A Surgical Robot Removed Gallbladders on Its Own—and No Human Touched the Controls

Source: NVIDIA | Published: 2026-08-10T21:00:15Z

Trained on 600 repetitions across 30 pig gallbladders, a Johns Hopkins surgical robot can now autonomously clip, swap instruments, and sever the bile duct without any human intervention.

Surgeons Are Now Teaching Robots to Tie Knots — Not Residents

Filippo Filicori, director of the robotic surgery training program at Lenox Hill Hospital, said something on stage that silenced the room: "The young surgeons actually operating on people right now are sitting at the console, operating on real patients, under my supervision. That's not fair. And it's not ethical."

Surgical simulation has existed for years, but the fidelity isn't there. High-end simulators like Intuitive Surgical's SimNow still rely on pixel-by-pixel rendering — slow and limited in realism. Filicori's team is working with NVIDIA to take a different approach using the Cosmos world model: rather than rendering every frame from scratch, they use the first frame of a surgery as a reference point, then incorporate robotic kinematic data to simulate instrument-tissue interaction. They've also developed a Gaussian model to replicate the deformation effects when instruments contact tissue.

The system is still in development, but the direction is clear: let robots learn to tie knots in simulation before they touch a real patient.

Gallbladder Removal: The First Framework to Complete Fully Autonomous Surgery

Axel Krieger from Johns Hopkins — who also serves as Chief Robotics Officer at Semaphor Surgical — presented what is currently the closest clinical result in autonomous surgery.

The team chose laparoscopic cholecystectomy as their proving ground, focusing on the most technically demanding phase: clipping and severing the bile duct and cystic artery. To build training data, they purchased 30 pig gallbladders, repeated each procedure 20 times, and accumulated 16 hours of surgical footage.

What they trained is a hierarchical policy model: a language policy at the top level continuously monitors the surgical video history, determines the current procedural phase, and decides whether correction is needed; below it, an action execution policy receives natural language instructions and outputs specific kinematic parameters. No human intervention required — in their ex vivo porcine tests, the robot autonomously completed the full sequence of clip placement, instrument swap, and duct transection, and self-corrected even when the initial grasp position was off.

The model also narrates its next intended step in advance, giving surgeons the ability to intervene at any point. Krieger noted that this design makes human-robot collaboration feel natural rather than like a black box operating blindly.

150,000 Trajectories: The Largest Open-Source Dataset in Surgical Robotics

Krieger's team released the OpenAGE dataset earlier this week: over 150,000 manipulation trajectories with kinematic data, totaling 1TB. It drew more than 1,000 downloads on launch day.

What makes the dataset significant is its design for cross-robot, cross-task training, formatted to the LeRobot standard. The team used it to fine-tune NVIDIA's Groot model and tested it on suturing tasks: with just 33% of the data, Groot already surpassed the previous state-of-the-art model on suturing.

They also used OpenAGE to train a Cosmos world model capable of generating photorealistic synthetic surgical video. In the demo, AI-generated footage was visually nearly indistinguishable from real surgery — a direct boon for policy testing and data augmentation.

What 3,000 Surgeries Taught Moon Surgical

Moon Surgical's Maestro occupies a distinct niche: not replacing the surgeon, but enabling one surgeon to control three instruments simultaneously — two hands on two instruments while the system autonomously manages the camera. They call this feature "Scope Pilot," and it's already deployed in live procedures.

After nearly 3,000 surgeries, they hit a predictable operational problem: table height, tilt angle, drape color, and surgeon height all vary between cases, meaning the robot arm's initial deployment position requires manual adjustment every time — slow and labor-intensive. They now run their full software stack in Isaac Sim to simulate various configurations, and use Cosmos Transfer to apply style transfer to real OR footage — generating synthetic scenes with different skin tones, lighting conditions, and drape colors without needing to actually operate under those conditions.

The team is candid about the gap that remains: generating photorealistic real-world images from Isaac Sim renders is still a work in progress. Cosmos Transfer automatically adds "hands" to the robot (because the base model's training data skews heavily toward human hands), requiring extensive fine-tuning to correct. They just received a version trained on Groot N1.6 — only to be told by their engineers that "that version is already outdated."

Why Hard-Tissue Surgical Robots Look Completely Different

Soft-tissue surgery — laparoscopic, da Vinci systems — follows an established architecture: surgeon at a console, two to four robotic arms operating remotely, no direct patient contact. That model has been validated over the past 30 years.

Orthopedic, spine, and joint surgery don't work that way. LEM Surgical founder Yossi explained the fundamental difference: the standard architecture for hard-tissue surgery is a single arm plus an infrared navigation camera, using bone markers for localization. "That's like hammering a nail with one hand," he said, "or tying your shoes one-handed."

LEM's solution is to replicate the human upper body: two manipulation arms plus one vision arm, operating in coordination. One arm stabilizes the bone and monitors position; another performs the grinding or cutting; the vision arm tracks everything in real time. This architecture enables things that a single-arm system structurally cannot do.

Another distinction Yossi emphasized — one often overlooked — is that an orthopedic OR always has 5 to 10 people present, some in the sterile field (who can touch the robot), some outside it (who cannot). The robot must coexist with hundreds of instruments from different vendors and make real-time determinations about who can approach and who cannot. "A decision matrix covering all those variables doesn't scale," he said. That's the core reason they're investing in AI and synthetic data training — not to make the robot smarter at cutting bone, but to ensure it understands how to collaborate with the people around it from day one.

The Data Bottleneck Isn't Volume — It's That the Annotation Infrastructure Doesn't Exist

Nearly every speaker raised the same issue: surgical data is difficult to extract from hospitals due to medico-legal risk; surgeons are nearly impossible to organize for annotation given how scarce their time is; and the field has long lacked basic ontological consensus — different departments and institutions don't even agree on what to call a given surgical movement.

A few months ago, Filicori's team convened a large group of SAGES members specifically to create a surgical gesture ontology. A month later, they applied that framework to the OpenAGE dataset. The ontology matters because the training quality of video-language-action models depends heavily on accurate classification of surgical gestures.

Video-language models play a key role here: embedded into the annotation pipeline, they can automatically identify critical moments in surgical video — the instant a vessel is inadvertently nicked, or the key anatomical landmark during hernia repair. That dramatically compresses the frame-by-frame annotation burden that would otherwise fall entirely on surgeons.

The roadmap for surgical autonomy is becoming clearer: camera control and instrument positioning will arrive first (lowest risk), followed by suction fixation, tissue retraction, and trocar placement, with full autonomous execution of complete surgical phases still further out. But 1,000 downloads on the day OpenAGE launched signals that the pace of progress across this field is accelerating — not slowing down.