Building physical intelligence for
endovascular surgery.

FDA-cleared surgical robot operated by GPT-6 Astra.

How it works.

Our physical harness enables frontier vision-language-action (VLA) models—including π_0.7 and GPT-6 Astra—to operate FDA-cleared surgical robots and associated peripherals. It provides a secure, low-latency interface for agentic surgery while accommodating human intervention and oversight.

Physical harness connects a VLA agent, clinician oversight, surgical and contrast robots, a patient, and fluoroscopy feedback.

How it performs.

We benchmarked Qwen3.8-27B, GPT-5.6 Sol, and GPT-6 Astra against the actions of trained residents across 512 endovascular procedures at the Corewell Health Cath Lab. GPT-6 Astra matched resident actions in 82% of decisions at both low and high thinking levels, taking approximately 5 and 15 seconds per decision, respectively.

Correct decisions by model and thinking level: Qwen3.8-27B scores 44%, 49%, and 56%; GPT-5.6 Sol scores 51%, 63%, and 65%; GPT-6 Astra scores 82% at both thinking levels.

Correct decisions versus latency per decision, comparing Qwen3.8-27B, GPT-5.6 Sol, and GPT-6 Astra at the tested thinking levels.

What's next?

We are building a reinforcement learning environment to further train surgical VLA agents on the same physical (VR) simulator used to train surgeons. Agents will learn to advance a catheter and autonomously request contrast from a separate robot that pushes a syringe. Agents will use fluoroscopy to detect when the catheter reaches its target and learn to minimize procedure time and contrast use.

A VLA agent controls catheter and contrast robots in a physical simulator, using fluoroscopy feedback to reach a goal while minimizing time and contrast.

Who we are.

Principal Investigators

Rachee Singh

Assistant Professor, Cornell University

Rachee is an Assistant Professor of Computer Science at Cornell University, where she leads the SysPhotonics group on systems and algorithms for efficient communication for machine learning and cloud applications.
Website
Emaad Manzoor

Assistant Professor, Cornell University

Emaad is an Assistant Professor at the Cornell Johnson Graduate School of Management and a graduate field member of Computer Science at Cornell, where he leads the Data Analytics Group on human-centered algorithm design.
Website
Dr. Ryan D. Madder

Interventional Cardiologist, Corewell Health, MI

Ryan is a board-certified interventional cardiologist and a fellow of the American College of Cardiology. He serves as the Program Director of the Interventional Cardiology Fellowship at Corewell Health, MI. During 14 years of practice at Corewell Health, he spent nine years as section chief of interventional cardiology and medical director of the cardiac catheterization laboratory. His research covers robotic coronary intervention, telestenting, and intracoronary imaging.
Website

Robotics in the cardiac cath lab (YouTube)

Research Team

Bhaskar Kataria Website
Jamal Hashim Website
Dr. Shuo Li Website

hello@projectwaldo.ai A LAUDE MOONSHOTS // ONE SEED AWARDEE

Building physical intelligence forendovascular surgery.