NVIDIA's GR00T-H Becomes the First AI to Complete Surgical Suturing on Its Own
Open-H-Embodiment releases 780 hours of surgical robot data across 50 institutions, enabling the first open foundation model to autonomously complete suturing end-to-end.
Takeaways
- Open-H-Embodiment is the largest open medical robotics dataset: 770 hours, 124,019 episodes, 20 robot platforms, 50+ institutions, all in LeRobot v2.1 format on Hugging Face.
- GR00T-H is the first open surgical VLA model, built by post-training NVIDIA's GR00T-N1.6-3B on 601 hours of surgical data; weights are public.
- GR00T-H is the only model to complete end-to-end autonomous suturing (25% of trials vs. 0% for all baselines including ACT and GR00T-N1.6), and achieves 64% average success across a 29-step ex vivo suturing sequence on porcine tissue.
- Cosmos-H-Surgical-Simulator is the first multi-embodiment surgical world model, spanning 9 robot platforms from a single checkpoint, enabling in-silico policy evaluation and synthetic data generation.
- GR00T-H shows strong data efficiency: with only 33% of fine-tuning data it matches ACT trained on the full dataset, and outperforms all baselines at 100% data (~73% vs ~50% for ACT).
- Current limitations include no live-tissue evaluation, no failure detection, and cutting steps reaching only 20-30% success; pre-clinical animal studies are the stated next step before any human use.
Surgical robots like the da Vinci have been in operating rooms for decades, yet they remain almost entirely dependent on a human surgeon at the controls. The gap between where medical robotics is and where it needs to be comes down to one thing: data. A new consortium of 50+ institutions, led by researchers at NVIDIA, Johns Hopkins, and Stanford, is trying to close that gap with Open-H-Embodiment, the largest open dataset of medical robotic video and kinematics ever assembled.
The Data Bottleneck Nobody Talks About
Autonomous medical robotics has been limited by a fundamental data problem: existing medical robotic datasets are small, single-embodiment, and rarely shared openly, restricting the development of foundation models that the field needs to advance. To put that in concrete terms: JIGSAWS, which served as the primary surgical manipulation benchmark for over a decade, contains just 3 hours of demonstrations on a single robot. The next best publicly available dataset, ImitateCholec, reached roughly 20 hours, also on a single platform. That is orders of magnitude smaller than the datasets driving general-purpose robotics.
The scarcity is not accidental. Collecting synchronized video and kinematics in a clinical environment means navigating institutional review boards, patient privacy regulations, and the challenge of instrumenting proprietary surgical hardware that was never designed for robotics research. The result has been a field stuck in a loop: no data, no foundation models, no foundation models, no generalization.
What Open-H Actually Contains
Open-H-Embodiment is the largest open dataset of medical robotic video with synchronized kinematics to date, spanning more than 50 institutions and multiple robotic platforms including the CMR Versius, Intuitive Surgical's da Vinci, da Vinci Research Kit (dVRK), Rob Surgical BiTrack, Virtual Incision's MIRA, Moon Surgical Maestro, and a variety of custom systems, spanning surgical manipulation, robotic ultrasound, and endoscopy procedures.
The numbers are striking compared to what came before:
| Dataset | Hours | Episodes | Platforms |
|---|---|---|---|
| JIGSAWS | ~3 | 103 | 1 |
| Exp. CRCD | ~7 | 80 | 1 |
| SutureBot | ~6 | 1,890 | 1 |
| ImitateCholec | ~20 | ~18,000 | 1 |
| Open-H (ours) | 770 | 124,019 | 20 |
The corpus spans 119 datasets totaling 770 hours, covering 20 distinct robot platforms, 33 task families, and 5 environment types ranging from digital simulation to live clinical procedures. Clinical data, predominantly contributed by CMR Surgical (489 hours of Versius procedures), accounts for 65% of the total corpus. Four clinical workflows each exceed 100 hours: prostatectomy (169 hours), cholecystectomy (172 hours), hysterectomy (122 hours), and hernia repair (119 hours).