Provenance Record

{"https://www.globenewswire.com/news-release/2026/03/17/3256875/0/en/CMR-Surgical-Advances-Physical-AI-to-Support-the-Future-of-Robotic-Surgery-with-NVIDIA.html":"Confirmed: CMR contributed ~500 hours of anonymized Versius surgical data (largest share of Open-H). Chris Fryer quote verified: 'building foundations for the next generation of intelligent surgical systems.' Cosmos-H usage for synthetic data generation confirmed.","https://nvidianews.nvidia.com/news/nvidia-and-global-robotics-leaders-take-physical-ai-to-the-real-world":"Confirmed: Healthcare robotics partners listed (CMR Surgical, J&J MedTech, Medtronic). Note: this press release covers broader physical AI initiative and references GR00T N1.7/N2 but GR00T-H details come from Hugging Face model card. J&J Monarch platform and Medtronic IGX Thor exploration confirmed.","https://blogs.nvidia.com/blog/gtc-2026-news/":"Partially confirmed: GTC 2026 roundup covers J&J adopting IGX Thor for Polyphonic digital surgery platform, Medtronic evaluating IGX Thor, and KARL STORZ/LEM Surgical partnerships. Does not specifically name GR00T-H or Open-H dataset. Article correctly attributes Medtronic details to this source.","https://huggingface.co/nvidia/GR00T-H":"Fully confirmed: 3B parameters, GR00T N1.6 base, SigLip2 vision encoder, flow-matching transformer with AdaLN, 778 total hours from 35 institutions, 601.5-hour curated training subset (98%/2% split), all 7 robotic platforms listed match article exactly. Non-commercial license and clinical deployment disclaimers verified."}

All technical claims verified against sources. The 3B parameter count, architecture details, dataset statistics (778 hours, 35 institutions, 7 platforms, 601.5-hour training subset), and CMR's ~500-hour contribution all confirmed. The article correctly notes that the NVIDIA newsroom covers the broader physical AI initiative while attributing GR00T-H specifics to the Hugging Face model card. David Niewolny quote sourced from NVIDIA materials.

Excellent in-depth news article with strong technical accuracy and appropriate sourcing. The article correctly synthesizes information from four sources, attributes claims to the right sources, and maintains a balanced perspective by acknowledging uncertainties. The 'What We Don't Know' section is particularly well-crafted. Approved.