The Surgeon’s New Hands

There are two very different machines a person can mean when they say “surgical robot.” The first steadies a surgeon’s hands: it scales motion, filters tremor, and rotates a wristed instrument inside a body cavity through an incision the width of a pencil — but every millimeter of that motion originates in a human hand on a console a few feet away. The second machine knows where to cut. It holds a plan, perceives the tissue in front of it, and executes a step of the procedure while the surgeon watches. Almost every robot in clinical use today is the first kind. Almost every dollar of ambition in the field is aimed at the second.

The distance between those two machines is the subject of this essay. I write it as someone who works inside this industry rather than above it, so take what follows as informed postulation about a trajectory, not prophecy. The trajectory itself, though, is unusually legible right now. The taxonomy that describes it is published and cited, the companies climbing it report their numbers every quarter, the startups attacking its hardest rungs raise rounds that hit the wire with dates and dollar figures attached, and — for the first time — academic labs have shown a robot performing a full phase of a real procedure on living-adjacent tissue with no hand on the controls. The pieces are on the table. What remains is to read them in the right order.

Part I

The autonomy ladder

The most useful frame for the whole field is borrowed, deliberately, from cars. In 2017 a group of medical-robotics researchers led by Guang-Zhong Yang published a short paper in Science Robotics proposing a six-level scale of surgical autonomy, explicitly modeled on the SAE levels that the automotive world uses to talk about self-driving (Yang et al., 2017). It has become the field’s shared vocabulary, and it is worth stating precisely because almost every confusion in the popular coverage of “robot surgeons” comes from sliding between its rungs.

• Level 0 — No autonomy. The robot does exactly what the surgeon commands. Motion scaling, tremor filtering, and tool tracking all live here, because the output is still the surgeon’s intended motion, merely cleaned up. This is teleoperation.
• Level 1 — Robot assistance. The robot provides continuous physical or cognitive guidance while the human stays in control throughout — a virtual fixture, an active constraint, a “wall” the instrument cannot cross.
• Level 2 — Task autonomy. The robot autonomously executes a discrete, surgeon-initiated task — a suture line, a bone cut — under supervision. The human approves at the start; the robot carries it out.
• Level 3 — Conditional autonomy. The robot generates strategy, not just motion: it plans a task and adapts the plan during execution, while the surgeon selects, approves, and supervises.
• Level 4 — High autonomy. The robot makes clinical decisions and executes them under the oversight of a qualified surgeon.
• Level 5 — Full autonomy. A robot performs the whole operation with no human input. The “robotic surgeon” of the headlines.

Two things make this ladder the right backbone for everything that follows. The first is that we can say, with citation rather than vibe, exactly where the commercial field sits. A 2024 systematic review in npj Digital Medicine examined every FDA-cleared surgical robot — forty-nine systems — and classified them against this scale. Eighty-six percent are at Level 1. The single highest level reached by any cleared system is Level 3, attained by a small number of orthopedic systems that mill bone against a patient-specific plan (Han et al., 2024). Nothing on the market is at Level 4 or 5. The flagship soft-tissue robots that define the public image of robotic surgery — the ones doing prostatectomies and hysterectomies by the millions — are Level 1. They are, in the strict sense, not autonomous at all.

The second reason the ladder matters is that the rungs are not evenly spaced, and the gap that matters is the one between rigid anatomy and soft tissue. A femur does not move while you cut it. Bowel, fat, and blood vessels deform continuously, bleed, and slide out of plane; a plan computed a second ago is wrong now. This is why the only fielded systems above Level 1 work on bone, and why a robot that autonomously sutures soft tissue in a living body remains, in 2026, a research result rather than a product. Surgery is the hardest possible domain for the autonomy ladder for reasons that compound: the cost of an error is irreversible, the workspace is deformable and occluded, the regulator sits in the room in the form of a clearance file and a liability regime, and the demonstrations you would need to train a policy are scarce, proprietary, and recorded inside an environment that punishes experimentation.

It is worth situating this against the rest of robotics, because surgery is both behind and ahead. Behind, because warehouse arms and even cars have reached commercial Level 2–4 autonomy in their domains while surgery has not. Ahead, because the value density of a single surgical task is so high that the field can fund a decade of research toward automating a few minutes of work. The broader robotics world has, over the last three years, gone through a methodological turn that surgery is now importing wholesale: the shift from hand-coded control to learning from demonstration, and then to vision-language-action models — large neural policies, trained on demonstrations and web-scale data, that map what a robot sees and is told directly onto what it does. Google DeepMind’s RT-2 in 2023, then Physical Intelligence’s π₀ in 2024, showed that a single learned model could generalize across tasks and embodiments. The reason this matters for the operating room is not analogy. It is personnel: the same researchers are now publishing surgical results. Hold that thread; Part IV pulls on it.

Part II

The landscape today

If almost everything in clinical use is Level 1, then a survey of the leaders is really a survey of two things — who controls the installed base of teleoperated systems, and who has pushed furthest up the bone-cutting branch of the ladder. Here are the five companies that define the field, each read against the autonomy scale and against the question that turns out to matter most by the end of this essay: what data does the machine leave behind?

Intuitive Surgical is the incumbent by a margin that is hard to overstate. Its da Vinci system — now in its fifth generation, cleared by the FDA in March 2024 — had an installed base of 11,106 systems worldwide at the end of 2025, up twelve percent on the year, alongside 954 of its Ion robotic-bronchoscopy systems. Surgeons performed roughly 3.15 million da Vinci procedures in 2025, pushing the cumulative lifetime count toward seventeen million, and the company booked about $10 billion in revenue (Intuitive Q4/FY2025 results; figures preliminary). On the autonomy ladder, da Vinci is Level 0 — pure teleoperation. Its most notable recent addition, Force Feedback on da Vinci 5, gives surgeons a sense of touch for the first time but adds no autonomy. What it adds, as we will see, is a new stream of data.

Medtronic is the most credible challenger in soft tissue, and its big moment arrived late in 2025: on December 3, the FDA cleared its Hugo system for urologic procedures, on the strength of a 137-patient trial, with the first US commercial case performed at the Cleveland Clinic (Medtronic, Dec 2025). Hugo had carried a European CE mark since 2021 and is deployed across more than thirty countries. Architecturally it is modular — independent arm carts and an open console — but on the ladder it is Level 0, a direct teleoperated competitor to da Vinci. The contest between Hugo and da Vinci is not about autonomy. It is about breaking a near-monopoly.

Stryker sits on the other branch of the tree, and it is the most genuinely autonomous of the mainstream systems — though with an important caveat. Its Mako platform, with more than three thousand robots installed and over 1.5 million procedures performed across forty-five countries, works in orthopedics: a preoperative CT scan becomes a three-dimensional model, which becomes a patient-specific cutting plan, which the robotic arm then enforces. Inside the planned region the surgeon moves the saw freely; at the boundary the arm resists like a wall and the tool deactivates (Stryker). This is Level 1 — robot assistance through an active constraint — and it is the ceiling for mainstream commercial systems. The caveat matters enough to state plainly: Stryker is explicit that the arm moves in tandem with the surgeon’s hand and makes no autonomous motion. It is bounded, shared control, not a robot that cuts by itself. The line between “enforces a plan” and “executes a plan” is exactly the line between Level 1 and Level 2, and no cleared soft-tissue system has crossed it.

Johnson & Johnson MedTech is the giant that has taken longest to arrive, and 2025–26 is finally its inflection. Its soft-tissue robot, Ottava, distinguished by four arms integrated directly into the operating table, completed its first investigational human cases in early 2025 and was submitted to the FDA for clearance in January 2026 — not yet on the market (MassDevice). Its endoluminal system, Monarch, received clearance for robotic lung biopsy in March 2025, paired with an AI navigation update built with Nvidia and GE Healthcare. Ottava is Level 0; Monarch is Level 0–1. J&J’s bet is distribution and integration, not autonomy — but it brings the deepest pockets in medtech to the table.

The fifth slot belongs jointly to two systems that bracket the field’s two directions. CMR Surgical’s Versius, a British modular system designed around small independent arm carts, won FDA de novo authorization in October 2024 and a 510(k) clearance for its Versius Plus generation in December 2025, with a US launch planned for 2026 and more than forty thousand procedures already performed outside the US (MedTech Dive). It is Level 0 — the scrappy, portable, cost-focused challenger to the da Vinci/Hugo duopoly in soft tissue. Zimmer Biomet’s ROSA, by contrast, lives on Stryker’s branch: a robotic platform spanning knee, hip, and shoulder replacement plus neurosurgery, with roughly two thousand installations and an enhanced ROSA Knee cleared in November 2025 (Zimmer Biomet). Like Mako it is Level 1, but it reaches its plan differently — historically through intraoperative registration rather than a mandatory preoperative CT, a real technical contrast in how much imaging the autonomy depends on. Between them, Versius and ROSA show the whole shape of the market: one race to make teleoperation cheaper and more accessible, another to push bounded autonomy deeper into structured anatomy.

Step back from the table and a pattern asserts itself. The thing that separates the leaders is not kinematics — the arms are, by now, a solved and commoditizing problem. It is the installed base and the regulatory file: the thousands of placed systems, the millions of logged procedures, the trial data that turns into a clearance. Intuitive’s moat is not that da Vinci moves better than Hugo. It is that seventeen million procedures have flowed through da Vinci consoles, and every one of them left a trace. Hold that observation too. It is the hinge of the whole argument.

Part III

The open frontier

If the incumbents own teleoperation, the open frontier — where venture money and startup energy concentrate — is everything teleoperation does not yet do: reach anatomies the big systems ignore, drive the cost low enough for an outpatient surgery center, and automate the individual steps of a procedure. The market they are chasing is large and growing fast, though precisely how large depends on whom you ask: analyst estimates of the 2025 surgical-robotics market cluster between roughly $10 billion and $14 billion, with compound growth rates in the mid-teens, and a long tail of more aggressive projections that should be read skeptically (Precedence, MarketsandMarkets). Venture funding for robotics broadly crossed three billion dollars in 2024 even as deal counts fell — capital concentrating into fewer, larger rounds, a sign of a field consolidating around its more credible bets. It is more illuminating to organize the frontier not by company but by the problem each is attacking.

Cost and access. The first problem is that a da Vinci costs more than most hospitals on earth can justify, which confines robotic surgery to wealthy systems and high-volume procedures. A cluster of companies is attacking this with lighter, modular, cheaper machines aimed at ambulatory surgery centers. Switzerland’s Distalmotion raised a $150 million Series G in November 2025 for its Dexter system; CMR raised more than $200 million in April 2025; Moon Surgical, backed by Sofinnova and Nvidia’s venture arm, built its Maestro system around existing laparoscopic tools; and China’s Ronovo took a $67 million round led by J&J’s venture arm in September 2025. The thesis is unglamorous and enormous: robotic surgery has barely penetrated its addressable market, and whoever makes it affordable inherits the rest.

New anatomies. The flagship systems concentrate on the abdomen and pelvis. Whole specialties — the eye, the brain, the vasculature, the microscopic structures of reconstructive surgery — remain barely touched, and they are where some of the most ambitious startups live. The clearest example is ophthalmic surgery: ForSight Robotics, building its Oryom platform for cataract and other eye procedures, raised a $125 million Series B in June 2025, bringing its total to $195 million, with backing that includes da Vinci’s own founding figure Fred Moll (Business Wire, The Robot Report). Cataract surgery is the most common surgical procedure in the world and faces a global shortage of surgeons — exactly the conditions under which automating a delicate, repetitive, high-volume task becomes both valuable and humane. Elsewhere on the anatomical frontier: Noah Medical (robotic bronchoscopy, past five thousand procedures), Medical Microinstruments (the Symani microsurgery system, a $110 million round in 2024), Remedy Robotics (the first remotely operated endovascular procedures), and Caranx (the first FDA-authorized AI for real-time guidance of transcatheter heart-valve implantation, cleared in 2025). Each is a wager that a specialty the incumbents skipped is large enough to build a company around.

Autonomous sub-tasks and the perception layer. This is the frontier that bears most directly on the autonomy ladder, and it splits in two. One half is hardware and policy: research groups and a few companies trying to automate discrete steps — suturing, knot-tying, needle steering, biopsy. A Vanderbilt-led effort won up to $12 million from ARPA-H in September 2024 explicitly to build toward a fully autonomous surgical robot. The other half is the software that any autonomy must stand on: machines that understand a surgical scene. Caresyntax raised a $180 million Series C extension in August 2024 for its surgical-intelligence platform; Proximie and Theator built businesses on capturing and analyzing surgical video; Activ Surgical’s intelligent-light module reveals blood flow and hidden structures in real time. These companies are, in effect, assembling the eyes and the memory that an autonomous step will eventually require — and, not incidentally, accumulating the surgical-video corpora that any learned surgical policy will need to be trained on.

Adjacencies. At the edges sit the rehabilitation exoskeletons — France’s Wandercraft raised a $75 million round in mid-2025 — and the long-horizon dream of capsule robots, micro-robots, and targeted nano-delivery, which remains almost entirely academic, with no fundable startup category yet. Worth naming as the field’s true frontier, worth discounting as a near-term commercial force.

Read across these four problems and a single bottleneck keeps surfacing under different names. The cost-reducers need to make robots that more surgeons can drive. The new-anatomy companies need to learn to perform delicate tasks well enough to trust a machine with them. The autonomy and perception companies need, explicitly, surgical data at scale. The constraint, again and again, is not actuators or optics. It is validated demonstration — recordings of expert surgeons doing the task correctly, enough times, labeled well enough, that a machine can learn the task and a regulator can trust the result. Which returns us, by a different road, to the observation at the end of Part II.

Part IV

What the future holds

So where does this go? Four claims, in ascending order of consequence.

Data is the real moat

The first and least appreciated truth is that the durable advantage in this industry is not the robot. It is the corpus the robot generates. Intuitive understands this better than anyone: more than twelve thousand installed systems performing over twenty million cumulative procedures, increasingly captured as synchronized video, instrument kinematics, and — on da Vinci 5 — force data, all funneled into an analytics layer the company calls Case Insights (Intuitive). No startup corpus, and no academic dataset, comes within orders of magnitude of this. When the field eventually trains the surgical equivalent of a foundation model, the question of who owns the training data has, in large part, already been answered by a decade of teleoperation that nobody at the time framed as data collection. The teleoperated systems of today are, in retrospect, the most expensive and best-instrumented demonstration-gathering apparatus ever built. The autonomy they did not provide, they were quietly making possible.

This reframes the entire competitive map. Hugo, Ottava, and Versius are not merely fighting da Vinci for procedure share; they are fighting for the right to instrument procedures at all, because every system placed is a data tap that compounds. It also explains the strategic logic of the independent perception companies — Caresyntax, Theator, Proximie — as an attempt to build the one surgical-data corpus that is not locked inside a single robot vendor. The next decade of competition in surgical robotics will be, underneath the hardware announcements, a contest over who holds the largest, cleanest, best-labeled record of surgery actually being done.

The GPT moment for surgery

The second claim is that the methodological turn that produced large language and vision-language-action models is now reaching the operating room — and that we have already seen the first convincing demonstrations. In 2022, a team at Johns Hopkins led by Axel Krieger showed STAR, the Smart Tissue Autonomous Robot, performing autonomous laparoscopic bowel anastomosis in vivo in pigs — planning and placing sutures in deformable tissue more consistently than an expert surgeon (STAR, Science Robotics 2022). STAR was built on classical planning and control. What came next was the learning generation. In 2024, the Surgical Robot Transformer applied imitation learning to a da Vinci, teaching it tissue manipulation, needle handling, and knot-tying from demonstration (SRT, 2024). And in 2025, its successor SRT-H — a hierarchical, language-conditioned policy — performed a full phase of a cholecystectomy, autonomously clipping and cutting the cystic duct and artery, succeeding on all eight unseen ex vivo gallbladders and recovering from its own errors through a high-level policy that reasons in language (SRT-H, Science Robotics 2025). The names on those papers — Chelsea Finn among them — are the same names on the general-robotics foundation-model papers. This is not analogy; it is the same research program, pointed at tissue.

The honest framing is bounded. These are research results on animal tissue, not products, not humans, and the leap from a clean ex vivo gallbladder to a bleeding, breathing patient with anomalous anatomy is exactly the kind of leap that has humbled self-driving cars for fifteen years. A surgical foundation model — a single policy that understands many procedures, generalizes across anatomies, and proposes the next step — is plausibly the thing that carries the field from Level 1 to supervised Level 2 and 3. It is not, on any near horizon, the thing that produces a Level 5 robot operating unattended. What it unlocks first is narrow and supervised: the robot performs the suture line, the clip, the routine sub-step, and the surgeon supervises and intervenes — autonomy as an extra pair of trained hands, not a replacement for the head.

The human as teacher

The third claim is the strangest, and it generalizes a pattern visible across all of robotics. To climb the ladder, these systems must learn from demonstration — which means that, for a transitional period whose length nobody knows, the surgeon’s job quietly acquires a second nature. The surgeon is no longer only operating. The surgeon is teaching, frame by frame, every time an instrument moves under a recording system that will later be used to train a policy to do what that instrument just did. The apprentice in this arrangement is the machine, and the master, for now, is the human whose skill is being slowly transcribed into a model. It is an unusual contract, and it runs against the grain of how we imagine automation: the expert is not displaced so much as conscripted as a teacher of their own eventual stand-in.

This has consequences worth naming. It changes surgical training, because a model that has absorbed the technique of the best surgeons can, in principle, coach the median one toward that standard — compressing the long apprenticeship of surgery the way a good textbook compresses a field. It changes the geography of expertise, because a specialist’s skill, once captured in a policy, becomes deployable to a rural hospital that could never recruit that specialist — the same flattening that put a world-class chess engine on a phone. And it presses on the surgeon’s identity in a way the profession has not yet metabolized: what does it mean to spend a career mastering a craft while simultaneously, with every case, recording the demonstrations that will let a machine perform parts of it? The likely near-term answer is reassuring and the long-term answer is open: for years, the human supervises, corrects, and handles the cases the model cannot, while the model absorbs the cases it can.

The shape of the climb

Which leaves the forward claim, stated as sharply as I can make it. Medicine will be among the last domains to reach full autonomy and among the most transformed by partial autonomy. The two halves of that sentence are not in tension; they are the whole point. Full autonomy — a Level 5 robot operating a human being unattended — is held back not by any single missing technology but by the convergence of everything hard at once: deformable tissue, irreversible error, a liability regime with no answer yet to the question of who is responsible when a learned policy is wrong, a reimbursement system that pays for procedures and not for autonomy, and a trust threshold that medicine, rightly, sets higher than any other field. Those forces will set the pace, and the pace will be slow.

But they will not change the direction, and the transformation along the way will be profound long before Level 5 ever arrives. Supervised autonomy at Levels 2 and 3 — the robot performing trusted sub-steps under a watching surgeon — is within reach of today’s methods, as the 2017 paper predicted and the 2025 results begin to confirm. The inflection that decides when it arrives is not a breakthrough in hardware; the arms are good enough already. It is the corpus — the accumulated, validated record of surgery being done, which the incumbents have been gathering for a decade without calling it that, and which the startups and the academic labs are now racing to build outside the incumbents’ walls. The surgeon’s new hands already exist on a thousand consoles. What they are still waiting for is not better hands. It is the memory of every operation that came before — and the patience of a field that, more than any other, has earned the right to climb its ladder one careful rung at a time.

Sources are linked inline. Company figures are as reported in 2025–2026 filings and press releases; Intuitive’s full-year 2025 numbers are preliminary. The surgical-autonomy levels are an academic framework, not a regulatory one, and constraint-based orthopedic systems should not be read as performing autonomous cutting.