NVIDIA Isaac GR00T-H-N1.7

NVIDIA Isaac GR00T-H-N1.7 is the surgical branch of NVIDIA's open robot foundation model: a post-trained variant of Isaac GR00T N1.7 adapted to operating-room robots using the Open-H-Embodiment dataset. Its significance is licensing as much as engineering. The first GR00T-H, built on GR00T-N1.6-3B, shipped under NVIDIA's OneWay Noncommercial License and was explicitly not approved for commercial use; this version carries the NVIDIA Open Model License and is marked ready for commercial use. In a field where every robot maker has so far had to train from scratch or negotiate proprietary data agreements, that is the whole point of the release. The architecture follows the parent model. A Cosmos-Reason2-2B vision-language backbone with a SigLip2 vision encoder takes in camera images and the text instruction; a flow-matching diffusion transformer produces continuous action vectors, interleaving self-attention over proprioception and actions with cross-attention to the vision-language embeddings, and conditioning each diffusion step through adaptive layer normalisation. The model has 3 billion parameters and ships in BF16. Inputs are a variable number of RGB frames at any resolution, a proprioception vector and a text instruction; the output vector is sized to the degrees of freedom of the particular robot. What makes it surgical is the data. Open-H-Embodiment, published by NVIDIA with 35 partner organisations under CC BY 4.0, covers 770 hours, 124,019 episodes and 119 datasets across 20 robot platforms from more than 50 institutions — Johns Hopkins, Stanford, UC Berkeley, UC San Diego, TU Munich, Vanderbilt, Balgrist and Northwell Health on the clinical side, CMR Surgical, Moon Surgical, Rob Surgical and Tuodao among the manufacturers. Post-training used only the real-world surgical part of it: 601 hours, roughly 63,930 episodes, 58 datasets and seven platforms — CMR Versius (described separately in this catalogue), dVRK, dVRK-Si, Rob Surgical BiTrack, KUKA LBR iiwa, USTC Torin and UR5e — with 98% of the material for training and 2% held out. Ultrasound, endoscopy and simulation data were left for future work. One detail says a great deal about how a model like this can go wrong: the Versius-500 contribution is capped at 20% of training steps so that no single embodiment dominates the loss signal, with the remaining datasets sampled in proportion to their size. Action and camera conventions were standardised to make the mixture trainable at all: actions as relative end-effector positioning, and camera setups reduced to either a single third-person monocular view or that view plus wrist cameras and additional modalities such as ultrasound images. The robots behind the recordings were teleoperated either programmatically or by engineers, researchers, medical students and professional surgeons, in simulation, benchtop, ex vivo, in vivo and clinical settings. Evaluation is done on real robots rather than on a benchmark table, and NVIDIA publishes no scores for this checkpoint — only the 2% validation split used during training. The one public result in the line belongs to a prototype of the earlier GR00T-H, which completed an end-to-end suture in the SutureBot benchmark. The limitation is stated by NVIDIA in plain terms, and is repeated here for the same reason: the model is meant for robotics research and development, including surgical and robotic-ultrasound policies, benchmarking and method development. It is not intended for clinical deployment, patient care or medical decision-making. Two figures published by NVIDIA do not agree. The model card puts the Open-H-Embodiment dataset at 770 hours; NVIDIA's own announcement of the dataset says 778. This catalogue records both rather than picking one. NVIDIA also gives no release date for this checkpoint — the dated event in the line is the announcement of GR00T-H and the dataset on 16 March 2026.

▸Videos

No videos yet.

▸Reviews