Surgical robotics
In some operations the robot is already the normal way of working. It gains millimetres and minutes, and that is measured. That those millimetres turn into a benefit for the patient is still an undemonstrated claim.
It is worth starting by dismantling the wrong image. In neurosurgery, the robot does not operate. It does not decide, it does not dissect, it replaces no one. It is an arm that holds, with a steadiness the human hand does not have, a trajectory the surgeon planned, checked and approved. Its virtue is of an uncinematic kind: it does the same thing a hundred times in exactly the same way, without tremor and without fatigue.
That is why its natural place is where geometry is everything: implanting electrodes to study an epilepsy, stereotactic biopsies, placing deep brain stimulation electrodes, pedicle screws in the spine. There it is no longer a promise. It is how the work is done.
What is measured
The figures behind that adoption are mostly about process and safety, and they are solid. In stereoelectroencephalography, a classic series of five hundred procedures reports a major complication rate below three per cent, with a clear improvement in entry accuracy on moving from the traditional workflow to the robot-assisted one1. In brain biopsy, a series of more than nine hundred patients across twenty-one years shows that only about one biopsy in sixty fails to reach a diagnosis, and that this figure held steady for two decades2. In paediatric deep brain stimulation under general anaesthesia, the distance between planned point and actual electrode position stays under two millimetres, with no major perioperative complications3.
In the spine the evidence is different and stronger, because there are randomised trials. One of them, with more than a thousand screws, shows a clearly higher proportion of screws in perfect position with robotic assistance than with fluoroscopy, and a reduction of roughly two-thirds in the radiation dose the surgeon receives4. That last figure should be read carefully: in the same trial, cumulative radiation emission time was longer with the robot. The surgeon is irradiated less because they step away and because the workflow changes, not because fewer rays are emitted.
The two uncomfortable findings
The first. When robot and classical stereotactic frame are compared head to head in stereoelectroencephalography, the best available meta-analysis finds no accuracy advantage: differences across the four measured error types are statistically indistinguishable from zero5. What the robot does gain, consistently, is time: more than half an hour per operation and several minutes per electrode. There is also a meta-regression over more than six thousand trajectories that does attribute to the robot an error reduction of nearly a millimetre6 — but that is an adjusted association across heterogeneous studies, not a randomised comparison. Reporting one of the two numbers and omitting the other would be advertising.
The second, and the weightier. In the meta-analysis of nine randomised spine trials, with better screw accuracy and fewer adjacent facet violations, the outcomes that matter to the patient — pain, disability, days in hospital — came out equivalent7. Nobody has yet shown fewer reoperations, less neurological deficit or better seizure control attributable to the robot. The improvement is real and it is procedural. That it translates into a better life is a reasonable, unproven hypothesis.
Three cautions of method
"The robot" does not exist. That same meta-analysis shows the accuracy advantage concentrated in one specific system and absent with another7. Generalising from a device to the whole category is a methodological error, and it is exactly the one promotional material makes.
The learning curve is real and underestimated. Studies that measured it place the plateau between a few and several dozen cases8. And the published series come almost always from centres that have already climbed it: what one reads is the best case, not the average case.
Industry bias is structural. The most cited review of learning curves in robotic spine surgery is signed, among others, by the inventor of the device studied, who declares royalties, consultancy and shares in the manufacturer8. That does not invalidate the work. It does oblige us to read any claim of superiority knowing who signs it. And in several of the available papers the conflict-of-interest statement cannot even be accessed, which is itself a datum about the field's transparency.
The horizon, and where it really is
There is talk of teleoperated endovascular robots able to navigate to a cerebral artery under remote guidance, and of microrobotics able to work through ever narrower corridors. Both exist. The magnetic manipulation platform published in Science Robotics navigated anatomical models and animals9: zero patients. And a recent review of the state of the art warns that the microscale poses its own fabrication and control problems that are not solved by simply miniaturising what already works10.
One issue remains, and it is not technical. The word "autonomous" circulates far too freely. Today, in neurosurgery, nothing is autonomous: there is a teleoperated positioner under continuous human control. Saying so precisely matters, because on that precision depends who answers when something goes wrong. The US regulator is explicit: it neither supervises nor accredits physician training on these devices, and places that responsibility with the manufacturer, the professionals and the institutions11. There, and not in the robotics, is where the hard problem lies.