Humanoid Robot: Definition, How They Work, Types and Examples (2026)

A humanoid robot is a physical robot with a human-inspired body structure, often combined with human-like movement or interaction. Most have a torso, head and paired limbs, but the exact form varies. The practical reason is compatibility: a human-scale machine can potentially use doors, stairs, shelves, tools and workstations built for us. The International Federation of Robotics offers a narrower reference point: human-like appearance, the ability to perform tasks in environments designed for people, and performance enhanced by human-like sensing. Definitions vary at the edges, but conversational behavior alone does not make software or a machine humanoid.

That promise is easy to overstate. A demonstration is not a dependable worker, and some systems remain remotely operated or limited to a narrow task. This guide covers the definition, machinery, types, examples, uses and limitations that matter in 2026.

What is a humanoid robot?

Humanoid is a design strategy: give a physical robot enough human-like structure, often paired with human-like movement or interaction, to work in spaces built for people. In the clearest cases, it has two legs, two arms, a torso and a head. Human-scale reach and mobility can let it approach a bench, pick up a tool, carry a container or pass through a standard doorway.

The word humanoid describes form and function, not intelligence. Some humanoids can plan multi-step actions and respond to speech. Others repeat pre-programmed motions or depend on an operator. A human-shaped shell does not prove that a machine understands its surroundings, and advanced AI does not require a human-shaped body.

“General-purpose” describes an ambition, not current capability. Most deployments constrain broadly capable hardware to one workflow, such as moving known parts between fixed stations.

Purpose-built automation is usually better for a stable, high-volume process. Humanoids are most interesting where tasks change and using existing carts, shelves and controls could reduce integration work. Even then, the label says nothing about payload, accuracy, autonomy, availability, price or whether the robot performs a useful task. Those claims need separate evidence.

Humanoid meaning: what the word covers beyond robots

Humanoid means having human form or human characteristics. The word predates robotics and still carries uses outside it. Anthropologists apply it to extinct hominins in the fossil record, and fiction uses it for any human-shaped being, biological or mechanical.

A humanoid robot is therefore one specific kind of humanoid, and the distinction matters when you read a specification. “Humanoid” on its own describes shape. “Humanoid robot” tells you the thing is a machine under some form of control. Shape is also the easiest part of any claim to confirm from a photograph, which is why the rest of this guide treats capability as a separate question.

Defining characteristics and important exceptions

There is no single visual checklist that covers every humanoid. Researchers and manufacturers make different trade-offs, but several characteristics recur.

• Human-like structure: a torso with paired limbs, often including articulated hands or grippers.
• Human-scale reach: the ability to work with objects, controls and surfaces positioned for people.
• Legged or upright mobility: usually two legs, although some upper-body humanoids use wheels for stability and efficiency.
• Whole-body control: coordinated motion across multiple joints so the robot can balance, reach and manipulate an object at the same time.
• Human-facing interaction: cameras, microphones, displays, gestures or speech that help people understand and direct the system.

Two legs are common, not mandatory. A social robot may have a head, torso and arms on a wheeled base. A warehouse system may be unmistakably humanoid above the waist but use a non-human lower body. Conversely, a biped with no face can still be humanoid if its structure and working envelope are designed around human environments.

Hands are another exception. Five-fingered hands are useful for tools built around human grip patterns, but they add mechanical and control complexity. Many industrial humanoids use simpler grippers, interchangeable end effectors or task-specific hands. The better question is not “Does it copy a person exactly?” but “Which parts of human anatomy help it do the intended job?”

The boundary also excludes some technologies that use human language or imagery. A voice assistant, animated avatar or conventional industrial arm is not humanoid simply because it communicates with people. An exoskeleton follows the human form but is worn and powered through a person rather than acting as an independent robot. Conversely, a faceless biped can be humanoid when its body, reach and movement are designed around human spaces. The useful test is embodied structure and function, not whether the machine appears friendly or lifelike.

Humanoids versus androids, cobots and industrial robots

A humanoid resembles the human body in its overall structure or movement. An android is a narrower idea: a humanoid designed to look convincingly human, often with a face, skin-like materials and expressive behavior. Every android is humanoid, but an exposed mechanical biped such as Atlas is not usually called an android.

A cobot, or collaborative robot, is defined by how it is used around people rather than by its shape. Many cobots are fixed robotic arms. A humanoid could form part of a collaborative application, but its human-like appearance does not make it safe to share an unguarded workspace. Safety depends on the complete system: the task, payload, tooling, speed, controls, environment and foreseeable contact.

A conventional industrial robot is commonly a fixed arm or specialized machine inside a designed cell. It can be faster, more precise and more economical than a humanoid at a repetitive, stable task. Humanoids are being developed for the awkward middle ground where work changes, the environment was built for people, and rebuilding the process around fixed automation is difficult.

Humanoids are not replacements for every robot. A gantry, conveyor, autonomous mobile robot or six-axis arm is often the better tool. The form earns its place only when human compatibility and flexibility outweigh its extra complexity.

How humanoid robots work

A humanoid combines a mechanical body, sensors, actuators, real-time control, task software and a power system. These parts form a continuous loop: sense the world, estimate what is happening, choose an action, move, measure the result and correct the next movement. The loop must run quickly because a delayed correction can turn a small balance error into a fall.

Perception and sensing

Cameras provide color, depth and motion information. Depth cameras or lidar can help build a three-dimensional map. Microphones capture speech and environmental sounds. Joint encoders report the angle and speed of each joint, while inertial measurement units estimate the robot’s orientation and acceleration. Force, torque and tactile sensors reveal contact with the floor, a tool or a person.

No sensor tells the whole story. The robot combines several streams to estimate its body and surroundings, identify objects and track them through a task. Lighting, reflective surfaces, clutter and partial occlusion can undermine a system that worked in a clean demonstration.

Actuators and mechanical design

Actuators create movement at the joints. Electric motors are now common in commercial designs, typically paired with gearing and control electronics. Other research platforms have used hydraulic systems. Designers balance strength, speed, weight, efficiency, impact tolerance, heat and backdrivability—the extent to which an external force can move a joint.

Degrees of freedom describe the independent ways a body can move. More joints can expand reach and dexterity, but each adds components, wiring, calibration and control work. A useful humanoid is therefore not the one with the largest specification number. It is the one whose joints, hands and range of motion reliably match the task.

Control, autonomy and AI

Low-level controllers regulate joint position, velocity and force many times per second. Higher layers coordinate balance, walking, reaching and grasping. A task planner then connects those skills into actions such as locating a tote, walking to it, lifting it and placing it at another station.

Machine learning can improve perception, grasp selection and motion policies. Reinforcement learning in simulation lets a system practice large numbers of movements without repeatedly damaging physical hardware. Demonstration data from teleoperation can teach a robot how a person performs a task. Language and vision models may help translate an instruction into a plan, but safety-critical motion still needs constraints, monitoring and fallbacks.

Autonomy is a spectrum. A robot may work independently inside a known cell, request help when confidence drops, or be directly teleoperated. Remote assistance is not inherently a flaw; it can be a practical way to recover from rare situations and collect training data. What matters is whether a vendor clearly states where the operator remains in the loop.

Locomotion, balance and power

Honda’s development history distinguishes static and dynamic walking. In static walking, the center of gravity stays within the supporting sole; during dynamic walking it may move outside that support area. In either case, the robot must shift its mass, place a foot, absorb contact and prepare the next step while accounting for the payload and ground. Model-based control, learned policies or a combination of both can generate the motion. Whole-body control matters because moving an arm or lifting a box changes the forces that keep the robot upright.

Power is a hard constraint. Onboard batteries must supply computers, sensors and many motors while remaining light enough to carry. Demanding motion shortens operating time and creates heat. Current designs address this with scheduled charging, replaceable batteries, automated battery swaps or work patterns that alternate active periods and charging. An impressive movement is only commercially useful if the system can repeat it for the required shift.

Main humanoid robot types

Industrial humanoids are built for factories, warehouses and logistics sites. They emphasize payload handling, repeatability, integration with business systems and serviceability. Their first tasks tend to be narrow but valuable: material movement, machine tending, parts sequencing or tote handling.

Research humanoids are platforms for studying locomotion, manipulation, perception and human-robot interaction. They may perform remarkable movements without being products designed for continuous customer operation. Research results often transfer into later commercial machines.

Social and service humanoids communicate with people through speech, gaze, gestures or screens. Some use wheeled bases because a stable platform is more practical for reception, education or guided interaction than legs. Their “human” qualities are concentrated in the upper body and behavior.

Telepresence and avatar robots extend a remote person’s sight, voice and actions into another location. Their value is presence and manipulation rather than independent decision-making. Entertainment and expressive androids prioritize lifelike appearance or performance, sometimes at the expense of general mobility.

These categories overlap. A research platform can become an industrial product, and an industrial machine may use teleoperation during training or exception handling. Classifying a robot by its intended deployment is usually more informative than judging it only by appearance.

A short history of humanoid robots

Humanoid machines appeared in stories and mechanical automata long before electronic robotics. Modern humanoid research emerged when engineers could combine sensing, computation and powered joints in one system.

Waseda University describes WABOT-1, completed in 1973, as the first full-scale humanoid robot. It combined a limb-control system with vision and conversation functions, could walk on two legs and could grip objects. WABOT-2 followed in 1984 as a specialist robot designed to read music and play an electronic organ.

Honda began bipedal research in 1986. Its P-series machines developed dynamic walking and self-contained operation before ASIMO was introduced in 2000. ASIMO brought stable movement, stairs, interaction and an approachable scale into public view. Honda’s later account is also a useful reality check: decades of demonstrations exposed unresolved questions about falls, autonomy and safe operation around people.

During the 2010s, research platforms pushed dynamic balance and whole-body movement, while social robots explored public interaction. The current phase is more commercial. Companies are narrowing the problem to repeatable work in customer facilities, using electric actuation, improved perception, simulation, teleoperation data and learned control. The emphasis has moved from “Can it perform this once?” toward “Can it deliver a safe, measurable result every shift?”

Representative real-world examples

The examples below illustrate different stages and design choices. They are not a ranking; for a broader comparison, see our regularly updated guide to the best humanoid robots and the RoboZaps humanoid database.

• Boston Dynamics Atlas: an electric industrial humanoid built around dynamic mobility and material handling. Boston Dynamics says the product version can operate autonomously, by teleoperation or through a tablet interface, and is being introduced through selected industrial deployments.
• Agility Robotics Digit: a biped optimized for logistics rather than copying every detail of a person. GXO describes a multi-year service agreement integrating Digit with other warehouse automation at its SPANX facility after a 2023 pilot.
• Figure robots: general-purpose industrial humanoids being developed through factory work. BMW Group says Figure 02 supported production of more than 30,000 BMW X3 vehicles over ten months in 2025. On June 30, Figure reported that Figure 03 had arrived at Plant Spartanburg and completed its first demonstration of a logistics-sequencing workflow.
• Apptronik Apollo 2: Apptronik’s current data-collection and training platform, operating in bipedal and wheeled configurations at Robot Park and selected partner sites. The company describes Apollo 3 as its upcoming commercial fleet.
• Unitree G1: a compact biped sold in base and EDU configurations. Unitree lists secondary-development support for G1 EDU, making that configuration the relevant choice for research and custom development.
• Ameca: an expressive social humanoid focused on communication and human-robot interaction rather than factory locomotion.

Product names alone hide important differences. Some machines are commercial products, some are early-access platforms, some are pilots, and others remain research systems. Buyers should look for evidence of the exact task, environment, duty cycle, supervision and support model—not just a polished video.

How to judge a humanoid robot claim

Start by naming the stage. A laboratory demonstration shows a behavior under those conditions. A customer pilot adds integration and real-world variation. A recurring production deployment is stronger evidence of reliability and economic value. None should be described as another.

Ask whether the task is continuous or edited into highlights. Useful details include operating time, completed cycles, intervention rate, payload and recovery procedure. A robot may complete the visible motion autonomously while a person selects objects, resets failures or drives it between takes.

Check who supplied the number. Manufacturer specifications and customer statements are primary sources for what those organizations claim, not independent tests. Because hardware and software change quickly, dates matter. We label company-reported results and remove claims that cannot be traced to a suitable source.

Customer evidence can still improve confidence. It can confirm that a robot worked at a named site, on a defined task and for a stated period rather than only in its maker’s laboratory. It does not replace an independent benchmark: both companies may approve the release, and intervention rates or failed cycles may be omitted. Strong evaluation states what the customer confirms, what the manufacturer reports and what remains unknown.

The same discipline applies to availability. “Unveiled,” “in development,” “accepting inquiries,” “in pilot,” and “shipping to customers” are different states. A deposit page does not prove general delivery, and a forecast is not a shipment. This guide avoids a giant specification table because those fields age at different speeds; the database and individual reviews are better places to maintain model-level detail.

Uses and deployment environments

Factories and warehouses are the leading proving grounds because the tasks can be valuable and the environment can be controlled. A humanoid may move parts between stations, feed a machine, build an order or perform repetitive handling at a workstation originally laid out for a person. Existing tools and aisles create the strongest argument for the form factor.

Useful deployment metrics go beyond whether the robot completed the motion. Operators need throughput, intervention frequency, uptime, charging time, safety events, maintenance effort and the cost of connecting the robot to existing equipment and software. A pilot becomes more persuasive when the customer expands it after observing those measures over time. Without that context, a successful cycle is evidence of technical capability, not proof of a dependable business case.

Hazardous inspection and response are another long-term fit. A human-compatible robot could enter a site built for people, climb steps, use equipment and keep an operator away from heat, chemicals or unstable structures. The difficulty is that these environments are less predictable than factories, so reliability and remote operation remain crucial.

Service settings include reception, education, guided interaction and research into care support. In these roles, speech, gaze and gestures may matter more than walking. Physical assistance around vulnerable people sets a much higher safety bar than delivering information from a fixed position.

Home use attracts attention because homes are the ultimate human-designed environment. They are also full of clutter, pets, children, thresholds, delicate objects and tasks that change constantly. Today’s realistic consumer options are limited compared with the general-purpose helper imagined in marketing. Our guide to humanoid robots for home use separates available products from announced ambitions.

For a sector-by-sector view, see the dedicated guide to humanoid robot applications. The important test in any setting is whether the robot improves a specific workflow after integration, supervision, maintenance and downtime are included.

Current limitations, safety and ethical concerns

Reliability and economics

Real work requires thousands of uneventful cycles, recovery from variation and predictable maintenance. An unfamiliar object, snagged cable or changed surface can expose weaknesses hidden by a scripted demo. Fleet support, spare parts, training and software integration matter as much as movement.

The economic comparison must include the complete system. Hardware is only one line item; integration, site preparation, supervision, charging, maintenance and lost output during faults also count. A specialized machine may deliver a better return where the task is stable. For current purchasing context, use the separate guides to humanoid robot costs and where to buy a humanoid robot.

Physical and operational safety

A human-scale robot has moving mass, pinch points and stored energy. A fall, unexpected reach or dropped payload can injure someone even when the system is marketed as collaborative. NIST’s task-based framework evaluates the application, tooling, possible contact and environment rather than treating the robot alone as the risk. For industrial applications, ISO 10218-1 covers the robot as a machine, while ISO 10218-2 covers its integration into complete applications and cells. Other environments may fall under different standards or regulations.

Practical controls include speed and force limits, monitored separation, safe stops, guarded zones, fall planning, payload limits and clear procedures for faults. Operators also need to know when the system is autonomous, remotely assisted or awaiting input. Human-like appearance should never substitute for validated safety behavior.

Privacy, work and accountability

Humanoids can carry cameras, microphones and operational logs through spaces where people work or live. Organizations need rules for what is recorded, why it is retained, who can access it and whether remote operators can see sensitive environments. Cybersecurity is physical safety when a connected system can move.

Workforce effects depend on deployment choices. A robot may remove repetitive lifting, create technical roles or change staffing, but those outcomes are not automatic. Employers should involve workers early, measure actual ergonomic and productivity results, and provide training for changed roles. Claims about wholesale job replacement are forecasts, not facts.

Responsibility must also be explicit. When a robot’s action involves a manufacturer, software provider, integrator, operator and customer, incident reporting and decision authority cannot be left vague. These issues are explored further in our guide to the challenges in humanoid robotics.

What comes next

Developers are concentrating on a small number of industrial tasks, improving reliability through customer work, and using teleoperation and simulation to gather training data. Hands, perception, battery management, safe recovery and fleet software matter more than theatrical new movements.

Hardware will also become easier to operate as a fleet rather than as a one-off machine. That means remote diagnostics, controlled software releases, shared task libraries, maintenance records and clear escalation when a robot reaches an unfamiliar state. A technically capable body without those operating systems is difficult to deploy at scale.

Expect uneven progress. A robot may become dependable at parts handling while remaining poor at household cleanup. Skills learned in one facility may transfer imperfectly to another with different lighting, containers or layouts. The general-purpose vision will be assembled from many constrained successes, not delivered as a single leap.

The clearest evidence will be operational: repeat customers, longer unsupervised runs, documented task performance, maintainable hardware and deployments that expand after a pilot. Our humanoid robot company guide tracks the organizations building these systems, while this page stays focused on the definition and technology.

Humanoid robots are already real, but that includes research platforms, pilots and production tools. Near-term progress depends on proving safe, reliable and economically useful performance in bounded jobs.

Keep exploring: the cheapest humanoid robots you can buy, and our humanoid robot news tracker for what changed this week.

Compare on RoboZaps: Figure 03, ASIMO, Ameca, Apptronik Apollo, Figure 02 and Digit. Full specifications and a quote route for each, alongside every humanoid robot we carry.