Open-H-Embodiment: A Large-Scale Dataset for Enabling Foundation Models in Medical Robotics

Figure 5 : Multi-Embodiment Performance Comparison. Evaluation of the GR00T-H foundational VLA (post-trained on Open-H) versus the GR00T-N1.6 base policy across three surgical platforms: the da Vinci Research Kit Si (dVRK-Si), CMR Versius, and Virtual Incision MIRA. GR00T-H demonstrates significant performance gains across all robot embodiments and sub-tasks, with the overall average success rate showing a statistically significant improvement ( p<0.001 ). Error bars represent the Clopper-Pearson 95% confidence intervals across trials.

Figure S2 : Movie S1 : Successful end-to-end suturing rollout by GR00T-H on the SutureBot benchmark. GR00T-H controls the dVRK-Si to pick up and hand over the needle, perform needle throw and extraction, and complete the task with a knot tie.

Figure S3 : Movie S2 : Successful sub-task rollouts by GR00T-H for ex vivo suturing. The policy is conditioned on current observations and sub-task prompts, with the prompts shown in the top left of the video.

Figure 1 : Open-H-Embodiment overview. ( A ) Geographic distribution of the 49 participating institutions across North America, Europe, the Middle East, and Asia. ( B ) The 20 healthcare robotic platforms represented in the dataset, spanning surgical systems (da Vinci Si, da Vinci Xi, dVRK, dVRK-Si, MIRA, Versius, BiTrack, Maestro, Torin), general-purpose manipulators adapted for clinical use (Franka Panda, UR5e, Kuka Med 14), and emerging platforms. ( C ) Representative frames from the dataset illustrating the diversity of tasks, viewpoints, and tissue types covered, including robotic surgery

Figure 2 : Composition of the Open-H-Embodiment dataset. ( A ) Dataset hours by robot platform. ( B ) Distribution of dataset hours by environment type. ( C ) Distribution of dataset hours across task families. Together, these panels summarize the current distribution of contributed data across embodiments, collection environments, and task families in Open-H-Embodiment.

Figure 7 : Quantitative evaluation of Cosmos-H-Surgical-Simulator. Per-frame L1 and SSIM for benchtop vs. tissue-based datasets. Mean L1 error and SSIM as a function of generated frame index across 72 autoregressively generated frames (6 chunks × 12 frames each). Mean over 3 generation seeds, each averaged across all evaluated episodes within the category; shaded bands indicate 1 standard deviation across seeds. Left: benchtop datasets (phantom and bench procedures). Right: tissue-based datasets (clinical, cadaver, and ex vivo tissue).

Figure S1 : Experimental setups for GR00T-H policy evaluation. SutureBot: The primary evaluation environment, utilizing a da Vinci Research Kit Si (dVRK-Si), silicone phantom pad, and suture needle. Generalization: An out-of-distribution testbed featuring wound geometries absent from the training data and external lighting sources to replace standard endoscope illumination. MIRA and CMR: Robotic platforms used to validate multi-embodiment policy performance. Ex vivo suturing: Performed with skin-on pork belly to assess performance on real tissue.

Figure S4 : Movie S3 : Qualitative results from Cosmos-H-Surgical-Simulator across 30 Open-H datasets, 9 institutions, and 9 embodiments. Each panel shows ground-truth observations (left) alongside model-predicted frames (right), conditioned on recorded kinematic action trajectories.

Figure 3 : SutureBot end-to-end and out-of-distribution evaluation. Top: Task survival for GR00T-H on the SutureBot end-to-end suturing task compared to GR00T-N1.6, ACT, and LingBot-VA. GR00T-H improves end-to-end performance with a long-horizon success rate of 25%, showing better handling of compounding errors. Bottom: Out-of-distribution evaluation on SutureBot with an unseen wound configuration under varied lighting ( n=20 per subtask). GR00T-H achieves a 3-task average of 54%, outperforming GR00T-N1.6 (30%) and ACT (5%). Clopper-Pearson 95% confidence intervals are represented as error bar

Figure 4 : SutureBot data efficiency evaluation. Task success rate at 33% and 100% fine-tuning data on SutureBot ( n=10 per subtask). At 33% data, GR00T-H matches ACT while GR00T-N1.6 underperforms both. At 100% data, GR00T-H outperforms all baselines, indicating that Open-H post-training enables both data-efficient learning and stronger scaling. Clopper-Pearson 95% confidence intervals are represented as error bars.

Figure 6 : Ex vivo suturing evaluation across 29 subtasks. ( n=10 per subtask, n=290 total). Tasks span needle manipulation, wound opening, suture passing, knot tying, and suture cutting. GR00T-H achieves an average success rate of \approx 64%, with near-perfect performance on structured manipulation primitives and lower success on fine-contact and cutting steps. The rightmost bar shows the overall average across all subtasks. Clopper-Pearson 95% confidence intervals are represented as error bars.