What Is a Humanoid, and at What Point Does It Stop Being One?

What makes a humanoid useful, and when does it become just another robot? 

Decorative image, inspired by Neo from 1X and Figure03 from Figure AI Humanoids, and modified by the author (GenAI used). Mehrdad Farimani, Sep. 2026, Sweden

As you probably know by now, I am not a humanoid lover and a pure optimist. I am not a pessimist either. I try to look at humanoids realistically, based on my years of working on the technical side of humanoid robotics.

This sounds like a simple question: what is a humanoid?

It is not.

It can be a philosophical question, a biological one, a technical one, a system-design question, or even an investor and finance question. The label changes what people expect, which machines they compare and, sometimes, the investment story.

So before trying to define a humanoid, we should ask why the definition matters. What difference does it actually make? What is meaningfully different between a humanoid and a non-humanoid robot?

I think the first notable thing is that a humanoid has a ground truth: a typical adult-sized human living in the modern world. This does not mean that the robot must copy human anatomy. In fact, as I argued in The Wrong Way to Design a Humanoid, blindly copying biology can be a very bad engineering strategy. Human anatomy grew, heals and remodels itself. Machines are manufactured, wear out and are repaired from the outside. The useful target is often the human-compatible interface, not the biological implementation underneath it.

But the reference still exists. Human height, reach, hand size, field of view, walking geometry and interaction patterns give humanoid designers a baseline that other robot categories do not have.

Humanoids are on a spectrum

First, let us agree that being humanoid is a spectrum, so as a human :D

A robot does not need to reproduce one idealized anatomy down to every joint to belong to the humanoid category.

Engineering is compromise. We simplify a bio-inspired system to make it manufacturable, serviceable, controllable, safe and economically useful. The direction in which those compromises should be made is still an open discussion in humanoid robotics.

Also, when I say humanoid, I mean the complete hardware-software system. A human-shaped shell without the perception, control and intelligence to use that body is not much of a humanoid in the functional sense.

So the better question is not whether a robot has every human feature. It is whether its remaining human-likeness still provides the main values for which we built a humanoid in the first place.

Why build a humanoid at all?

I see five main arguments.

1. The world is built for humans

This is the most common and most widely accepted argument. Most environments in which we want humanoids to work assume a human form factor.

Doors, stairs, shelves, tools, handles, workstations, vehicles, kitchens, warehouses and factories contain dimensions and interaction points based on human proportions. If a robot must enter these environments without rebuilding them, human-like height, reach, mobility and manipulation become useful interfaces.

This does not prove that a humanoid is optimal for every task. Wheels can be more efficient on flat floors, and a fixed arm can be better at repetition. A humanoid is an argument for accessing many tasks in an environment that already exists. As an example, Humanoid feet have the smallest footprint to keep a system with that size stable. Compare this with a wheel-based system that requires a heavy and fairly large base to be able to stabilize and bend.

2. Human data becomes more transferable

The second reason is training and robot learning. The closer a robot's morphology is to a human's, the more directly it can benefit from human demonstrations, egocentric video, motion capture and teleoperation.

This transfer is not automatic. Different dynamics, strength, sensing and control create a serious embodiment gap. But similar kinematics reduce the translation required, giving a robot with roughly human arms, hands, head position and reach a potentially more scalable path to learning from human data. Read about the State of Data Collection for Robotic Manipulation. 

3. Generalization can justify the complexity

We cannot prove that the humanoid form is the most generalizable possible robot. That would be a very large claim. But among the machines we have already built, the human form is our strongest demonstrated example of a body that can perform a huge range of physical tasks in human environments.

That makes it a reasonable reference for a general-purpose robot, or GPR.

Generalization can also make a complex robot economically viable because it combines highly valuable tasks with necessary but low-value ones. Carrying a component might not justify an expensive machine, but carrying it, finding the tool, performing the operation and cleaning the area might.

As I discussed in Robotics Complexity vs Versatility: Why Robots Are Converging, complexity has to earn its keep. The humanoid form matters only if the versatility it creates is worth its mechanical, computational and reliability cost.

4. Human interaction can be more natural

For complex machines working near people, natural interaction can be better than forcing every instruction through an app, joystick or specialist interface.

Speech, gaze, gesture and readable body language let people use interaction patterns they already know, without an intermediary modality.

But appearance is also a promise. In A Humanoid Robot Is Judged Against a Human, Not a Machine, I argued that a human-like feature raises expectations about the capability behind it. A face implies social understanding. Five fingers imply dexterity. Eyes imply attention. Human-likeness helps interaction only when behavior supports what the design appears to promise.

5. The body may help create the intelligence

This is my favorite argument, although it is also the most hypothetical.

Biology suggests that the human body did not simply appear after the human brain. The two developed together. Standing on two legs freed the hands. Multi-jointed fingers enabled toolmaking. Flexible reach, touch and manipulation created new possibilities for acting on the environment, and those possibilities placed new demands on the brain.

A dolphin may have considerable intelligence, but its body does not allow it to build and manipulate tools in the same way. What a system can learn is partly shaped by what its body lets it do.

This does not prove that giving a robot five fingers will produce general intelligence. It suggests something more modest and useful: embodiment affects the space of experiences from which intelligence can develop.

To be a Humanoid, what capabilities does that require?

If these are the reasons for building a humanoid, we can work backwards and ask what is necessary to achieve them.

Dexterous, bimanual manipulation

A general-purpose humanoid should have two articulated arms, typically with six or seven degrees of freedom each, positioned in roughly adult-human proportions.

Two arms increase reach and allow the robot to carry medium-to-large objects, stabilize with one hand while working with the other, multitask and use whole-body strategies for balance.

At the end of the arms, the robot needs human-compatible dexterity. In the strongest version, this means compact three-to-five-finger hands, with at least two useful active degrees of freedom per finger, a capable wrist, and fingers close enough to human size to use human tools. Flexion and extension are basic. Abduction, opposition and good thumb placement make the hand much more useful.

Fingernails sound cosmetic, but support precision pinching, scraping and separating thin objects. Small anatomical details can have functional value.

The hand should also be compliant and, where possible, backdrivable and empowered by impedance control. Manipulation in the real world depends on tolerating contact rather than treating every collision as a control failure. There’re many good arguments for why softness and compliant materials matter in a humanoid hand. 

Haptic feedback can come from joint-torque sensing, tactile sensors, motor-current estimation and other force sensing. Vision alone is not enough when contact is occluded. Only implementing a touch sensor is not also enough. Practice this right now; try to move an object through another object (move a cup with a pen in hand), close your eyes, and ask yourself: what sensing is telling you that the cup is moving? Is it the touch? Force through your joints? Generally haptic, or something else. 

Bipedal locomotion

Bipedal locomotion is compatible with stairs, ladders, narrow passages and workspaces designed around a standing human. It gives an adult-sized machine a relatively small footprint while keeping its upper body at working height.

This does not mean bipeds are superior on every terrain. Wheels are dramatically more efficient on smooth floors, and quadrupeds can offer a larger support polygon and good mobility on rough ground. The advantage of bipedalism is compatibility across the broad mix of spaces humans already navigate, not universal superiority.

The feet should be close enough to human shoe dimensions to fit stairs, walkways and workstations. Active or passive toes may improve gait, contact and push-off, but they must justify their extra complexity.

Hip/Pelvis is much more significant than we think. Imagine the hip as the fixed point of our body, with the upper limbs and lower limbs attached to it. In classical humanoid robotics control and dynamics, ZMP (Zero Moment Point) was and is the most important reference for designing a system with locomotion. 

Torso and waist 

The torso is not just a box connecting arms and head. Its proportions determine reach, tool access and compatibility with human spaces. An articulated waist, often with two or more useful degrees of freedom, can increase reach, improve balance and push recovery, and let the robot reposition its upper body without moving its feet.

That matters in cluttered spaces, where moving the feet for every small extension of reach is slow. 

Head, perception and face

The head provides a human-compatible sensor viewpoint. Much egocentric human data is collected from approximately this position, which helps align training data with the robot's perception.

A two-degree-of-freedom neck lets the robot inspect its workspace, look toward people and scan its own body. Movable eyeballs are not necessary because multiple cameras can create a wide field of view. A face is not mechanically necessary, but readable gaze and expression matter when human interaction is central.

Again, the face should not promise more understanding than the system can deliver. However, it can also depend on the application of the robot. For example, a human head, in a mechanical sense, has more than two DoFs. We often rely on flexion, extension, and rotation. But articulated flexion/extention (pitch) or lateral flexion (roll) is also a significant movement when it comes to conveying specific interaction, especially for social robots such as Ameca. 

Multimodal understanding and communication

A useful humanoid needs more than a body. It needs natural-language understanding and speech, object recognition and scene interpretation, vision-language-action capabilities, navigation, path planning, localization and mapping, and useful auditory perception through microphones and sound localization.

These sensory and communication capabilities are what make the morphology usable. So far, smell and taste have not shown the same general value, although specialized applications may eventually justify them.

The less visible requirements

Several high-level qualities are just as important as body layout.

Power efficiency and runtime are obvious ones. A human-like working shift of six to eight hours on battery is a useful ambition, although most current humanoids remain far from it under continuous demanding work. The robot should normally carry its own energy, avoid tethering and charge or swap batteries with minimal help. The exact requirement remains application-dependent.

Local computation is also critical. Edge computing and cloud services can extend capability, but immediate perception, balance, collision response and basic task execution cannot depend on a stable internet connection or tolerate network latency. A useful humanoid needs enough onboard processing to remain safe and functional locally.

Backdrivability and impedance control matter most where interaction occurs. A knee does not necessarily need the same degree of backdrivability as a hand or arm, but the complete system needs controlled compliance. This improves safety, supports physical human-robot interaction, lets the machine adapt to imperfect contact and can reduce damaging peak loads that shorten hardware life.

Finally, the external shell is not decoration. It protects the robot and the people around it, carries identity and aesthetics, and creates affordance. The appearance should signal the system's intended use and actual capability. Softness matters because human environments assume bodies that deform, absorb impact and tolerate imperfect contact. I explored this more directly in Humanoids Can't Get Too Far Without Softness.

Can a humanoid have non-human features?

Of course.

Humans do not have lidar, but lidar may improve navigation and safety. A robot can have cameras behind its head, a shoulder that rotates beyond human range or a neck that turns backwards. Atlas has famously demonstrated non-human/super-human ranges of motion. These features can increase capability, although they may also make learning from human data more complicated.

Wheels under the feet are another interesting case. A robot with agile bipedal legs and wheels beneath its feet is not the same as a conventional wheeled mobile manipulator. It is closer to a human wearing roller skates. If the legs still provide human-compatible stance, stepping, reach and terrain access, the wheels extend the locomotion system rather than replace it.

All of these additions can be good if they do not make the system disproportionately complex compared with the capability they create.

So, where does a humanoid stop being a humanoid?

Is a human-like upper body on a wheeled base a humanoid? Is a biped with two-degree-of-freedom parallel grippers a humanoid? What about a robot with no face?

There is no honest boundary based on a single component.

A collaborative robot arm reproduces part of human morphology, but we do not normally call it a humanoid because it is not a complete embodied system. A prosthetic robotic hand may reproduce a full human hand, but it is still a humanoid subsystem rather than a humanoid robot.

For a general-purpose robot, I would use a functional threshold:

A humanoid is a complete embodied robotic system whose human-like morphology materially enables human-environment compatibility, learning from human data, broad physical generalization and natural interaction.

The fifth argument, co-development of body and intelligence, remains a supporting hypothesis rather than a practical test today.

Under this definition, a wheeled upper-body robot may be humanoid in appearance and contain humanoid subsystems, but it is better described as an anthropomorphic mobile manipulator if wheels remove much of its access to stairs, narrow footholds and other human locomotion spaces. A biped with simple parallel grippers can reasonably be called a humanoid, but it occupies a lower point on the spectrum because it sacrifices tool compatibility and dexterous generalization. A robot without a face can still be fully humanoid if social facial interaction is not central to its capability envelope.

The category stops being useful when too many of the human-derived advantages disappear. If the body can no longer access human spaces, transfer human demonstrations, generalize across human-scale physical tasks or communicate in naturally readable ways, then calling it humanoid describes its silhouette more than its function.

That distinction matters. Otherwise, humanoid becomes a marketing term for anything with two arms and a camera on top.

The boundary will also move. We may discover that ankle rotation, tactile skin, active toes or another capability currently treated as optional is essential to general-purpose performance. Today's definition is based on today's knowledge and today's systems. It is not a law of nature.

All in all, what we are seeking is human-like proportion and capability, just enough to make the system useful in a world built around humans.

Different humanoids will make different compromises, and specialized variants can absolutely be useful. But when discussing a general-purpose humanoid, the full system matters. The aim is not maximum anatomical similarity. It is the strongest practical combination of compatibility, learnability, generalization and interaction.

A humanoid stops being a humanoid not when it loses one human feature, but when the human form is no longer doing meaningful work for the system.

The objective is not to rebuild a human.

It is to build the most general-purpose machine we can for a world humans happened to build.

Cheers