Humanoid robot

A humanoid robot is a machine built with a human-like body: two legs, two arms, a torso, a head, and hands capable of grasping. The form is chosen so that the machine can operate in spaces and with tools that were designed for people, rather than requiring the environment to be rebuilt around it 1.

Humanoids are among the oldest dreams of robotics and among the newest commercial categories. Honda's ASIMO became a public symbol of what walking machines could do, Boston Dynamics' Atlas made dynamic movement look routine, and a wave of companies founded since 2021 has promised general-purpose labour in factories and homes.

This entry covers the case for the human form, the history, locomotion and manipulation, the role of learning, the current machines, the economics, and the gap between demonstration and deployment.

The strongest argument for humanoids is that the world is already built for humans. Doorways, stairs, shelves, vehicles, and hand tools assume a body of a particular shape. A robot that fits that shape can be deployed without modifying the building or the workflow, which is a large advantage in existing factories and warehouses.

A second argument is data. Human demonstration is the richest source of manipulation data, and a body that resembles a person's can learn from videos, teleoperation, and motion capture more directly than a radically different form. The hope is that humanoids inherit the scale of human-generated data 2.

A third is flexibility. A task-specific machine is efficient at one job and useless at the next. A general-purpose humanoid could be reassigned, which matters where product lines change and automation must be reprogrammed.

The counterarguments are equally strong. Wheels are more efficient than legs on flat floors, a fixed arm is more precise than a walking one, and special-purpose automation is cheaper and safer for repetitive work. The human form is a bet that generality will be worth the mechanical cost 3.

WABOT-1, completed at Waseda University in 1973, is usually described as the first full-scale humanoid robot: it could stand, walk a few steps, grasp objects, and communicate in a limited way. The project established the pattern of university laboratories building complete bodies rather than isolated arms.

Honda's programme began in the 1980s and produced a series of walking machines, culminating in ASIMO in 2000. ASIMO could walk, climb stairs, and respond to gestures, and it became a global ambassador for the field. Honda ended the programme in 2022, after two decades in which the robot remained a demonstration rather than a product 4.

Boston Dynamics' Atlas, unveiled in 2013 for the DARPA Robotics Challenge, shifted the emphasis to dynamic movement. The hydraulic machine could run, jump, and recover from pushes, and its demonstrations set a new expectation for what a legged robot could do. An electric Atlas was announced in 2024 1.

The commercial wave began around 2021, when advances in actuators, simulation, and learned control coincided with investor interest. Companies including Figure, Tesla, Unitree, Agility Robotics, 1X, and Apptronik announced humanoids aimed at logistics, manufacturing, and eventually homes 5 6.

Walking is a controlled fall. A bipedal robot must keep its centre of mass over a small support polygon, place its feet on uncertain ground, and absorb impacts, all while carrying a payload. Classical approaches used the zero-moment point, a measure of where the ground reaction forces must act to keep the robot balanced, and model-predictive control to plan footsteps 1.

Modern controllers increasingly use reinforcement learning. A policy is trained in simulation to produce joint torques from observations of the body and the terrain, with rewards for staying upright and moving toward a target. The policy is then transferred to the real machine, a process that requires modelling actuator delays, friction, and sensor noise accurately enough that the simulation is a useful proxy 3.

Actuation has changed with the controllers. Hydraulic actuators gave Atlas its power and speed but were heavy, leaky, and expensive. Electric quasi-direct-drive actuators, which trade gear reduction for transparency and controllability, have become the default for new humanoids, and they make torque control and backdrivability practical 1.

Battery life remains a constraint. Legged locomotion is energetically expensive, and most humanoids operate for a few hours between charges. That is enough for a demonstration and often insufficient for a shift, which is one reason deployment pilots tend to be short and closely supervised.

A human hand has around twenty degrees of freedom, and no production robot hand matches it. Most humanoids use three-fingered or five-fingered hands with six to twelve actuated degrees of freedom, sacrificing dexterity for cost, weight, and reliability. The result is a machine that can grasp a box and struggles with a key.

Grasping has been studied for decades, and the classical pipeline — perceive the object, plan a grasp, execute — works in structured settings. The harder problem is in-hand manipulation: rotating, reorienting, and adjusting an object without dropping it, which requires touch sensing and fast feedback that most platforms lack.

Tactile sensing is improving. Cameras inside fingertips, capacitive arrays, and force-torque sensors at the wrist give a robot some sense of contact, but the coverage and durability are far below human skin. Without reliable touch, a robot must grip harder than necessary, which is slow and can damage objects.

The manipulation gap is the main reason humanoid demonstrations concentrate on carrying, picking, and placing. Those tasks are valuable in logistics, and they are a small fraction of the work that human hands do 7.

The dominant architectural question is whether to build a modular pipeline or an end-to-end learned policy. A modular system separates perception, planning, and control, which makes failures diagnosable and safety arguments tractable. An end-to-end system maps observations to actions with a single network, which can absorb large datasets and generalise in ways that hand-written pipelines do not.

Vision-language-action models sit between the two. They take an image and an instruction, produce a sequence of actions as tokens, and are trained on both web data and robot trajectories. RT-2 demonstrated that knowledge from internet-scale training could improve manipulation, including on objects and instructions not seen during robot training 7.

Data is the bottleneck. Robot trajectories are expensive to collect, and the largest public efforts pool data across institutions and robot types to train generalist policies 2. Teleoperation, in which a person controls the robot to generate demonstrations, is the standard collection method, and it is slow.

Simulation fills some of the gap. Policies trained in physics simulators can transfer to real machines if the simulation is calibrated and the policy is robust to variation. The approach has worked well for locomotion and less well for contact-rich manipulation, where small modelling errors matter.

The present generation is diverse. Boston Dynamics' electric Atlas is a research and industrial platform with a strong emphasis on dynamic movement 1. Tesla's Optimus is presented as a general-purpose robot for manufacturing and eventually homes, with the company's own claims standing in for independent evaluation 6. Figure's robots have been piloted in automotive and logistics settings 5. Unitree sells low-cost platforms, including the G1, to researchers and developers, which has expanded access to humanoid experimentation 8.

Agility Robotics' Digit is designed for warehouse work and has been deployed in logistics pilots. 1X has focused on home environments, and Apptronik on industrial applications. Several Chinese manufacturers, including Unitree, UBTech, and Fourier, have announced platforms at prices well below earlier research machines.

The technical specifications vary widely: height from about one metre to 1.8 metres, weight from roughly 30 to 90 kilograms, payload from a few kilograms to around 25, and battery life from about two to five hours. These figures come from manufacturers and are not measured to a common standard, so comparisons between platforms should be treated cautiously.

What the machines have in common is that they are prototypes or early production units. None has been deployed at the scale of industrial robots, and the number of units in commercial operation is small relative to the announcements.

The business case depends on cost, uptime, and the value of flexibility. A humanoid that costs as much as a car and works for a few hours a day is not competitive with a person on hourly labour, and it is far from competitive with a fixed robot on a repetitive task. Companies are betting that volume manufacturing will lower prices and that software improvements will raise uptime.

Pilots are the current stage. Manufacturers have tested humanoids for moving totes, loading parts, and inspecting equipment, usually in structured environments with close supervision. A pilot demonstrates that a task is possible; it does not demonstrate reliability, maintainability, or safety at production scale.

Teleoperation is often present in demonstrations and rarely disclosed in marketing. A robot that appears autonomous in a video may be operated by a person wearing a headset, and the gap between a teleoperated demo and an autonomous shift is the central uncertainty in the sector. Companies that publish autonomy metrics are the exception.

Safety certification is another barrier. Industrial robots are governed by standards that assume fixed machines behind fences. A legged robot that shares space with people requires a different risk assessment, and standards specific to humanoids are still being developed.

A humanoid is a heavy machine that moves on two legs near people. Falling is the obvious hazard: a 70-kilogram robot that loses balance can injure someone nearby. Collision avoidance, force limiting, and emergency stop systems are necessary, and they are harder to design for a machine that is supposed to work in unstructured space.

Beyond physical safety, there are questions about surveillance and dual use. A humanoid with cameras and microphones in a workplace or home collects data about the people around it, and the privacy implications are rarely addressed in product announcements. Police and military interest in humanoid platforms raises additional questions about accountability.

The labour question is contested. Advocates describe humanoids as filling shortages in dangerous and repetitive work; critics describe them as a way to replace workers while externalising the costs of displacement. The evidence so far is too thin to settle the argument, because deployment is too small to measure.

Finally, there is the question of trust. A machine that looks like a person invites expectations that it cannot meet, and the design of that expectation is an ethical choice as much as an engineering one.

The first disagreement is about form. One camp argues that the human body is a general-purpose solution to a general-purpose environment, and that humanoids will dominate once the mechanics are good enough. The other argues that generality is better achieved by fleets of specialised machines, and that the human form is a public-relations choice rather than an engineering one 3.

The second is about autonomy. Demonstrations are often teleoperated, cherry-picked, or both, and companies rarely publish failure rates. Researchers who work on locomotion and manipulation tend to describe the remaining problems as substantial; marketing materials tend to describe them as solved. The truth is visible in deployment numbers, which are small.

The third is about timelines. Predictions of humanoids in homes within a few years have been made repeatedly since ASIMO, and the machines have improved without becoming products. The current wave has better software and cheaper hardware than its predecessors, which may make the difference, and the burden of proof is on the deployment rather than the demonstration.

The fourth is about the role of foundation models. If vision-language-action models continue to improve with scale, humanoids could inherit capability from general AI research; if contact-rich manipulation requires data and methods that language models do not provide, progress will be slower. The field is betting on the first and has not yet demonstrated it 7.