- Nvidia’s framework models device-anatomy contact, flexible instruments, sensor inputs and robot policies using classical and generative simulation. [1] - The main workflow and tutorial repositories use Apache 2.0, while Cosmos-H-Dreams model weights carry Nvidia’s separate Open Model License. [2] - Nvidia’s cited paper supports a large GPU-training speedup, but it reports 32 parallel environments—not the 8,192 stated in Nvidia’s announcement—and did not test on a real robot.
[4] Nvidia has released Medical Physics Simulation, an open-source, GPU-accelerated layer within Isaac for Healthcare designed to let developers train and test medical robots before moving to physical prototypes and laboratory experiments. The software combines anatomy, device mechanics, sensor simulation and robot learning, including classical physics solvers and the Cosmos-H-Dreams generative video simulator.[1]
The release is distributed across Nvidia’s Isaac for Healthcare repositories rather than as a single package. The i4h-workflows repository provides end-to-end reference implementations, while i4h-tutorials includes standalone catheter and vascular-digital-twin components. Both use the Apache 2.0 license. Cosmos-H-Dreams also licenses its source under Apache 2.0, but its downloadable model weights use the Nvidia Open Model License.[2]
From CT scans to simulated instruments #
The most detailed new workflow targets endovascular procedures. It can ingest DICOM or NIfTI CT data, reconstruct patient-specific vessels and simulate catheters or guidewires using an XPBD/Cosserat-rod solver. A GPU renderer produces fluoroscopy and digital-subtraction angiography with configurable noise, scatter, blur and detector effects. Nvidia lists TAVR guidance, coronary intervention and general endovascular navigation as intended research applications.[3]
Nvidia said CMR Surgical and Cambridge Consultants are applying Cosmos-H-Dreams to soft-tissue surgery; Johnson & Johnson MedTech is modeling anatomy and kidney-stone scenarios for its MONARCH platform; XCath is training endovascular autonomy policies; and Medtronic Structural Heart is exploring simulated X-ray data for catheter-navigation research. These are development and research uses, not evidence of clinical validation.[1][5]
The validation record needs qualification. Nvidia’s post says a benchmark ran 8,192 environments and cut training from more than five hours to under two minutes. The linked FF-SRL paper reports the time reduction, but achieved it with 32 parallel environments on an RTX 2060 Mobile GPU. Its evaluation covered a state-based instrument-reaching task over five runs, and the authors said real-robot tests remained future work. The study supports the case for GPU-native training speed, but it does not substantiate Nvidia’s 8,192-environment figure or establish clinical performance.[4]
Companies mentioned #
Further sources #
[[1] Nvidia announcement describing Medical Physics Simulation, its architecture, ap… ↗](https://blogs.nvidia.com/blog/medical-physics-simulation-open-source/)
[[2] Isaac for Healthcare workflow, tutorial and Cosmos-H-Dreams repositories, inclu… ↗](https://github.com/isaac-for-healthcare/i4h-workflows)
[[3] Endoluminal workflow documentation covering CT ingestion, vascular reconstructi… ↗](https://github.com/isaac-for-healthcare/i4h-workflows/blob/main/workflows/catheter_navigation/README.md)
[[4] FF-SRL paper reporting the surgical reinforcement-learning benchmark, hardware,… ↗](https://arxiv.org/abs/2503.18616)
[5] CMR Surgical announcement describing its Versius Plus simulation work with Nvid… ↗
The stories that matter, in one email. Free — unsubscribe anytime.