Nvidia pivots to physical AI for healthcare robotics

The framework treats healthcare robotics as physical AI systems, shifting the focus from text-based learning to contact, force, and consequence. By combining classical physics modeling—which handles the mechanical rules of catheters and tissue resistance—with generative AI that simulates visual scene dynamics, Nvidia allows developers to generate edge cases on demand. Instead of waiting for a rare surgical complication to occur in an operating theatre, teams can now simulate a guidewire catching on a vessel wall or a kidney stone lodged at an unusual angle.

Scaling clinical simulation

Running these simulations on GPUs using Nvidia’s Warp and Newton libraries enables massive parallelization. In benchmarks, the company reported that executing 8,192 parallel environments reduced training time from over five hours to under two minutes. Early adopters are already experimenting with this scale: CMR Surgical has contributed 500 hours of anonymized clinical data to the Open-H Embodiment dataset, while Johnson & Johnson MedTech is building a digital twin of its MONARCH platform to refine urological procedures. XCath and Inner Logic are similarly using the synthetic data to train endovascular autonomy and validate device mechanics.

Despite these gains in throughput, significant hurdles remain. While parallel simulation accelerates the exploration of failure modes, it does not inherently guarantee clinical reliability. The gap between simulated physics and the chaotic reality of a surgical suite remains unbridged. Furthermore, none of these systems are currently deployed on patients; they remain training exercises designed to build an evidence trail for regulatory bodies like the FDA. By open-sourcing the framework, Nvidia aims to allow clinical review boards to inspect the underlying physics assumptions, a necessary step for any system intended to operate within a human body.