Nvidia Releases Surgical Robotics Models for Research

Nvidia surgical robotics models arrived on 20 October, when the company published two new models on Hugging Face for research use. The models are called Cosmos3-H-Embodied-Nano and Cosmos3-H-Dreams, and both are dated 20 October 2026 on their model cards. They are released under the Nvidia Open Model License Agreement, which sets out how researchers and developers can use them. The models are designed for surgical robotics, a field where AI can help robots understand their surroundings, predict movements, and support surgeons during procedures. Nvidia describes the work as part of its Cosmos family of world models, which are trained to understand physical environments and how actions change them.

Cosmos3-H-Embodied-Nano is described as a post-trained version of Cosmos3-Nano, trained on open surgical robotics data for forward dynamics, inverse dynamics, and policy, so that one checkpoint can predict robot actions and endoscopic video. Cosmos3-H-Dreams is a real-time, action-conditioned surgical world model for the CMR Surgical Versius system, built on the Nano model and distilled into a faster student model that generates endoscopic video from a camera frame and robot commands. The models are not a finished medical product, and they are not a replacement for a surgeon, but they give researchers a starting point for building and testing surgical robotics systems.

Nvidia surgical robotics models support research

Nvidia surgical robotics models are part of a wider release on 20 October, which also included Isaac Lab 3.0 for robot simulation and an open release of TorchNeuron for AWS Trainium chips. Together, the releases show Nvidia’s focus on robotics as a practical field, from factory robots to surgical systems. For researchers, open models are valuable because they can be studied, tested, and improved by the wider community, rather than remaining inside a single company. Surgical robotics is a sensitive area, where safety and accuracy matter more than speed, and any real use would need extensive testing and approval. The models are a research step, not a clinical deployment, but they show how world models are moving from general robotics into specialised fields such as surgery, where understanding the physical world is essential.

Source: USASI