The Wrong Way to Design a Humanoid

An analysis of how biological a humanoid robot should actually be

Decorative photo, inspired by Atlas Humanoid from Boston Dynamics and edited by the author. Mehrdad Farimani, Aug. 2026, Sweden

There is something extremely seductive about a tendon-driven robotic hand.

You look at the human hand and see muscles pulling tendons, tendons routing through the wrist and fingers, lightweight distal joints, natural compliance, and an absurd amount of dexterity packed into a very small volume. Then you look at a conventional robotic hand filled with motors, gearboxes, bearings, cables, and electronics and think: nature already solved this. Why don't we just copy it?

Maybe.

But I think there is a trap hiding inside that logic.

Not because tendon-driven hands are necessarily bad. There are very legitimate engineering reasons to use tendons in robotic hands. The more important question is whether, when we copy human anatomy, we are copying a good mechanical principle or copying a mechanism whose advantages depend partly on being biological.

Those are not the same thing.

As humanoid robotics becomes a more serious engineering field, I think we need to become much more selective about where biological inspiration is useful and where it starts becoming baggage.

A human finger isn't really a conventional machine

Take your index finger.

From a robotics point of view, it is relatively easy to describe. Bones provide structure, joints constrain motion, muscles generate force, tendons transmit it, ligaments stabilize the system, nerves provide sensing and control, and skin provides friction, protection, and compliance.

It is tempting to translate this directly into engineering terms. Muscle becomes a motor, tendon becomes a cable, joint becomes a bearing, proprioception becomes an encoder, and skin becomes a tactile sensor.

Congratulations, human hand reverse-engineered.

Except we haven't really reverse-engineered it at all.

A human finger has at least three characteristics that almost none of our conventional robotic mechanisms have: it grows, it heals, and it physically adapts.

I don't think these are secondary biological features. They may be deeply connected to why human anatomy can tolerate the complexity it does in the first place.

First, it grows

A human hand is not manufactured at adult size.

It develops from a tiny structure into a full-sized hand while all of its systems develop together. Bones grow, tendons lengthen, muscles change, nerves extend, skin expands, blood vessels develop, and joint geometry changes. The final hand is not an assembly of independently manufactured components. It is the result of a long developmental process in which the components and their interfaces grow together.

As an engineering exercise, this is pretty ridiculous.

Imagine designing a robotic gearbox that starts tiny and gradually becomes several times larger while preserving useful function and alignment along the way. Most mechanical engineers would politely ask why you are doing this to them.

Biology does it routinely.

Development matters because the final anatomy we observe is partly possible because it grew into that configuration.

This is easy to forget when reverse-engineering biology. We examine the adult mechanism and treat its geometry as if it were independent from the process that produced it. But some structures may be elegant to grow while being terrible to manufacture, assemble, calibrate, inspect, or service using conventional engineering methods.

That difference alone should make us cautious about treating anatomy as a ready-made CAD file.

Then it heals

The human body is also maintaining its mechanical systems continuously.

Skin repairs itself, bone remodels, muscle recovers from loading, and connective tissues respond to repeated stress and microdamage. We tend to notice biological repair only when something goes seriously wrong, such as a cut, fracture, or major injury, but a great deal of maintenance happens below our threshold of attention.

The machine is being repaired while we are using it.

Compare that with a robot. A cable starts fraying, a bearing develops play, a seal leaks, or a gear wears. The robot does not generate new fibers around the damaged cable or rebuild the worn surface. At best it detects the problem and asks a human to fix it.

Mechanically, this difference is huge.

Biology can tolerate certain kinds of complexity partly because that complexity comes with an embedded maintenance system. Robotic mechanisms usually do not get that maintenance system for free.

So when we admire a biological mechanism, we should ask whether its reliability comes only from its geometry or partly from the fact that the system around it is constantly maintaining it.

And the hardware itself adapts

This third property may be the most interesting.

Imagine someone who works with tools, glass, masonry, climbing equipment, or rough materials every day. Their hands do not remain physically identical to the hands they started with. Repeated friction can change the skin, sustained loading can change muscle, bone remodels according to mechanical demands, and connective tissues respond over time to how the body is used.

The important point is that the person is not only becoming better at controlling the hand.

The hand itself is changing.

This distinction gets blurred in robotics because we use the word adaptation constantly. We have adaptive control, adaptive policies, online calibration, reinforcement learning, and models that can generalize to new tasks. All of that is useful, but most of it happens in software.

The aluminum link still has the same geometry.

The gearbox still has the same ratio.

The fingertip still has the same material properties.

We make the robot's intelligence increasingly adaptive while keeping its physical body surprisingly static.

Biology does both.

A biological body is continuously manufactured

Taken together, these three characteristics can be described more precisely as development, regeneration, and remodeling.

They point toward a larger idea: a biological body is not simply manufactured once and then operated. It is being manufactured, maintained, and modified throughout its lifetime.

A conventional machine generally goes through something like manufacture, operation, wear, external repair, more operation, and eventually replacement.

A biological system combines development, operation, stress detection, repair, and remodeling as overlapping processes.

That changes what kinds of mechanical architectures are viable.

And this brings us back to tendons.

Is a tendon actually a cable?

Mechanically, a tendon certainly behaves like a tensile force-transmission element, so using cables to reproduce part of that function makes obvious sense.

But a biological tendon is not simply a cable with better marketing.

It is living tissue embedded within a living mechanical system. Its properties can change in response to loading, it interfaces with surrounding structures that are themselves compliant and adaptive, and the body around it performs maintenance and repair. The actuator driving it, muscle, is also nothing like a conventional electric motor. Muscle changes force output, stiffness, geometry, and behavior depending on activation and loading.

So when someone says, "Humans use tendons, therefore robotic hands should use tendons," I think the immediate follow-up should be:

Which property of the tendon system are we actually trying to capture?

If the goal is remote actuation, reducing distal mass, compliance, force transmission, underactuation, mechanical coupling, or compact packaging, there may be a very strong engineering case.

But at that point we are no longer copying anatomy because it is anatomy.

We are extracting a useful engineering principle.

That distinction matters.

Every piece of biological complexity should earn its keep

I wrote previously about the complexity-versatility tradeoff in robotics. The basic argument was that complexity is neither good nor bad on its own. What matters is how much useful capability and versatility we get in return for adding it.

I think exactly the same test should apply to biomimetic mechanisms.

A tendon-driven hand can introduce routing complexity, tensioning requirements, friction, hysteresis, stretching, wear, calibration difficulties, and additional maintenance considerations. None of those automatically make tendon systems a bad idea. They simply mean the architecture comes with a complexity bill.

The interesting question is what we receive in return.

If the tendon routing allows lighter fingers, safer contact, better packaging, useful compliance, or robust underactuation, the bill may absolutely be worth paying.

If we are paying that bill mainly because this is how the human hand works, then I become much more skeptical.

The human body receives benefits from its tendon architecture that a robot may not receive. It gets biological repair, remodeling, distributed sensing, tissue adaptation, compliant interfaces, and continuous maintenance.

If we copy the anatomical complexity but remove most of those supporting properties, we may end up copying the invoice while leaving half the purchased features behind.

This is where biomimicry gets tricky

I am definitely not arguing against learning from biology.

Quite the opposite. Biology contains an enormous number of mechanical ideas that robotics can learn from.

But I think there is an important difference between copying a biological principle and copying a biological implementation.

A camera is an easy example. A camera and an eye share some fundamental ideas: light enters through an aperture, optics focus it, and a photosensitive surface converts that light into information.

But industrial cameras do not need to recreate every anatomical feature of an eyeball. Nobody insists that a camera should contain the equivalent of vitreous humor simply because human vision does.

We extract the parts of the solution that remain useful under our own engineering constraints.

Aircraft are another obvious example. Birds demonstrated extraordinarily sophisticated solutions to flight, and aviation learned deeply from the physics of wings, lift, stability, and control.

Then we built airplanes that do not flap.

That does not mean birds are badly designed. Birds are extremely well adapted to the constraints of being living, growing, reproducing organisms.

Airplanes operate under a different constraint set.

Once the constraints change, the optimal implementation can change too.

Wheels are possibly biology's funniest missing feature

The wheel is one of the clearest examples.

It is arguably one of the most useful mechanisms humans have created, yet large-scale animal locomotion basically does not rely on continuously rotating joints.

Imagine trying to put a wheel on a biological leg. Mechanically, the rotation itself is easy enough.

Now route the blood vessels and nerves across a joint that rotates continuously.

Things become interesting pretty quickly.

Machines do not have the same continuity problem. We can put a bearing on an axle and allow one component to rotate relative to another indefinitely.

This illustrates something important: once biological constraints disappear, new mechanical optima become available.

Evolution had to build humans using cells. It had to maintain vascularization, nerves, metabolism, reproduction, growth, repair, and development without access to an external assembly line.

Robots do not share most of those constraints.

So it would actually be surprising if the optimal internal architecture of a humanoid robot turned out to be identical to human anatomy.

What should we copy from humans then?

This is where I think the distinction becomes especially useful for humanoid robots.

One of the strongest arguments for humanoid morphology is that the world is already designed around humans.

Doors assume roughly human dimensions. Stairs assume human legs. Shelves assume human reach. Tools assume human hands. Workstations assume human height. Vehicles, kitchens, warehouses, construction sites, factories, and homes contain an enormous number of design decisions made around the capabilities and proportions of the human body.

That creates a powerful reason for robots to resemble us at the interface level.

Human-like height and reach make sense when the robot needs to work in human spaces. Two arms make sense because much of our environment assumes bimanual manipulation. Human-compatible hands matter because nearly every tool we have created was designed around our hands.

But notice what this argument actually tells us to copy.

It tells us about the interface between the robot and the human world.

It does not automatically tell us what should exist underneath the covers.

Copy the interface, not necessarily the implementation

This connects directly to something I explored in Humanoids Can't Get Too Far Without Softness.

When we say the world is designed around humans, we often think only about dimensions. But our environment is also designed around a body that deforms.

Human fingertips conform slightly around objects and generate useful friction. Skin and soft tissue tolerate imperfect contact. Shoulders, arms, hands, and legs can brush against the environment without every small contact becoming a precise rigid-body collision problem.

Softness is therefore part of the interface between the human body and the world.

But reproducing that effect does not require us to reproduce human tissue.

A compliant mechanism, silicone structure, foam, flexure, variable-stiffness actuator, or some material architecture we have not invented yet might provide the behavior we actually care about.

The environment wants the robot to behave in a human-compatible way.

It does not particularly care whether collagen is involved.

That gives us a design principle I think is useful far beyond hands:

Copy the interface, not necessarily the implementation.

Human proportions may be useful.

Human-compatible contact behavior may be useful.

Human reach, hand size, and locomotion capability may be useful.

But the internal mechanics can be completely different if a different solution works better.

The same principle applies to human-likeness

In A Humanoid Robot Is Judged Against a Human, Not a Machine, I argued for what I called functional human-likeness rather than simply maximizing resemblance.

The closer a robot becomes to a human, the more expectations we attach to it. Five fingers imply dexterity. Eyes imply attention. A human face implies social understanding. Human appearance becomes a promise about capability.

I think there is a mechanical version of the same problem.

Human anatomy can also become a promise.

If we introduce a complex anatomical mechanism because it looks sophisticated or biologically elegant, that mechanism should provide an actual functional advantage. Otherwise we risk creating something like anatomical theater: a machine that imitates biological structure without receiving the capability that made the original structure interesting.

Perhaps humanoid design needs functional human-likeness internally too.

Copy what creates capability.

Ignore what merely creates resemblance.

There is a larger embodied-intelligence problem hiding here

In Robotics Fails Where Humans Improvise, I argued that many real-world environments have effectively evolved around human adaptability.

Construction is an obvious example. Humans constantly compensate for imperfect geometry, inconsistent material behavior, incomplete information, tolerances, friction, unexpected contact, and thousands of small deviations that were never explicitly planned.

We often describe this ability purely as intelligence.

But some of that intelligence is not happening in the brain.

A fingertip can deform before the nervous system has to calculate a perfect contact trajectory. Friction helps stabilize an object without solving a new optimization problem every millisecond. Soft tissue absorbs small errors in positioning. Muscles continuously modulate stiffness depending on the task.

In other words, some intelligence is embedded in the mechanics and materials of the body.

This is one reason I think humanoid robotics will eventually need much more adaptive hardware.

But adaptive hardware does not necessarily mean anatomically accurate hardware.

That distinction is important.

Maybe we are learning the wrong lesson from biology

Today, we are making robot intelligence dramatically more adaptive. Policies can learn from experience, vision systems can generalize across unfamiliar situations, controllers can recalibrate, and robots can increasingly acquire new behaviors from demonstrations or interaction.

Yet while the software becomes more flexible, the physical machine underneath it remains surprisingly static. The body is usually designed once, manufactured once, and then expected to operate with essentially the same geometry and material properties until something wears out or gets replaced.

Perhaps the deeper lesson from biology is not that we should use tendons.

Perhaps it is that we should stop assuming robotic hardware has to remain static.

Imagine robotic structures that gradually change stiffness based on loading history, contact surfaces that become more abrasion-resistant where repeated contact occurs, mechanical systems that detect wear and redistribute forces, or structures capable of some degree of self-repair.

Eventually, we might even see robot morphology itself change slowly according to the work the machine performs.

That would be a much deeper form of biomimicry than reproducing anatomical layouts.

Instead of copying the final form of the human body, we would be copying one of biology's more fundamental strategies: the body itself participates in adaptation.

So, are tendon-driven hands a dead end?

Probably not.

Otherwise this article would have been much easier to write.

Tendon-driven mechanisms solve genuine robotic problems. Remote actuation can reduce distal mass, cable routing can help with packaging, compliance can improve interaction, and underactuation can produce useful mechanical coordination without placing an actuator at every joint.

If tendon-driven hands win because those advantages outweigh their complexity, calibration, wear, and maintenance costs, they should win.

The potential dead end is something else:

assuming that increasing anatomical similarity automatically moves us closer to human capability.

I don't think it does.

Sometimes anatomy contains a brilliant physical principle that should be copied almost directly. Sometimes anatomy is primarily a consequence of constraints that biology could not escape. And sometimes the best robotic solution may be a mechanism that evolution could never produce at all.

That should be exciting rather than disappointing.

The real humanoid argument

This leaves humanoid robotics with a larger philosophical fork.

One direction is to make the machine increasingly human internally as well as externally. That could mean artificial muscles, synthetic tendons, human-like joint arrangements, soft tissue, skin, anatomically inspired feet, and increasingly accurate reproductions of the structures found in our own bodies.

There is a compelling logic behind this direction. Human anatomy demonstrates a level of dexterity, robustness, efficiency, adaptability, and embodied intelligence that no humanoid robot currently matches.

But there is another possible direction.

We could make the interface increasingly human-compatible while allowing the underlying machine to become increasingly nonhuman.

Keep human proportions where the environment demands them. Keep five fingers if five fingers provide the best compatibility with human tools. Keep softness where human environments assume compliant contact. Preserve human-scale reach, mobility, and interaction.

But underneath the surface, use whatever engineering gives us.

Electric motors where biology uses muscles.

Rotary mechanisms biology cannot easily produce.

Replaceable modules where organisms rely on healing.

Non-anatomical transmissions.

Strange linkages.

New materials.

Internal geometries that would be impossible to grow inside an embryo.

The result could be human-compatible without being mechanically human.

This also connects back to the complexity question. Humanoids already sit at the extreme end of robotic system complexity because mobility, manipulation, balance, sensing, dexterity, autonomy, perception, and safety all have to coexist in one machine.

Every additional layer of biomimetic complexity should therefore answer a very simple question:

What capability are we actually buying?

If there is a strong answer, use it.

If the answer is mostly "because that's how humans work," I think we should be much more skeptical.

I suspect the most successful future humanoid may look fairly familiar from across the room. Its proportions, reach, hands, and contact behavior may be recognizably human because it needs to operate in a world that humans built.

But take the covers off and perhaps it should look completely alien.

Not a mechanical reconstruction of the human body, but a machine optimized for performing human-scale work.

And that, to me, is the more interesting question for humanoid robotics:

Should we try to make robots as human as technically possible, or should we copy only the parts of humanity required to interface with our world and make everything underneath as unapologetically machine-like as possible?

I increasingly suspect the second path will take us further.

The objective is not to rebuild a human.

It is to build the best possible machine for a world humans happened to build.