Back to timeline

Research · July 9, 2025

SRT-H removes a gallbladder by itself

On 9 July 2025 a group from Johns Hopkins University and Stanford published in Science Robotics SRT-H, a system trained on demonstrations. On eight pig gallbladders outside the body it performed, without human intervention, a phase of cholecystectomy: finding the cystic duct and artery, placing clips and cutting them. All eight attempts succeeded, 17 tasks in each.

Why it matters

Until then autonomous surgical systems had carried out single short actions on a rigid plan; here a machine for the first time went through a whole phase of an operation lasting several minutes. This is an editorial assessment. The system corrected its own mistakes without a person: its upper level followed the course of the operation and gave instructions in plain language. Learning from demonstrations, already at work with household robots, reached surgery.

How it is built. Two levels. The upper level decides which task comes next and phrases it, or a correction, in plain language, for example «move left arm towards me». The lower level carries out the motions for that instruction. Both were trained on demonstrations: 16,000 trajectories, about 17 hours, recorded on 34 gallbladders; verbal corrections of common mistakes were added to the data during training. The result. The phase took 317 seconds on average, and in each operation the system corrected itself 1 to 14 times. Reloading clips and changing tools were done by the operator and are not included in that time. In a preliminary comparison with an expert surgeon the quality was comparable, but the machine was slower. da Vinci. The system runs on a research version of the da Vinci robot (da Vinci Research Kit Si) with two extra cameras on the instruments: the same machine a surgeon steers in the clinic here moves by a trained program. What the record does not claim. The tissue is pig organs outside the body, not a live animal. One phase of the operation was automated, not the whole cholecystectomy. The release says the robot learned from videos of surgery and responded to voice commands while operating; in the article the language corrections come from the upper level itself, and human voice corrections appear only in the training data. The journal version was not read, behind a Cloudflare check; the figures come from the third version of the preprint.

Event record

Event date
July 9, 2025
Timeline date
Event date
Verification
Sources gathered automatically · September 29, 2026
Lines
ID
evt-0935

The day the article appeared in Science Robotics 10(104), per Crossref and Europe PMC, the same day as the Johns Hopkins release («published today»). The preprint arXiv:2505.10251 has three versions, of 15 May, 17 June and 8 July 2025; the figures here come from the third, dated 9 July 2025 on its first page.

Sources

Related events