Can a Robot
Learn Surgery
by Watching?

Surgery has an old teaching rhythm: watch an expert, try it with help, then do it yourself. Nobody hands a trainee a rulebook for every millimetre of movement.

For decades, robots were taught the opposite way. Engineers wrote code for each move. One Johns Hopkins roboticist said someone could spend a decade just modelling suturing — for a single type of operation. [5]

In July 2025, his team reported a different route. A system called SRT-H (Surgical Robot Transformer–Hierarchy) learned a long phase of gallbladder removal by watching videos of surgeons doing it on pig cadavers, with captions describing each task. [1]

Then it did the job itself: a string of 17 tasks — finding the duct and the artery, grabbing them precisely, placing clips, and cutting with scissors. [2] It succeeded on all eight gallbladders it was tested on, though it took longer than a human surgeon. [2] [3]

The gallbladder was not a random pick. It is the operation that started the keyhole-surgery era.

On 12 September 1985, a German surgeon, Erich Mühe, removed a gallbladder through tiny cuts using a scope he built himself. [9] When he presented it in 1986, colleagues met him with ridicule; some called it “Mickey Mouse surgery.” [10] He had done 94 such operations before Philippe Mouret in France did his first in 1987 — and for years it was the French who got the credit. [9] Mühe was formally recognised as the first only in 1999. [9]

The Johns Hopkins team says the same thing about their choice: gallbladder removal was the first operation done the minimally invasive way, so it has a “lighthouse” role for any new technique. [8]

There is a practical reason too. Keyhole and robotic surgery already happen on a screen, through a camera inside the body. The operation is filmed by its very nature. That is what makes “learning by watching” possible at all.

The robot doesn’t “see” a video the way we do. Each demonstration becomes two streams: the camera picture, and the kinematics — the angles and positions of the instruments, turned into numbers. [5]

The brain is a transformer, the same family of AI that runs ChatGPT. ChatGPT predicts the next word. This model predicts the next movement. The method is called imitation learning: do what experts did in situations like this one. [4]

A small trick mattered a lot. When the team trained the robot to copy movements as fixed positions in space, some tasks failed almost every time. When it learned each move relative to where its own instrument tip already was — “a little left of here” rather than “go to point X” — results improved sharply. [6] Human hands work much the same way.

The payoff is speed of teaching. The earlier version learned needle handling, tissue lifting and suturing, and the team said a new procedure could be taught in a couple of days rather than years. [4]

Short tasks take seconds. A whole phase of an operation takes many minutes and dozens of decisions. One small slip early on can derail everything after it. That is why the new system has two layers — the “Hierarchy” in its name.

The top layer plans in plain language, choosing the next step or a correction; the lower layer handles fine motion. [7] Because the planner works in words, a person can talk to it. In tests it took spoken commands such as “grab the gallbladder head” and corrections like “move the left arm a bit to the left”, and learned from that feedback. [1] [2]

A co-author compared this to training residents, who master different parts of an operation at different rates. [2]

The result is real, and so are its limits. The organs came from pig cadavers: no heartbeat, no breathing, no fresh bleeding. The team made it harder by moving the robot’s starting position and adding blood-like dyes that changed how the tissue looked — and it still coped. [2] But a lab bench is not a living patient.

The same group had already crossed a different line in 2022, when their earlier robot, STAR, operated on a live pig. That robot, though, needed specially marked tissue and followed a rigid plan — like driving a carefully mapped route. SRT-H is the attempt to drive on any road. [2]

The team’s stated next step is studies in live pigs. [8]

If robots learn by watching, the next question is obvious: how much is there to watch?

A lot. Robots made by one company alone, Intuitive’s da Vinci, were used in about 31.5 lakh operations in 2025 (3.15 million), up from about 26.8 lakh in 2024. [11] More than 1,000 surgical cases a day now flow into the company’s digital system, from an installed base of over 12,000 machines. [12]

Every robotic operation records both things the robot needs: the camera view and the movement of every instrument, because the surgeon moves the instruments through the machine. Researchers have pointed to this huge repository of clinical recordings, with approximate movement data, as a training source. [6]

The catch, in our view: those recordings come from real patients and belong to hospitals and companies. Who may use them, with whose permission, and how mistakes in the recordings are filtered out are questions the technology alone can’t settle.

Robotic surgery is no longer rare in India. Data presented at the Society of Robotic Surgery India 2026 conference showed robotic-assisted operations rising from 8,912 in 2021 to 44,857 in 2025 — about five times in four years. [13] India has over 200 da Vinci systems. [14]

India also builds its own. The made-in-India SSI Mantra has 224 systems installed across 12 countries. [15] It is being used for telesurgery: more than 173 operations with the surgeon in another place, including a 2026 cardiac case across the Guyana–India distance. [16] In Bengaluru, Narayana Institute of Cardiac Sciences did India’s first robotic mitral valve replacement on the same system. [17]

Notice what all of these have in common: a human surgeon moves every instrument, whether from the next room or another continent.

Telesurgery moves the surgeon. Autonomy tries to copy the surgeon. They are often mixed up in headlines, but they solve different problems.

In 2017, a group of researchers proposed self-driving-car-style levels for medical robots, from Level 0 (the robot only follows the operator) to Level 5 (a robot surgeon needing no human), which they called science fiction. [18] They flagged a sharp legal point: at the top levels the robot is not just a medical device but is effectively practising medicine — and the US regulator, the FDA, regulates devices, not the practice of medicine. [18]

Where are real robots today? A review of 49 surgical robots cleared by the FDA from 2015 to 2023 found most at Level 1 (robot assistance), three at Level 3, and none at Level 4 or 5. [19] The FDA itself prefers the phrase “robotically-assisted surgical devices”, to stress that cleared systems still need the surgeon’s direct, continuous control. [19]

So the near future probably looks less like a robot surgeon and more like a modern car. The lead researcher expects small self-driving features, like brake assist and park assist, to arrive inside robots that surgeons still control. [8] One reason he gives: surgeons’ caseloads are expected to keep climbing as populations age. [8]

And the gallbladder has a lesson here. The keyhole method that was laughed at in 1986 became routine within a few years, once surgeons saw the results. [20] Whether autonomous steps follow that path will be decided by live-patient evidence, not by a video.