewire tip in the live 2D image and projects its position onto the centerline path to obtain a continuous progress value. The controller uses this value, combined with the graph topology at the current position, to select one of five behavioral modes: delivery (advance the guidewire forward); branch selection (at a vessel split, choose the correct branch); probing (small exploratory motions when tip position is ambiguous); rollback (retract to the last confirmed correct position when off-path entry is detected); and recovery (resume navigation after an operator-review stop).

Operators can pre-mark specific locations as interaction-risk points — positions where vessel wall injury risk is elevated. When the guidewire approaches these points, the controller prioritizes retreat before the standard off-path detection threshold would trigger a stop. The entire controller is rule-based, with no neural network inference at navigation time. This architecture is deliberately chosen for a medical device context — rule-based behavior is transparent and auditable in ways that reinforcement-learned policies are not, which matters significantly for regulatory review.

The Competing System Requires Hardware This One Does Not

The most directly comparable published system is VascularPilot3D, described in 2024 as the first fully autonomous endovascular robot on physical hardware, reporting 100% success across 25 trials. The structural difference is precisely the 3D-2D registration requirement: VascularPilot3D constructs a 3D vascular model and registers it to the intraoperative 2D fluoroscopic view — a process requiring a preoperative 3D imaging scan and registration infrastructure. It achieves a higher reported success rate in a smaller trial (100% versus 96.9%, 25 versus 65 runs), but at the cost of additional hardware that limits where it can be deployed.

The trade-off is real. A 3D model provides richer spatial information about branching geometry and vessel depth — information that is genuinely ambiguous in a 2D projection. But the present paper's answer is that a well-designed centerline graph with a rule-based controller can resolve that ambiguity well enough to navigate successfully in 96.9% of runs while accepting zero new hardware requirements.

Zero New Hardware Is the Largest Practical Implication

Every catheterization lab worldwide already has fluoroscopy and DSA imaging capability. No new capital equipment purchase is required to deploy this navigation framework. For researchers planning clinical translation, that difference is fundamental: a first-in-human trial using this system could proceed at any facility with an existing robotic endovascular platform and standard DSA imaging — essentially every major catheterization lab currently evaluating endovascular robotics. A trial using a 3D-reconstruction-based system must recruit centers with matching 3D integration infrastructure, a far smaller pool.

The 65-run phantom study also clarifies what the success metrics actually mean. Even the two runs that triggered the failure criterion reached at least 95% of the planned path before their third stop was counted. The system never became catastrophically lost — it became cautious, halting for operator review rather than forcing the guidewire into incorrect anatomy. The 96.9% pass rate measures how often the operator was needed three or more times, not how often the guidewire went somewhere it should not.

What the phantom results cannot establish is in-vivo performance. Real vascular anatomy involves elastic deformation, calcified plaques, respiratory motion, and patient-specific variation that phantoms do not fully replicate. In-vivo animal study and, ultimately, first-in-human trials represent the required next steps before any regulatory submission. In the United States, an autonomous surgical navigation system would require 510(k) clearance or De Novo review; the FDA's Software as a Medical Device framework and its evolving AI/ML guidance would apply. The rule-based controller's inspectable decision logic may prove an asset in that process, given regulators' documented preference for AI systems whose behavior can be directly analyzed rather than inferred from statistical outputs.

The catheterization lab of the near future does not yet look different from today's. But the set of things a robot can do in it just expanded — and the infrastructure that limits when those things are possible did not get larger.