(PRECISION WITHOUT CONTROL)
Autonomous Orthopedics Project
Mechanical safety depends on design, materials, and assembly
Imagine that you are sitting on an airplane with your family, leaving for a vacation. You look through the window at the wing. Now imagine being told that the fasteners securing that wing to the fuselage were installed without a measured specification. The mechanic tightened each one until it felt right, relying on experience, sound, resistance, and instinct. There is no verified endpoint and no record of the forces used.
The mechanic may be exceptionally skilled. You would still want to leave the airplane. Aviation does not ask passengers to trust a mechanic’s tacit, tactile skill when a measurable assembly process can take its place.
Every mature mechanical industry depends on three disciplines: design, materials, and assembly. A sound design made from the right materials can still fail if its parts are joined incorrectly. Aviation and automotive manufacturing therefore extend standardization through the final act: assembly. They specify how parts are fitted, how fasteners are tightened, what limits apply, how the result is verified, and what record remains.
Now walk into an operating room. Total hip replacement is a mechanical assembly performed inside the human body. Its implants are the products of exacting engineering. Their alloys, dimensions, surface textures, tolerances, sterilization, and design testing are tightly controlled. Then those standardized parts reach the patient, and the discipline changes.
The surgeon prepares a cavity in living bone, chooses an implant size, and press-fits a metal component into place. The final fit depends on patient-specific bone geometry, stiffness, friction, preparation, alignment, and the history of every load already applied. Yet the surgeon must infer that hidden mechanical state from resistance, sound, vibration, advancement, recoil, and feel. The surgeon decides how hard to strike, whether the implant is gaining useful fixation, and when to stop.
The loads can reach kilonewton scale. The endpoint is not directly measured.
This is not surgical whim, and it is not a failure of skill. It is highly developed sensorimotor judgment forced to compensate for missing instruments. The surgeon is being asked to control a condition that the tools cannot reveal. The method varies because the information remains trapped in individual surgeons’ hands.
The contradiction becomes sharper when robotics enters the room. A three-dimensional plan may locate an implant to a fraction of a millimeter. Cameras track the instruments. A robotic arm guides position. Then the surgeon lifts a four-pound mallet and applies force to an interface the robot cannot interpret.
The technological timeline makes the contradiction almost absurd. Vacuum tubes gave way to transistors, integrated circuits, personal computers, navigation, robotics, and artificial intelligence. Yet the decisive act of assembly still returns to the mallet—a tool whose basic principle is often dated, rhetorically, to roughly 200,000 BC. The exact date is not the point. The discontinuity is: twenty-first-century computation ends in Paleolithic force delivery.
The next blow may improve fixation. It may add almost nothing. It may create microdamage or a fracture that no one can see. Too little fixation can end in motion and aseptic loosening. Too much interference, force, energy, or continued impaction can strain or break the bone. The computer knows where the implant should go. It does not know what the bone is experiencing now.
Total hip replacement may be the only mature, high-force mechanical procedure whose final assembly still lacks a measured control loop. What would be unacceptable in the wing outside the window remains routine inside the patient.
In Foundation Design: Principles and Practices, professor and author Donald P. Coduto teaches engineers to “maintain the same level of precision throughout the different levels of the project”—in other words, to carry sound judgment from calculation through construction. In plain language, the admonition is: do not measure with a micrometer, mark with chalk, then cut with an axe.
The warning is not an argument against precision. It is an argument for consistency. Precision in analysis is of little value if the final act of construction is uncontrolled. Press-fit hip replacement has become the medical counterpart of Coduto’s warning. We measure position and alignment with extraordinary accuracy, then apply kilonewton-scale loads without a reliable account of what those loads are or have done to the implant or the bone.
Feel is valuable, but it is private. It cannot be calibrated across surgeons, audited after a failure, or accumulated into a common body of knowledge. When the operation ends, most of its mechanical history disappears.
Autonomous Orthopedics begins with a simple proposition. Measure the physical act. Estimate the state of the bone and implant. Predict the consequence of the next action. Keep the action inside a safe boundary. Verify the result. Once that loop exists, force can become a controlled variable rather than an event reconstructed from memory.
Surgeons are expected to control conditions their instruments cannot reveal.
A press-fit implant must be tight enough to remain stable while bone grows into it. Too little engagement can lead to motion, subsidence, and loosening. Too much interference or continued impaction can injure the bone or fracture it. The safe region lies between those outcomes, but it changes with the patient, the implant, the quality of the bone, preparation, the direction of loading, and the history of every preceding strike.
The surgeon is asked to find that region without a direct measurement of it. How much bone should be removed? How hard should the surgeon strike? When should impaction stop? Has the last strike increased stability, or has the interface reached the point of failure? Should the next strike be smaller, redirected, replaced by vibration, or omitted?
The standard tools answer none of these questions directly. Sophisticated systems can see and guide implant position. A mallet or automated impactor can deliver force. The surgeon’s senses must bridge the gap between them. Even the force that reaches the implant is not the same as the force applied to the impactor. The impactor bends, absorbs, and redirects energy. Sensor location changes the measured force. A precise readout in the wrong place can be another form of false precision.
Before surgery, visual templating is the surgeon’s principal quantitative guide to implant size. But the template is a geometric estimate drawn from an image; press fit is a mechanical condition that emerges inside living bone. An X-ray may indicate a 54-millimeter acetabular cup. In the operating room, the reamer’s chatter and resistance may tell the surgeon that 54 feels too small. The surgeon must then decide whether to stop or ream farther for a 56-millimeter cup. There is no quantitative instrument to reconcile the visual plan with the tactile signal.
That is not a routine change of mind. Reaming removes living bone and cannot be undone. If the tactile impression is right, stopping at 54 may leave inadequate fixation and invite loosening. If it is wrong, proceeding to 56 may sacrifice the dense peripheral bone that provides fixation and move the patient toward excessive strain or fracture. The X-ray and the hand are giving different answers, and the surgeon must choose without being able to see the mechanical state. That is the anxiety at the operating table: one path risks too little fixation; the other risks taking away bone that cannot be put back.
The same conflict appears on the femoral side. Even experienced surgeons may select a stem one adjacent size—about one millimeter—above or below their own preoperative template. One millimeter sounds trivial. Mechanically, it can move the construct in opposite directions. An undersized stem can lose stability; an oversized stem can raise cortical strain and fracture risk.
Loosening and periprosthetic fracture together account for nearly half of the recorded reasons for revision across large hip registries. They have many causes, and no honest analysis should attribute all of them to force. Yet they mark the two ends of the same mechanical problem: too little useful fixation and too much damage. The portion that can be controlled has never been measured well enough to know its true size.
This is why the tooling deficit matters more than the familiar debate over surgeon variation or experience. The surgeon has not failed to use the tool. The tool has failed to reveal the state the surgeon is expected to control.
The skill is real precisely because it has been difficult to describe.
Michael Polanyi was a distinguished physical chemist who became one of the twentieth century’s influential philosophers of science. In The Tacit Dimension, he began with a deceptively simple observation: we can know more than we can tell. He gave a name to knowledge that is real and useful but cannot be fully reduced to instructions: tacit knowledge.
Orthopedic surgery offers a physical example. An experienced surgeon may hear the cup change pitch or feel a broach change resistance and know that the interface has entered a new state. Ask for the complete rule, and words become inadequate. The judgment lives partly in the surgeon’s hand—in muscle memory.
That knowledge is earned through repetition. It is also difficult to transmit. A trainee can watch the operation and learn its sequence. The trainee cannot simply download the expert’s internal calibration of sound, resistance, recoil, advancement, and danger. Surgeons who perform the procedure infrequently may never collect the same range of physical experience. Even high-volume experts cannot feel a crack that has not yet declared itself or see the full stress field inside the bone.
Hans Moravec is a pioneering roboticist and artificial-intelligence researcher at Carnegie Mellon University’s Robotics Institute. The observation associated with his name—Moravec’s paradox—is that machines often master abstract calculation and symbolic reasoning before they master the low-level perceptual and sensorimotor abilities that humans perform almost without thought. What feels easy to us may be the harder engineering problem.
Moravec’s paradox helps explain why press-fit arthroplasty has survived the computer revolution. Reading a ten-thousand-page report is easy for a modern model. Understanding changing contact conditions inside a human femur or acetabulum while deciding whether another impact is safe is hard. Cutting, reaming, broaching, and impaction are not merely movements to be copied. They are fast exchanges between tool, implant, and living material whose state is only partly visible.
Polanyi explains why the skill is hard to state. Moravec explains why it is hard to automate. Together they define the frontier: capture the physical signals beneath tacit expertise, convert them into a measurable state, and make that state transferable without pretending that a surgeon’s judgment can be replaced by a paragraph or a generic algorithm.
That difficulty is the opportunity. The operating room already produces the information. Force, displacement, vibration, sound, recoil, motor current, and motion all change as the interface changes. The surgeon senses a compressed version of those signals. A properly instrumented system can record them together, compare them with physical ground truth, and determine which patterns actually predict fixation and damage.
The goal is not to turn intuition into a paragraph. It is to turn its physical sources into data. Some expert habits will prove meaningful. Others may prove to be noise, tradition, or compensation for poor tools. Measurement allows us to tell the difference.
More information does not help if the surgeon must carry all of it.
Surgical technology has often responded to uncertainty by adding another screen and more information for the surgeon to process. The surgeon receives a plan, an image, a registration check, an alignment value, a warning, and a new workflow. Each addition may be reasonable. Together they ask one human brain to absorb more information while performing the operation.
The better design principle comes from the distinction between central cognition and extended cognition. Central cognition asks the brain to acquire, process, store, and use nearly all relevant information. It must continually construct and update an internal model of the external world. Extended cognition distributes part of that work into the body, the tool, or the environment, allowing the surrounding system to carry information in a form that can be used directly.
The Martian is a metaphor for the first approach. Imagine a creature with an enormous central brain, able to watch every number, interpret every waveform, and reconcile every signal in real time. Much of contemporary surgical technology quietly asks the surgeon to become that Martian. It makes the machine more informative by making the human responsible for more computation.
The spider demonstrates the alternative. A spider does not rely on its brain alone. Its web is both a sensing surface and an external memory. The spider feels vibrations in the threads of its web, locates activity through the pattern of tension, and can alter the web after a productive capture by tugging and tightening particular threads. The tightened region becomes more sensitive to future vibrations.
If the largest prey has repeatedly arrived from the southwest side of the web, the spider need not preserve that fact as an abstract map inside its head. By changing the tension of those threads, it writes the information into the environment. The web now helps remember where valuable prey came from and helps detect the next event. The spider-and-web system can therefore do more than the spider’s central nervous system could do alone. That is extended cognition in physical form.
The operating room does not need a Martian. It needs the logic of the spider. The instrument should perform the fast local work: measure the interaction, recognize a state change, remember what the preceding action produced, limit an unsafe action, and present only the decision that deserves the surgeon’s attention. The surgeon remains responsible for purpose, context, exceptions, and authority. The tool and environment carry more of the sensing, memory, and control.
The surgeon should see a concise answer. The implant is still gaining useful fixation. The estimate is uncertain. The next action should be smaller. Loading has become asymmetric. Stop. The underlying system may be technically sophisticated, but the experience of the surgeon at the operating table should become simpler.
This is where much of today’s robotic complexity misses the point. A machine can improve the accuracy of a plan and still fail to change the outcome that matters. Geometry is necessary. It is not the whole operation. If a costly system leaves the decisive mechanical uncertainty untouched, it has automated the perimeter while the surgeon still manages the center by feel.
The crowded market is in language and services while the harder frontier remains physical.
Most of the current excitement in artificial intelligence concerns agents that read, write, search, schedule, recommend, and transact. These systems can be useful and commercially important. Their world is made of representations. When they fail, a sentence can be corrected, a draft can be restored, or a transaction can be stopped.
Physical AI acts on matter. It must infer a hidden state from imperfect signals, move through imperfect hardware, and remain safe while the world changes under load. It cannot ask the bone to return to the previous version. A fractured acetabulum has no undo command.
Orthopedics therefore offers Big Tech something rare: a bounded problem that is difficult for the right reasons. The system must interpret several signals at once while anatomy changes from patient to patient and the action carries kilonewton-scale consequence. Ground truth is hard to obtain, which makes each well-characterized procedure unusually valuable. Success can be measured in mechanics before it is measured in marketing.
AO is also a very different animal from conventional surgical haptics. Haptic systems developed for general and soft-tissue surgery commonly operate in a world of delicate interactions measured in single-digit newtons, often around five to ten newtons. Press-fit hip surgery can produce ten to fifteen kilonewtons. That is a force discontinuity of roughly three orders of magnitude.
This is not conventional haptics with the gain turned up. At kilonewton scale, the structure of the instrument, its compliance and damping, the direction and duration of each pulse, the location of the sensor, implant advancement, changing friction, local bone strain, and emergency stopping all become part of the measurement problem. AO is not trying merely to reproduce a sensation at a console. It is trying to estimate and control the hidden mechanical state of a bone-implant interface before energy becomes injury.
The valuable asset will not be a general model dropped into an operating room. It will be a system that can connect a physical action to the state it produced and the outcome that followed. That requires instruments, biomechanics, control theory, clinical judgment, careful validation, and disciplined model governance. It is less fashionable than another digital agent. It is also more defensible.
A company that solves this problem will have done more than automate a surgical task. It will have learned how to build intelligence that makes high-load contact with the human body and the physical world, knows when it is uncertain, and stops before uncertainty becomes injury.
AO is an architecture that can live inside many instruments and robots.
Autonomous Orthopedics is not a proposal for one more large robot. It is a control layer for force-mediated surgery. It can begin in a handheld instrument, a powered inserter, or a robotic end effector. The embodiment may change. The logic does not.
The loop begins with the interaction itself. Instruments capture force, motion, displacement, vibration, sound, electrical behavior, and the timing of each action. A patient-specific model estimates what cannot be seen: contact, stiffness, seating, stability, remaining margin, and proximity to damage. The system predicts what another action is likely to do. A safety layer narrows the available choices. The surgeon authorizes the level of assistance. The instrument acts, and the result is measured again.
The first version need not strike the mallet or apply the force itself. It can simply tell the truth about what happened. The next can advise when to continue or stop. Later versions can limit an unsafe command, adjust a powered pulse, or perform a tightly bounded task under supervision. This is the path to full autonomy, a progression that appears increasingly inevitable. Automation is a permission earned by evidence, not a word placed in the product name.
AO’s earlier systems anticipated parts of this loop. Electronic Signature Sizing of Bone sought information in the electrical behavior of powered tools. The Automated Prosthesis Installation Method addressed controlled impaction. The Vibratory Implantation and Osteointegration Instrument explored vibration as an alternative way to advance an implant. Related concepts examined screw insertion, torque, and mechanical feedback. Those systems were largely analog or electromechanical. They were pieces of an architecture waiting for better sensors, faster chips, and stronger models.
The architecture now extends beyond the instrument. Each case can produce a mechanical record. Within the operation, new measurements update the estimate for that patient. Across operations, de-identified data can improve candidate models after those models are tested, locked, and approved. The system will learn from every case without improvising on the patient. Just as autonomous vehicles became possible through millions of miles of data and training, Autonomous Orthopedics can improve through the disciplined accumulation of mechanical cases. Over time, it should learn patterns that no individual surgeon, however experienced, could encounter or retain—and may ultimately outperform even elite tactile judgment in the bounded tasks it has been trained and validated to control.
Standardization does not mean using the same force on every patient .
A fixed force would be simple and wrong. Bone is not manufactured stock, nor is it a uniform grade of wood on a carpenter’s bench. The correct action depends on anatomy, bone density, implant design, preparation, friction, alignment, and the response to the preceding action. The patient-specific target must change. The method for finding it should not.
That distinction matters because the industry has already begun to automate the strike. DePuy Synthes’ KINCISE Surgical Automated System is described as delivering constant and consistent energy. Zimmer Biomet’s HAMMR Automated Hip Surgical Impactor System offers three selectable energy levels and consistent output within the selected setting. These devices can reduce manual variation, off-axis blows, and surgeon fatigue. Those are legitimate advances.
But consistent input is not the same as closed-loop control. Energy is not force, and neither a constant input nor a choice among a few settings reveals what that input has done to this patient’s bone. If the tissue response is not measured, a powered impactor can repeat an inappropriate action with greater consistency. The danger is not automation itself. The danger is mistaking repeatability for appropriateness.
Standardization should not prescribe the same blow. It should standardize how the right action is determined for each patient.
A useful standard would define where force is measured, how instruments are calibrated, how signals are synchronized, which mechanical states are estimated, how uncertainty is reported, what safety boundaries apply, when the system must abstain, how an endpoint is verified, and what record remains after the operation. It would also define the evidence required before a device may move from measurement to advice, from advice to limitation, and from limitation to action.
This standard would connect domains that now sit apart. Implant designers know the component. Surgeons know the procedure. Robotic systems know geometry. Sensors know a local signal. Regulators know the device claim. None alone holds the full account of how an implant became fixed to this patient’s bone.
Once the mechanical state has a common language, instruments and implants can be designed for observability and control. Training can compare decisions rather than reputations. Failures can be reconstructed from an interaction record rather than assigned through inference. Manufacturers can learn how specific designs behave across bone conditions. Robots can control something more consequential than pose and alignment.
The platform value follows from the standard. The first company to define the mechanical state, the measurement rules, and the interfaces will influence how every compatible instrument reports and acts. The data will deepen the standard, and the standard will make the data more valuable. That cycle is far harder to copy than a sensor.
The old ideas have met the technology needed to combine them.
The mallet did not suddenly become dangerous, and tacit skill did not suddenly become imperfect. What changed is our ability to observe the event. Compact sensors can now survive near the point of action. Edge processors can read multiple signals with little delay. Modern actuators can shape a pulse rather than merely deliver one. Physics-based models and learning systems can work together at higher speeds. Simulation can expose a controller to rare conditions before a patient does. Earlier analog and electromechanical systems described ingenious ways to interrogate the implant-bone interface, but the resources required to unite those measurements in a practical real-time control system did not yet exist. The ideas were not wrong. They were early.
Technological transitions often look difficult until the enabling system proves itself. Propeller aircraft did not become useless when the jet engine arrived, but they no longer defined the frontier of speed, altitude, and long-range travel. Once force-aware implantation can measure the interface, adapt the action, and verify the endpoint, unmeasured impaction will not disappear overnight. It will simply stop defining the frontier of responsible orthopedic assembly.
The commercial timing has changed as well. MedTech companies have built robotic ecosystems around geometry. Big Tech companies are searching for credible uses of Physical AI. Hospitals face pressure to make outcomes more reproducible. Surgeons are increasingly asked to adopt complex systems whose clinical advantage may be modest or difficult to see. The market is ready for a simpler proposition: solve an important uncertainty that the current tools do not measure.
There is also an uncomfortable incentive problem. Revision surgery is more complex and more expensive than the first operation. Device companies participate in both markets, while responsibility for failure can disperse among the patient, surgeon, hospital, and implant. This does not mean manufacturers seek to benefit from failure. It means the existing market does not automatically reward a cross-platform system built to prevent mechanical failure. The company that develops the first credible platform will not merely lead a product category. It may define the standard by which every later implant and instrument is judged—and force the rest of the market to build around it.
That choice will become harder to postpone. Once a competitor can document the force history, verify the endpoint, and connect the record to outcomes, an unmeasured procedure—whether performed with a mallet or an automated impactor—will begin to look less like tradition and more like avoidable risk.
The force domain will become a platform whether incumbents lead it or inherit it.
For MedTech, the threat is architectural. A company can own the implant, the instrument, and the robot yet still depend on another company for the intelligence that decides how hard to act and when to stop. If the mechanical control layer becomes the source of evidence and improvement, the hardware beneath it risks becoming interchangeable—a commodity. The data and the mechanical platform become the differentiators.
For Big Tech, the opportunity is equally clear. The market for language agents is crowded. AO offers a difficult Physical AI problem with direct human value and a measurable path from sensing to safe action.
The company that helps solve it gains more than a healthcare application. It gains a working model for intelligence in high-force contact with the human body and the physical world.
For venture capital, the timing is unusual. The first experiment is modest compared with the size of the eventual platform. It can answer a decisive question before large capital is committed: can the relevant mechanical state be observed well enough to improve the decision? If the answer is no, the program stops early. If the answer is yes, the value begins to compound through data, validation, workflow, intellectual property, and standards.
Late entrants will not be missing only a product. They will be missing the force histories, failure boundaries, implant-specific behavior, operating-room integrations, and trust accumulated by the first serious system. By the time the category has a familiar name, its architecture may already be embedded.
Build the smallest system that can prove the hidden state is real and useful.
The first project should not attempt an autonomous operating room. It should instrument one bounded press-fit task and ask a short sequence of questions. Can practical signals reveal seating, fixation gain, and proximity to damage? Can the estimate survive changes in bone, implant size, alignment, and instrument? Does acting on the estimate improve fixation or reduce damage compared with expert technique and simple fixed settings? Is the benefit large enough to justify the added system?
The work should begin in controlled models, move to cadaveric validation, and enter clinical use only when each preceding step holds. Simple methods should be tested before complex ones. If a force threshold performs as well as a sophisticated model, use the threshold. If the state cannot be observed, redesign the instrument or stop. Complexity must earn its place.
The first commercial product can be a recorder and endpoint adviser. The surgeon remains in control. The value is immediate: a calibrated account of the action, a clearer stopping decision, and a record that can be studied. Greater authority follows only after the system proves that it can predict, constrain, and verify.
This program requires a small coalition: an orthopedic and biomechanical team to define the problem and ground truth; a controls and robotics team to build the sensing and safety system; a MedTech partner with implants, instruments, and a path into the operating room; a Physical AI partner capable of multimodal models and edge inference; and capital tied to measurable milestones rather than the appearance of scale.
Physical orthopedics cannot be financed by pretending it is software. A digital service can release an inexpensive minimum viable product, collect feedback, and revise the code. A powered hip inserter with a safety-critical control system must be engineered, built, tested, and documented before it touches a patient. You cannot create a minimum viable intelligent hip installer and repair it with the next software update.
A serious first AO development program on the order of five million dollars may look large beside a software prototype. It is modest beside the value of defining a new surgical control layer. The capital purchases the evidence: sensorized instruments, prototypes, controlled bench work, cadaveric validation, synchronized multimodal data, state-estimation and control methods, human-factors testing, and regulatory-quality documentation.
That capital should not be blind. It should be staged against falsifiable milestones. Can the system observe fixation gain? Can it distinguish productive advancement from diminishing return? Can it identify asymmetric loading or proximity to damage? Can a simple rule perform as well as a complex model? Does the information improve the surgeon’s decision? A failed milestone should stop or redirect the program. A successful milestone should release the next stage.
Demanding complete validation before financing the work that creates validation is a category error. Every major physical invention begins as an unconventional idea that must be made real before the conventional evidence exists. The earliest capital accepts technical uncertainty in exchange for architectural ownership. Investors who wait until the control signal, workflow, regulatory path, and clinical value are obvious will face less uncertainty—and a far higher price, after the defining data and partnerships have already been claimed.
The Autonomous Orthopedics architecture is defined. Its analog and electromechanical predecessors exist, its patent history exists, and the control logic is now technically credible. The first integrated clinical system still must be built and validated. That is not an ambiguity hidden behind the word autonomous. It is the present opportunity for the MedTech company, technology company, academic group, or investor prepared to help establish the field.
The surgeon is not the weak link, and the mallet is not the central enemy. The failure lies in an incomplete mechanical system that asks a human being to apply large forces without revealing the condition those forces create. AO is intended to complete that system: to preserve human judgment while giving it a measurable state, a patient-specific boundary, and a verified endpoint.