Issue 1- Hands Without a Surgeon

Welcome to the first issue. If you subscribed before any issue existed, you did so on faith, and I will try to earn it. The format here is the one I intend to keep: a four-minute brief in three parts, written by one surgeon for others.

SRT-H: A Hierarchical Framework for Autonomous Surgery via Language-Conditioned Imitation Learning — Kim, Chen, Hansen, Shi, Goldenberg, Schmidgall, Scheikl, Deguet, White, Tsai, Cha, Jopling, Finn & Krieger — Science Robotics, July 2025.

For a decade, "autonomous surgery" has meant a robot doing one tidy thing in a rigged-up environment: a peg transfer, a single suture throw, a needle pass through scaffolded bowel held still with fixtures and markers. The honest reader filed these under demonstration, not surgery. This paper is the one worth clearing the desk for. A robot built on the da Vinci Research Kit completed the clipping-and-cutting phase of a cholecystectomy — clip the cystic duct, clip the artery, divide both — across eight unseen ex vivo porcine gallbladders, fully autonomously, at a 100% success rate. Not navigation. Manipulation, on tissue that varied from specimen to specimen, with the system reading each new anatomy and recovering from its own errors as it went. Anyone who has spent years watching this field promise more than it delivered should sit up: a real threshold has been crossed, and crossed cleanly.

The approach is the oldest one in surgery — learn by watching, then by doing. The group collected roughly 16,000 trajectories (about 17 hours) across 34 porcine gallbladders, then trained a two-tier system: a high-level policy that watches the scene and issues instructions in plain language ("clip duct," "move left arm to the right"), and a low-level policy that turns each instruction into robot motion. The part that should make a surgeon smile is the correction loop. When the low-level policy drifts, the high-level policy notices and talks it back on course — the way one might murmur a trainee's hand back to the right plane. On test, it averaged 5 minutes 17 seconds per case and corrected itself roughly six times per procedure. That is the genuine advance here: not a machine following a script, but one that knows when the script has gone wrong and fixes it unprompted.

The methodology earns every bit of the conclusion. The ex vivo variability is real, the test organs were unseen, and the ablations show the headline result is no fluke: strip out the corrective language and recovery-scenario success falls from 100% to 66.7%; an end-to-end model with no hierarchy lands at 33.3%. The team even handed the planning job to GPT-4o and showed it stumbled on basic surgical sequencing — a disciplined, almost generous negative result that quietly refuses the easy story that a general model can simply be pointed at an OR. This is careful work by people who plainly respect both the robotics and the surgery, and it reads that way.

There are three things a working surgeon should notice that follow directly from the results — not to temper the achievement, but to see exactly how good it is.

First, the system has learned to execute the clipping-and-cutting phase, and it does so beautifully — what it has not learned is to perform a cholecystectomy. The phase begins after the critical view of safety has been established, by a human, off-camera. The decision that prevents bile duct injury still belongs to the surgeon. That the team automated the motor sequence downstream of that judgment this cleanly is the accomplishment; the judgment is the next summit, and naming it is no knock on this one.

Second, the 100% is real, and it is young. Eight cases on pre-dissected ducts in an open dome, without bleeding, smoke, or respiratory excursion, is a forgiving bench — and exactly the right place to prove a concept first. What earns respect is that the concept held across anatomy the system had never seen. How it behaves when the field turns hostile is the work ahead, and the authors say so plainly rather than papering over it.

Third, "comparable to an expert surgeon" is the honest framing, and the team holds it with real grace. On a single gallbladder, SRT-H drew smoother, shorter trajectories than the surgeon, who was in turn substantially faster. The paper declines to claim superiority and calls the sample insufficient — restraint that does the authors credit, and that deserves to survive into a trade press far too eager to crown a winner.

The bottom line: this is the most convincing demonstration of step-level surgical autonomy to date, and it earns that standing the hard way — by throwing out the fixtures and markers that made every prior result easy to wave off. It sits at LoA IV, autonomous execution under a surgeon's eye, for a single phase of a common operation, on porcine tissue, with wrist cameras still to be shrunk for laparoscopic use. Every one of those is a tractable engineering problem, not a wall — which is precisely why this is exciting. The field's question has quietly changed from "can a learned system do this at all?" to "how far can we take it?" That is a far better question to be asking, and we owe this group for getting us to it.

The harder problem, as ever, is not severing a clean cystic duct on a quiet bench. It is knowing, in a fibrotic Calot's triangle full of blood, which tubular structure to divide — and for the first time, getting there looks like a question of when, not whether.

Theator

The obvious thing to say about Theator is that it records operations and lets surgeons review them — an indexed, searchable film library of one's own cases. That is real and useful, but it is not why the company is worth watching. The reason is that Theator has quietly repositioned itself from a review tool into a documentation tool, and documentation is where surgical video stops being optional and starts touching the medico-legal and billing infrastructure of the hospital. The platform now describes itself as turning surgical video into structured clinical data across nine specialties and 180-plus procedure types, with the automatically generated operative report as the headline output. (theator.io)

That shift matters structurally. A highlight reel is a convenience a surgeon can take or leave. An operative note is a legal document, a billing instrument, and the permanent record of what was done. If an AI can produce that note from the video, it inserts itself into a workflow every surgeon is obligated to complete and most resent completing. The wedge is administrative burden, not clinical insight — and burden sells into a hospital far more reliably than the promise of better decisions.

The evidence is better than the marketing. In a Mayo Clinic study in the Journal of the American College of Surgeons (Khanna et al., May 2025), AI-generated operative reports for robotic radical prostatectomy were compared against surgeon-written ones, with expert video review as ground truth. Across 158 cases, 53.2% of the surgeon-written reports contained at least one discrepancy with what the video showed. (journals.lww.com) The human note is already unreliable.

But hold the line. The system has learned to transcribe the steps it can see, not to author the narrative a surgeon is accountable for. One procedure, one centre, scored on pre-specified steps. The note's hardest content is what the camera misses: why a plane was abandoned, what the frozen section changed, the judgment that converted the case.

What confirms the thesis: notes surgeons sign without rewriting, across open cases too, and a regulatory posture that survives the first contested note in a deposition. What kills it: surgeons treating the output as a draft they reconstruct anyway. Theator has raised $42.5M, with no round since the 2022 Series A extension — a runway question against a long sales cycle. (evtoday.com)

The company that owns the operative note owns the most defensible position in surgical AI. Whether Theator holds it depends less on the model than on whether a surgeon will put their name to what it writes.