🤖 machine learning

Toward Safe Autonomous Robotic Endovascular Interventions using World Models

This paper presents a world-model-based autonomous navigation framework using TD-MPC2 for mechanical thrombectomy that demonstrates superior robustness and safety compared to state-of-the-art Soft Actor-Critic algorithms in both simulation and in vitro experiments, offering a promising path toward generalizable AI-assisted endovascular interventions.

Original authors: Harry Robertshaw, Nikola Fischer, Han-Ru Wu, Andrea Walker Perez, Weiyuan Deng, Benjamin Jackson, Christos Bergeles, Alejandro Granados, Thomas C Booth

Published 2026-04-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Harry Robertshaw, Nikola Fischer, Han-Ru Wu, Andrea Walker Perez, Weiyuan Deng, Benjamin Jackson, Christos Bergeles, Alejandro Granados, Thomas C Booth

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a doctor trying to perform a delicate surgery inside a patient's brain. To do this, they have to thread a tiny, flexible wire and a thin tube (a catheter) through a maze of blood vessels, starting from the groin and winding all the way up to the brain. It's like trying to thread a needle while riding a unicycle on a tightrope, but the "needle" is inside a living body, and the "unicycle" is a robot arm controlled by a human.

This procedure is called a Mechanical Thrombectomy (MT), and it's used to clear clots that cause strokes. It's incredibly difficult, dangerous, and requires years of training. Doctors also get tired, and the X-ray machines they use to see inside the body expose them to radiation.

The Problem: Robots Need to Learn the Maze

Scientists have been trying to build robots that can do this navigation automatically. They use a type of AI called Reinforcement Learning (RL). Think of RL like training a dog: you give it a treat (a reward) when it does something right, and nothing when it does something wrong. Eventually, the dog learns the best path.

However, previous attempts had a big flaw. The "dogs" (AI agents) were trained on just one specific type of blood vessel map. When they were shown a new patient with a slightly different anatomy, they got lost or crashed. It was like training a driver only on a straight highway and then expecting them to drive perfectly through a complex, winding city they've never seen before.

The Solution: The "Video Game Simulator" (World Models)

This paper introduces a smarter way to train the robot using something called a World Model.

Imagine you are playing a video game. Before you actually play, you build a mental model of the game world. You know, "If I jump here, I'll hit a wall," or "If I turn left, I'll find a shortcut." You don't need to actually crash into the wall a thousand times to learn that; you just simulate it in your head.

The researchers used an AI called TD-MPC2 that acts like this super-smart simulator. Instead of just reacting to what it sees right now, it builds a mental map of how the blood vessels move and bend. It "dreams" of thousands of different paths in a virtual world before it ever touches a real robot. This allows it to plan ahead and adapt to new, unseen patients much better than the old methods.

The Experiment: Virtual vs. Real

The team tested two AI "drivers":

The Old Driver (SAC): The current standard, which learns by trial and error without a mental map.
The New Driver (TD-MPC2): The one with the "World Model" simulator.

They tested them in two ways:

In Silico (The Video Game): They ran simulations on 15 different patient blood vessel maps. The new driver was much better at finding the right path (58% success rate vs. 36% for the old one).
In Vitro (The Real-Life Test): They built a plastic, transparent model of a human aorta and brain vessels. They put a real robot arm in a mock operating room and used real X-ray cameras (fluoroscopy) to guide the robot, just like a real surgery.

The Results: Slower but Safer and Smarter

Here is what they found:

Success Rate: The new AI (TD-MPC2) was significantly better at navigating the complex, tricky parts of the maze, especially in the simulations. In the real plastic model, it did slightly better, though the difference wasn't huge yet.
Safety: A major concern is that the robot might push too hard against the delicate blood vessel walls, causing a tear. The new AI was very gentle. Even though it pushed slightly harder than the old AI, the force was still 10 times lower than the level that would cause damage. It was like a cat walking on a table vs. a dog jumping on it; both might bump the table, but the cat is much safer.
Speed: The new AI was slower. It took about twice as long to finish the task. Why? Because it was being cautious. It was thinking, "If I go fast here, I might hit a wall, so let me slow down and check my map first." The old AI was faster but more reckless, often crashing into dead ends.

The "Superhuman" Challenge

There was one specific task (turning into the left carotid artery) that was so difficult that even human experts struggle with it using standard tools. The new AI managed to solve this in the computer simulation (which is "superhuman"), but failed in the plastic model. The researchers realized that in the real world, doctors use special, curved tools for this specific turn, which the robot didn't have yet.

Why This Matters

This study is a huge step forward because it proves that an AI trained in a computer can successfully navigate a real, physical robot in a realistic setting.

Generalization: The AI didn't just memorize one map; it learned how to navigate, so it could handle new patients it had never seen before.
Safety: It proved that AI can be safer than humans by keeping contact forces low.
The Future: While the robot is currently a bit slow and needs better tools for the trickiest turns, this "World Model" approach is the key to making autonomous stroke surgery a reality. It could eventually allow doctors in small hospitals to perform these life-saving surgeries with the help of a smart robot, or let specialists operate remotely from a different city.

In short: The researchers taught a robot to "dream" about navigating blood vessels. This dreamer is slower than a reckless driver, but it's much smarter, safer, and better at handling new, confusing maps than any robot we've had before.

1. Problem Statement

Mechanical Thrombectomy (MT) is a critical procedure for treating ischemic strokes, but it faces significant challenges:

Clinical Limitations: Limited access to specialized centers, long transfer times, and high operator radiation exposure.
Automation Challenges: Autonomous endovascular navigation requires handling highly variable vascular geometries and ensuring real-time, accurate control.
Current RL Limitations: Existing Reinforcement Learning (RL) approaches, particularly model-free algorithms like Soft Actor-Critic (SAC), often struggle with:
- Generalization to diverse patient anatomies (hold-out data).
- Long time-horizon tasks required for full MT procedures.
- Safety metrics (e.g., excessive contact forces causing vessel rupture).
- Lack of rigorous in vitro validation using patient-specific phantoms and fluoroscopy.

2. Methodology

The authors propose a World Model-based framework using the TD-MPC2 algorithm (a model-based RL method) and compare it against the state-of-the-art model-free SAC algorithm.

A. Navigation Tasks & Environment

Tasks: The first phase of MT (navigating a guide catheter and guidewire from the femoral artery to the Internal Carotid Artery) was split into five distinct tasks (A1–A3R) covering the aortic arch and carotid arteries.
Dataset: 15 patient-specific CT angiography (CTA) scans were processed into surface meshes.
- 10 scans were used for training.
- 5 scans were held out for in silico evaluation to test generalization.
- Anatomical variations included Type-I (87%) and Type-II aortic arches.
Simulation (In Silico): Used the stEVE framework and SOFA physics engine. Devices (0.0441" catheter, 0.035" guidewire) were modeled with realistic stiffness and friction. Feedback was provided via 2D fluoroscopic tracking coordinates, mimicking clinical conditions without 3D geometry visibility.
Physical Testbed (In Vitro): A 3D-printed transparent phantom (Type-I aortic arch) was used in a mock operating room. A robotic manipulator with independent actuation for the catheter and guidewire was controlled by the RL agents using real-time fluoroscopic feedback.

B. Reinforcement Learning Agents

SAC (Baseline): A model-free algorithm with an LSTM layer for trajectory-dependent state representation. It optimizes a critic directly to maximize cumulative rewards.
TD-MPC2 (Proposed): A model-based algorithm that learns a latent dynamics model of the environment. It combines Temporal Difference learning with Model Predictive Control (MPC) and Cross-Entropy planning.
- Key Feature: It simulates environment transitions to plan action sequences over a short horizon, allowing for better handling of long episodes and complex dynamics.
- Training: Both agents were trained on the same 10 vascular anatomies with data augmentation (random scaling and rotation) to improve robustness.

C. Safety & Evaluation Metrics

Success Rate: Percentage of episodes reaching the target.
Path Ratio: Efficiency of the path taken (distance traveled vs. initial distance to target).
Procedure Time: Time to reach the target.
Safety Metrics:
- Tip Contact Forces: Measured in Newtons (N). A rupture threshold of 1.5 N was established as the safety limit.
- Tip Speed: Monitored to prevent rapid, unsafe movements.

3. Key Contributions

First World Model Validation: Demonstrated the first world-model-based agent (TD-MPC2) capable of performing multiple endovascular navigation tasks across hold-out patient vasculatures in simulation.
Rigorous In Vitro Translation: Validated the agents in a clinically relevant in vitro testbed using patient-specific phantoms and fluoroscopic guidance, bridging the gap between simulation and physical reality.
Safety-Centric Benchmarking: Provided comprehensive safety metrics (contact forces) showing that autonomous navigation can remain well below vessel rupture thresholds.
Comparative Analysis: Directly compared TD-MPC2 against SAC in both in silico and in vitro settings, highlighting the trade-offs between exploration safety and procedural speed.

4. Results

A. In Silico Evaluation (Hold-out Set)

Success Rate: TD-MPC2 significantly outperformed SAC (58% vs. 36%, $p < 0.001$ ).
Path Ratio: TD-MPC2 showed superior path efficiency (49% vs. 22%, $p < 0.001$ ).
Procedure Time: TD-MPC2 took longer (17.1s vs. 9.6s, $p < 0.001$ ), indicating a more cautious, deliberative planning strategy.
Safety: Mean tip contact forces for TD-MPC2 were 0.15 N, well below the 1.5 N rupture threshold. While forces were slightly higher than SAC, they remained safe.

B. In Vitro Evaluation

Success Rate: TD-MPC2 (68%) and SAC (60%) performed comparably ( $p = 0.567$ ), though TD-MPC2 showed a trend toward higher success in specific difficult tasks (e.g., A2R: 60% vs 40%).
Path Ratio: TD-MPC2 achieved significantly better path ratios in task A2L (71% vs. 47%, $p = 0.017$ ).
Procedure Time: TD-MPC2 was significantly slower (111.0s vs. 66.2s, $p < 0.001$ ) due to the physical constraints of the robot and the agent's cautious planning.
Failure Modes: Both agents struggled with the most complex task (A2L, left carotid access) due to anatomical tortuosity and friction differences between simulation and reality. SAC failures were often due to wrong branch catheterization; TD-MPC2 failures were often due to timeouts.

5. Significance & Conclusion

Generalization: The study proves that a single world-model agent can generalize to unseen patient anatomies better than model-free SAC, a critical step toward clinical applicability where patient anatomy varies wildly.
Safety: The research confirms that autonomous navigation can be performed with contact forces an order of magnitude below the vessel rupture threshold, addressing major safety concerns regarding AI in surgery.
Clinical Translation: By validating the system with fluoroscopic guidance and physical phantoms, the authors demonstrate a viable pathway from in silico proof-of-concept to physical robotic assistance.
Trade-off: The results highlight a trade-off where the model-based approach (TD-MPC2) sacrifices procedural speed for higher success rates and safety in complex, unseen environments.

Limitations: The study was limited to one in vitro model, and the agents still struggled with the most anatomically challenging tasks (A2L) in the physical world, suggesting a need for specialized catheter shapes or further training data diversity. Future work will focus on expanding the dataset and measuring forces in physical phantoms more directly.