Humanoid Robot

A humanoid robot is a robot whose body plan or functional arrangement is modeled on important features of the human body. A typical full-body design has a torso, two arms, a head or head-like sensor location, and two legs, but the boundary is not universal. Some researchers include machines with wheels instead of legs, simplified hands, or no humanlike face. The International Federation of Robotics uses a narrower statistical definition that requires a torso, arms, legged or wheeled mobility, humanlike perception, and autonomous operation, while an academic survey uses the broader idea of anthropomorphic form.^[1][2]

The humanlike robot form factor is intended to let a machine operate in spaces, at workstations, and around objects designed for people. That is an engineering objective, not a guarantee that any humanoid can climb ordinary stairs, manipulate human tools, or work safely beside people. The same morphology creates difficult control problems: the robot may have a high center of mass, a small support area, many coupled joints, and contacts that change as it walks and manipulates objects.

Humanoid robots range from remotely controlled research platforms to systems that execute a bounded workflow with human supervision. Morphology and autonomy are separate properties. A teleoperated machine can be humanoid in a research taxonomy even though it would be excluded from the IFR installation survey. As of August 2026, the field includes laboratory systems, industrial pilots, and reported deployments in bounded workflows, but the cited deployment records do not establish general-purpose autonomy.^[21][22] The IFR says mass adoption remains uncertain and expects humanoids to complement established robots rather than replace them.^[3] Reviewing China's robotics strategy in May 2026, the federation was blunter: humanoids are showcased to demonstrate competitiveness in embodied intelligence, but their "actual capabilities in real-world production scenarios are currently limited to demonstrators or pilot projects."^[30]

Definition and boundaries

There is no universally accepted minimum height, number of joints, exterior appearance, or autonomy level for a humanoid robot. Definitions instead combine several partly independent properties:

Scroll sideways for more →

A bipedal robot has two-legged mobility but may lack a torso, arms, or other humanlike features. An android is designed to resemble a human in outward appearance, while many humanoids expose mechanical structures and do not attempt visual realism. A social robot is classified by interaction role rather than body shape. It can be humanoid, animal-like, or non-figurative. An industrial robot and a service robot are classified mainly by use, so either class can include humanoid machines. The autonomy axis is separable enough that it supports its own taxonomy, described at humanoid robot autonomy levels.

The form factor is also distinct from the intelligence that controls it. A humanoid can use conventional state machines, optimization, machine learning, or a combination. Calling a robot humanoid does not establish that it understands language, generalizes to unfamiliar tasks, or possesses human-level cognition. Conversely, an artificial intelligence system does not become humanoid merely because it controls a voice or animated avatar. Marketing terms such as physical AI and embodied intelligence describe an ambition to connect learned models to bodies; they are not measurements of what a particular machine does.

The rationale for a humanlike body is strongest when the environment cannot be redesigned economically. Doorways, shelves, ladders, hand tools, controls, and work heights often reflect human reach and scale. A torso with arms and mobile support can potentially reuse that infrastructure. In a stable and repetitive process, however, a fixed manipulator, conveyor, autonomous mobile robot, or purpose-built machine may be faster, simpler, safer, or more energy-efficient. Humanoid morphology is therefore one design choice within robotics, not a universal endpoint.

Historical development

Humanoid robotics developed through a sequence of research platforms rather than one continuous product line. Waseda University began its WABOT project in the early 1970s and completed WABOT-1 in 1973. The university describes it as the first full-scale humanoid robot and reports that it could walk, grasp objects, and conduct simple communication in Japanese. WABOT-2, completed in 1984, was built to read a musical score and play an electronic organ.^[4]

Honda began biped research in 1986. Its early E-series machines studied walking, while the later P-series integrated legs, a torso, and arms. Honda introduced ASIMO in 2000 and subsequently demonstrated running, stair use, object carrying, and interaction functions. These were staged capabilities developed over multiple versions, not evidence that ASIMO could perform arbitrary household work.^[5]

Japan's Humanoid Robotics Project produced the HRP-2 prototype in 2002. The National Institute of Advanced Industrial Science and Technology reported a self-contained platform with 30 degrees of freedom that could lie down and get up, capabilities intended to support research in environments where falling and recovery matter.^[6] The iCub platform, described in 2010, took a different approach: it was an open research system designed for experiments in embodied cognition and developmental robotics. Its shared hardware and software supported reproducible work across several laboratories.^[7]

NASA and General Motors developed Robonaut 2 for work associated with human tools and spacecraft operations. It was launched to the International Space Station in 2011, becoming the first humanoid robot in space. Its original station work was largely experimental and did not make it an autonomous astronaut substitute.^[8]

The DARPA Robotics Challenge tested teams of robots and human supervisors on capabilities for future disaster response. For the 2015 finals, DARPA expected roughly 15 physical robot forms and deliberately degraded communication links to test partial autonomy, fall robustness, and battery management. The event showed the importance of mobility, manipulation, perception, and operator interfaces as a combined system.^[9]

These projects established recurring themes: dynamic bipedal locomotion, coordinated whole-body motion, safe contact, remote supervision, and the need to recover from errors. Later systems have added faster computation, better electric actuation, richer sensors, large-scale simulation, and learned control. The central challenge remains integration. A robot that performs one impressive motion under prepared conditions may still lack the reliability, task range, or recovery behavior needed for routine work. A fuller chronology of platforms and programs is kept at history of humanoid robots.

Mechanical architecture

A humanoid is a coupled mechanical system. Its links determine reach and mass distribution, its joints determine motion, and its actuators must produce enough force or torque while respecting temperature, speed, and energy limits. A count of degrees of freedom describes independent joint coordinates, not practical capability. Two robots with similar counts can differ greatly in payload, stiffness, sensing, control bandwidth, impact tolerance, and reachable workspace.

Most contemporary systems use electric motors with gear reductions because they are compact and can be powered from batteries. Hydraulic actuation can provide high power density, while pneumatic systems and artificial-muscle concepts appear in specialized research.^[2] Transmission choice affects friction, backlash, efficiency, compliance, and the accuracy of force control. Series elastic actuators deliberately place an elastic element between a motor and its load. The original 1995 work showed how this arrangement can improve force control and absorb shocks, with a tradeoff in motion bandwidth.^[10]

Transmission ratio and the impact problem

The gear ratio chosen for a joint is the single decision that most sharply divides a machine built for precise positioning from one built for dynamic contact. Inertia reflected through a transmission to the joint, which is what a collision must accelerate before any controller can respond, scales with the square of the reduction ratio. A high-ratio drive therefore buys stiffness and positioning accuracy at the price of making every impact more expensive.

Wensing and colleagues quantified this for legged machines and introduced a metric, the impact mitigation factor, to compare designs. Their proprioceptive actuator used a single-stage planetary reduction of 5.8:1; a model of the HUBO Plus humanoid built around 160:1 reductions carried roughly eight times the reflected inertia (about 0.085 against about 0.010 kg m2). Averaged across leg configurations, the low-ratio design retained about 90 percent of the impact mitigation of a hypothetical series-elastic version of itself; the high-ratio humanoid retained 52 percent.^[41] The low-ratio approach, now generally called quasi-direct drive, has a second benefit: joint torque can be estimated from motor current, which removes the need for a separate joint torque sensor or a physical spring. It is not free. The same paper notes that low ratios raise resistive heating for a given output torque, and concludes that the correct choice is application-dependent, since "a load-carrying robot walking at slow speeds could energetically benefit from a higher gear ratio in comparison to an agile robot running at high speeds."^[41]

The 2025 survey of humanoid locomotion and manipulation defines quasi-direct drive as a reduction below 10:1 and gives the same balance sheet from the other side: the advantages are backdrivability and high force-control bandwidth from reduced backlash and friction, and the penalty is that high torque demands high current, "which in turn causes strain on power electronics and overheating."^[2] A humanoid's torque budget is in that sense also a thermal budget.

Harmonic drive, or strain-wave, gearing occupies the opposite corner of that tradeoff, providing a high reduction with near-zero backlash in a compact, coaxial package. It is standard in industrial arms and in humanoids designed around position control, and its costs are poor backdrivability, friction losses, and sensitivity to shock loads. A series elastic actuator reaches good force control and impact tolerance by a third route, an explicit spring, at the cost of bandwidth. Real humanoids mix these choices by joint: an ankle that must survive heel strike and a wrist that must position a tool are not solved by the same actuator.

Mass distribution and remote actuation

Where the actuators sit matters as much as what they are. Rotational inertia in a limb sets the torque needed to swing it, so agile designs concentrate mass in the torso and drive distal joints remotely. The survey notes that the MIT Humanoid carries roughly 75 percent of its mass in the torso, achieved through belt drives and parallel mechanisms that actuate the ankle or knee from further up the leg. The costs are structural: four-bar transmissions tend to reduce joint range of motion, and belt drives add modeling and maintenance burden.^[2] A specification sheet that lists joint torques without saying where the motors are located leaves out much of what determines how the machine will move.

Mechanical compliance can reduce impact forces and improve contact sensing, but a compliant joint can also store energy and complicate precise positioning. Stiff transmissions can track position closely but may transmit higher collision forces unless sensing and control respond quickly. A design may combine rigid, elastic, and software-controlled compliance at different joints.

The sensing system usually has two roles. Proprioceptive sensors estimate the robot itself through joint encoders, motor currents, inertial measurement units, and force or torque sensors. Exteroceptive sensors observe the environment through cameras, depth sensors, microphones, range sensors, and touch. Whole-body tactile coverings, often described as robot skin, can localize contact outside the hands. Research published in 2015 covered the upper body of an HRP-2 with modular artificial skin and used contact feedback to adapt grasps of large, previously unknown objects. It illustrates how distributed sensing can support physical interaction rather than only object recognition.^[12]

Hands create a pronounced design tradeoff. Multi-fingered hands can reproduce more human grasps, but they add joints, cables or motors, sensors, control variables, and fragile contact surfaces. Simpler grippers are easier to protect and control but fit fewer tools and objects. The appropriate end effector depends on the workflow; humanlike appearance alone does not establish useful dexterity. The design space is treated separately at dexterous hand and humanoid robot hands.

Power and heat constrain every subsystem. Walking, balancing, computing, sensing, and manipulation draw from the same onboard supply. Larger batteries increase operating time but also add mass that the legs must move. Cooling hardware adds further mass and volume. The IFR's 2025 assessment identified battery duration and performance in speed, precision, reliability, and repeatability among current limitations.^[29] Two hardware responses are under way: embedding cells into structural components so the chassis carries energy rather than only mass, and moving to solid-state chemistries with higher energy density. The 2025 survey credits the latter with lifting operating times to figures such as four hours for Apptronik's Apollo and two hours for the Unitree G1, while noting that continued reliance on tethered power in laboratory work shows the constraint has not been removed.^[2]

Perception and state estimation

Humanoid control depends on estimating both the body and its contacts. Joint encoders describe relative angles but do not by themselves reveal the floating torso's position in the world. An inertial measurement unit measures angular motion and acceleration, while foot and wrist force sensors indicate contact loads. State estimators fuse these signals with kinematic and sometimes visual information to infer body orientation, velocity, foot contact, and drift.

Contact uncertainty is especially important. A foot may touch earlier than planned, slip, roll at an edge, or land on a compliant surface. A hand placed on a rail changes the support conditions. If the controller assumes a firm contact that does not exist, its balance calculation can be wrong even when individual sensors are functioning. Robust systems therefore detect contact transitions and update their model rather than treating a nominal gait as fixed.

Environmental perception supports motion planning, object manipulation, and interaction with people. Computer vision can estimate terrain, locate objects, or track a person, but performance depends on lighting, occlusion, camera motion, calibration, and the match between training data and the deployment environment. Depth sensing can improve geometry while still failing on reflective, transparent, distant, or partially hidden surfaces.

Perception for manipulation must be connected to action. Recognizing an object class is not enough to select a stable grasp, estimate its pose, predict whether it will move, and apply an appropriate contact force. Tactile sensing supplies information that vision cannot, including slip, contact force distribution, and whether a grasp has actually closed on an object, and the 2025 survey of humanoid locomotion and manipulation treats it as essential for contact-rich work.^[2] Likewise, speech recognition or language instruction does not establish that the requested physical action is safe or feasible. A task system needs to relate semantic goals to geometry, robot state, tool constraints, and allowed behaviors.

Balance, locomotion, and whole-body control

Standing and walking require the controller to keep the robot's motion compatible with available contact forces. In quasi-static motion, the projected center of mass can be kept within the support region. Dynamic motion also depends on momentum, inertia, contact timing, and ground reaction forces.

The zero moment point is a widely used concept for flat-foot contact. It identifies the point on the support surface where horizontal components of the contact moment vanish. Keeping the planned point within the feasible support area can prevent rotation about a foot edge under the model's assumptions. It is not a universal stability certificate: the concept becomes insufficient when contacts are slipping, highly dynamic, non-coplanar, or not well represented by a flat support polygon.^[13]

Kajita and colleagues used preview control in 2003 to generate center-of-mass motion from a future reference sequence of zero-moment points. Looking ahead lets the controller respond before a planned support change rather than correcting only after an error appears.^[14] Modern humanoid controllers also use model predictive control, centroidal or momentum models, trajectory optimization, and feedback policies. These are related tool families, not a single standard architecture.

Optimization-based work on the Atlas robot combined footstep planning, state estimation, whole-body planning, and control for locomotion over varied terrain. Its importance lies in the integration of multiple components and explicit constraints, not in proving that the resulting robot could operate without a specified model, sensors, or task plan.^[15]

Whole-body control coordinates more joints and contacts than are strictly necessary for one end-effector goal. A task hierarchy can give highest priority to constraints such as joint limits and established contacts, then control hand, foot, center-of-mass, gaze, or posture objectives in the remaining motion space. Sentis and Khatib formalized such hierarchical composition for simultaneous behavioral primitives and compliant interaction.^[16]

Balance recovery adds another layer. Small disturbances may be rejected by ankle or hip motion, while larger disturbances can require a step or an additional hand contact. Falling cannot always be prevented. Practical systems also need a controlled stop, fall detection, damage limitation, and a way to stand up or request help. Performance on a prepared floor does not imply equal performance on stairs, rubble, wet surfaces, moving platforms, or crowds.

Manipulation changes the balance problem because the arms move substantial mass and external forces pass through the body. Carrying an object shifts the combined center of mass. Pushing a door or inserting a part creates reaction forces. Useful locomotion and manipulation must therefore be planned together, often called loco-manipulation, rather than as independent leg and arm motions.

Learning, teleoperation, and data

Humanoid robots use a spectrum of control and data-collection methods. Teleoperation can provide direct remote control, shared control, demonstrations for learning, or recovery when autonomy fails. The H2O research system retargeted a person's motion to a full-size humanoid in real time and used reinforcement learning to help produce feasible whole-body motion. That result demonstrates a teleoperation method, not independent task autonomy.^[18]

Imitation learning trains a policy from demonstrations. HumanPlus combined whole-body shadowing with behavior cloning and reported learning selected tasks from tens of demonstrations. Its experiments show that a research platform can reuse human motion and collect real-robot examples efficiently; they do not establish unrestricted transfer to every object, operator, or environment.^[19]

Reinforcement learning can optimize locomotion or manipulation policies from rewards, often in simulation where many trials are cheaper and safer than on hardware. What changed the practical calculus was GPU-parallel simulation. Rudin and colleagues showed in 2021 that training thousands of simulated robots at once on a single workstation could produce a walking policy in under four minutes on flat ground and about twenty minutes on uneven terrain, a speedup of several orders of magnitude over earlier methods. The demonstration robot there was a quadruped, but the recipe, thousands of parallel simulated robots plus a terrain curriculum, carried over to bipeds.^[45]

Transfer works when the simulation is deliberately made uncertain. Domain randomization and system identification vary dynamics, latency, friction, mass, and observations so that a policy is forced to be robust rather than tuned to one physics model. Neither is free, and the cost is capability: He and colleagues observed in 2025 that these approaches "often rely on labor-intensive parameter tuning or result in overly conservative policies that sacrifice agility," and proposed instead learning a residual action model from real data and folding it back into the simulator, evaluated across two simulator-to-simulator transfers and a transfer to a physical Unitree G1.^[50] Radosavovic and colleagues trained a causal transformer over the history of proprioceptive observations and actions across an ensemble of randomized simulated environments and deployed it to a humanoid with no fine-tuning, reporting walking over varied outdoor terrain, robustness to external disturbance, and in-context adaptation without weight updates.^[46] Transfer of that kind is better established for locomotion than for contact-rich manipulation; the 2025 survey of the field frames the underlying difficulty as a tradeoff between model fidelity and computational efficiency and singles out tactile sensing as essential precisely for contact-heavy tasks, which are the ones a rigid-body simulator represents least well.^[2] A simulation still approximates actuators, contacts, sensors, and damage, so sim-to-real transfer requires separate real-world validation rather than an extrapolation from simulated scores.

HumanoidBench provides a simulated set of 27 whole-body locomotion and manipulation tasks in MuJoCo, on a Unitree H1 fitted with two dexterous Shadow Hands. Its 2024 evaluation found that state-of-the-art reinforcement learning struggled on most of them, while a hierarchical method built on robust low-level skills did better. The benchmark is useful for controlled comparison, but its results remain results in a specified simulator and task suite.^[17]

Robot foundation models

The most active research direction since 2024 has been the adaptation of foundation models to robot control. A vision-language-action model conditions actions on visual observations and a language instruction, so that a single network can in principle be pointed at many tasks and, with an embodiment identifier, at many robot bodies. The general pattern is described at robot foundation model.

A caveat belongs at the front of any summary of this work: most of the best-known generalist policies are not humanoid results. Physical Intelligence's pi-0.5, described in April 2025, was trained on roughly 400 hours of data from wheeled mobile manipulators working in about 100 home environments and evaluated in three real homes not seen during training, at ten trials per task. It is a strong claim about open-world generalization and it involves no legs.^[37] The same is true of most large manipulation datasets. Transfer from that literature to a walking machine is an assumption, not a finding.

The genuinely humanoid line of work is narrower. NVIDIA's GR00T N1, released in March 2025, paired a vision-language module with a diffusion transformer emitting continuous action chunks, at 2.2 billion parameters, trained on a mixture of real robot trajectories, human video, and synthetic data. On a Fourier GR-1 humanoid across ten trials per task, five for one industrial task, NVIDIA reported 82.0 percent on pick-and-place, 70.9 percent on articulated objects, 70.0 percent on industrial tasks, and 76.8 percent overall.^[20][40] Google DeepMind's Gemini Robotics, published in March 2025, trained on action data collected on ALOHA 2 bimanual arms; the humanoid appears only as an adaptation target, in what the paper itself calls "preliminary experiments" exploring transfer to a bi-arm Franka and to "Apollo from Apptronik, a full-size humanoid robot with five-fingered dexterous hands."^[52] Gemini Robotics 1.5, published in October 2025, added a motion-transfer mechanism intended to share behavior across robot embodiments and an intermediate reasoning step so a policy "thinks before acting."^[38]

These are laboratory results reported by the organizations that built the models, and the numbers they publish are the strongest available evidence of the gap between demonstration and reliability. Google DeepMind's Gemini Robotics 2, announced on 30 July 2026, is presented as controlling a humanoid from feet to fingertips rather than only an upper body. On an Apptronik Apollo 2, its published whole-body figures were 76.3 percent success picking from a shelf, 68.4 percent from a table, and 45.7 percent from the floor; on multi-finger dexterity tasks it reported 92 percent for unscrewing a bulb but 44 percent for tying a trash bag, 40 percent for a ziplock bag, 36 percent for screwing in a bulb, and 32 percent for using a dustpan. The developers' own summary is that "while Gemini Robotics 2 achieves a medium to high success rate for whole-body and gripper-based dexterous tasks, the multi-finger dexterous manipulation remains challenging."^[39] Numbers in that range are useful research progress and are far below what a production process normally tolerates.

One structural feature of this literature limits how far any of it generalizes: a great deal of it runs on the same small set of low-cost, openly sold Chinese platforms. HumanoidBench simulates a Unitree H1; the sim-to-real work above transferred to a physical Unitree G1; so did much of the whole-body control work published alongside it.^[17][50] Inexpensive research humanoids are what made the experiments possible at all, and they also mean that a finding presented as a property of humanoids may in part be a property of one 35 kg robot.

Data provenance matters because a learned controller inherits the coverage and errors of its examples. Human videos may not contain robot joint states or forces. Teleoperation data reflect the interface and operator. Simulation data reflect the simulator. Real-robot data are expensive and can overrepresent prepared environments. Cross-embodiment collections address the scarcity by pooling: Open X-Embodiment, assembled by 21 institutions and released in 2023, gathered demonstrations spanning 22 robot embodiments, 527 skills, and 160,266 tasks, and the accompanying RT-X models showed positive transfer, meaning experience from one platform improved another.^[43] Single-organization collection has since gone further in raw volume: AgiBot World Colosseo, released in March 2025, reported over one million trajectories across 217 tasks in five deployment scenarios, gathered through a standardized pipeline with human-in-the-loop verification.^[51] Pooling increases coverage but also mixes control rates, sensor suites, gripper types, and task conventions. It also does not address balance: manipulation trajectories recorded on a fixed or wheeled base contain no information about how a walking machine should keep its footing while performing the same motion. Dataset documentation should identify hardware versions, sensors, control frequency, task resets, failures, interventions, and train-test separation.

Evaluation and evidence

There is no single score for humanoid capability. Evaluation should separate the subsystem being tested, the test conditions, and the level of human assistance.

Scroll sideways for more →

Locomotion metrics can include speed, terrain, disturbance magnitude, falls, recovery rate, and energy use. Manipulation metrics can include completion rate, pose and force error, object variation, damage, and recovery after a failed grasp. System metrics include cycle time, intervention frequency, uptime, maintenance time, safety events, and performance across environmental changes. Reporting only successful trials hides the failure distribution.

The distinction between a skill and a workflow is essential. A robot may walk reliably while its object perception fails, or grasp objects while requiring a person to position them. A language-conditioned policy may select a task correctly but produce an unsafe trajectory. End-to-end evaluation must include task selection, physical execution, monitoring, recovery, and the human procedures around the machine.

Video evidence is the weakest common form and the most widely circulated. A recording cannot show how many attempts preceded it, whether the scene was arranged, whether the motion was scripted or learned, or whether a person was driving the robot. The default reading of an undocumented clip should be that the autonomy level is unstated, because vendors that do disclose it often disclose remote human involvement. 1X Technologies, for instance, states on its own product page for the NEO home robot that "for complex tasks NEO doesn't know, an Expert from 1X can remotely supervise its actions at scheduled times," alongside an offer to "pilot your NEO from anywhere in the world through your Mobile App and VR device."^[34] A consumer product marketed as "built for full autonomy" while documenting a remote-operator path is not evidence of household autonomy, and that disclosure is more informative than any demonstration reel.

What a video leaves out of frame can matter as much as what it shows. Reviewing industrial humanoid pilots in April 2026, A3 observed that these deployments run in "walled gardens," zones where "robot and human workers might share a workspace, but not at the same time," and that the striking feature of the promotional footage is "how rarely humanoids are operating in close proximity to humans."^[54] The lead researcher of the IEEE humanoid study group put it less diplomatically: industrial humanoids "are actually operating in semi-caged areas. You might not see the cage in the videos, but they're keeping people away from them, because even in industrial testing, they're not stable enough to approach."^[55]

The same caution applies to customer statements. BMW's June 2026 release about work at Spartanburg reports a bounded, repeated task and does not state the autonomy level at which the robot performed it.^[21] Absence of a claim is not a claim of absence, but it does mean the record supports "the task was completed under production conditions" rather than "the robot worked autonomously."

The shortage of repeatable protocols is a recognized problem inside the research community, not only a complaint about marketing. Reviewing reinforcement-learning controllers for humanoid standing and walking in 2024, van Marum and colleagues observed that prior work "lack[s] a clear method to systematically test new reward functions and compare controller performance through repeatable experiments," which "limits our understanding of the trade-offs between approaches and hinders progress," and proposed a low-cost quantitative protocol covering command following, disturbance recovery, and energy efficiency, validated on a Digit humanoid.^[44] Work of that kind is what turns a video into a measurement.

Where such audits have been done, they have not been kind to the benchmarks. Jiang and colleagues examined five widely used manipulation benchmarks in 2026 against four failure modes: shortcut solvability, lack of statistical significance, creeping overfitting, and data-source dependence. They report that two of the most-cited suites, LIBERO and CALVIN, fail multiple diagnostics; that on LIBERO "a 0.09B probe with no language encoder scores at or near reported SOTA" and most reported gains are not provably statistically significant; and that on CALVIN, randomizing block poses within the range already seen in training reduced the score of every policy tested.^[47] A benchmark that a tiny language-blind model can nearly saturate is not measuring language-conditioned manipulation, whatever its leaderboard says.

Sample sizes are the second structural weakness. A 2026 real-robot study characterized standard practice as "binary success rate at a fixed timeout with N less than or equal to 25 rollouts per condition, almost always without confidence intervals or paired statistical comparison," and proposed distributional alternatives instead: time-to-success curves and a throughput measure anchored to human teleoperation on the same fixture. Its headline result, on a fixed-base arm rather than a humanoid, is that the best model evaluated was roughly seven times slower per operation than the human reference.^[48] Speed is rarely reported alongside success rate, yet a robot that succeeds at human-comparable rates seven times more slowly is a different economic proposition from one that does not.

A third failure mode is specific to language-driven systems: they tend to attempt whatever they are asked. A 2026 benchmark of 6,069 instructions that were ambiguous, physically infeasible, or based on false premises found that an embodied planner, Gemini Robotics ER 1.6 Preview, declined to act on only 16.5 percent of them, against 39.0 percent for a general-purpose model on the same set; defensive prompting raised the planner to 90.8 percent and combining it with in-context learning reached 93.6 percent, but no method fully solved the problem.^[49] For a machine with mass and momentum, a policy that rarely refuses an impossible instruction is a safety property, not just an accuracy one.

Benchmark results should identify robot model, software version, sensors, payload, floor or terrain, task definition, number of trials, and intervention policy. Comparisons across different robots are unreliable when one result uses simulation, another uses a laboratory video, and a third reports operational hours. None of this is an argument that benchmarks are worthless. It is an argument that a benchmark number is a measurement of a policy on a task suite under a protocol, and that a claim of general physical intelligence is a different kind of statement requiring a different kind of evidence.

Applications and deployment status

Humanoids have long been used as research platforms for locomotion, manipulation, control, perception, and embodied AI. This role does not require a commercial business case. Shared platforms such as iCub and HRP systems let researchers compare algorithms on a common body, while specialized projects study disaster response, space operations, or human interaction.

Industrial interest concentrates on tasks in existing human-oriented facilities, including material handling, machine tending, parts sequencing, and repetitive manipulation. A humanoid can be attractive when fixed automation would require major layout changes or when one mobile platform might serve more than one station. The comparison must include not only purchase price but also integration, supervision, safety engineering, maintenance, energy, speed, and uptime.

Public evidence needs careful labels. BMW reported in June 2026 that a Figure 02 robot had supported production of more than 30,000 X3 vehicles over ten months by inserting sheet-metal parts in its Spartanburg body shop, and described a follow-on Figure 03 logistics sequencing project.^[21] This is an official customer statement about a bounded workflow, not proof that the robot could perform arbitrary factory jobs. GXO and Agility Robotics announced a multi-year robotics-as-a-service agreement in 2024 for Digit machines to move totes from cobots onto conveyors in a live logistics facility. Digit is described by the companies as a bipedal mobile manipulation robot, which illustrates the taxonomy boundary between a full humanoid and a task-oriented biped with arms.^[22] A site-by-site record of announced and verified placements, and the distinction between pilots, preorders, and paid operations, is maintained at humanoid robot deployments.

Professional-service concepts include inspection, hazardous-site support, and work in places built for people. Space and disaster projects show why human-compatible tools and access routes are appealing, but both also show the need for remote supervision and carefully bounded missions. Personal household assistance is a more open problem because homes contain unstructured objects, children, pets, stairs, liquids, and close physical contact. The IFR did not expect general household helpers to reach mass adoption in the near or medium term as of its 2025 assessment, and in May 2026 it placed broad commercialization of humanoid factory work toward the end of China's 2026 to 2030 plan period rather than early in it.^[3][30]

The appropriate question is not whether humanoids will replace all robots. It is whether a specific body and control system improves a defined workflow relative to purpose-built automation and human work, under measured safety and reliability requirements. Existing deployments establish feasibility for selected tasks, while the breadth, duration, and independence of evidence vary.

Cost and economics

Price information is unusually asymmetric in this field. Most companies developing full-size humanoids do not publish a price, and several sell only through pilots, service agreements, or reservation deposits, so the public record consists mainly of forecasts. The exception is the low end of the Chinese market, where robots are listed openly as research and education platforms. Unitree's own product pages, as of August 2026, illustrate how size, joint count, runtime, and price move together.

Scroll sideways for more →

Those are manufacturer list prices excluding tax and shipping, for one vendor, and they are not comparable to the delivered cost of an industrial system. They do establish something that forecasts cannot: electric humanoids under two metres are already being sold at prices closer to a car than to a traditional industrial robot cell, and the price scales with mass and joint count rather than with any measure of capability.

The labour-substitution argument that underwrites most humanoid forecasts requires more than a low purchase price. It requires that the machine work a large fraction of the hours a person would, at a comparable rate, on tasks that are actually assigned to it, with the supporting costs included. Those supporting costs are integration engineering, safety assessment and any required guarding, network and fleet infrastructure, spare parts and repairs for high-cycle joints and hands, energy, charging or battery swapping downtime, and the human staff who supervise, recover, and maintain the fleet. If a robot needs a remote operator for a share of its tasks, that operator's time is a direct cost of the deployment rather than a temporary artifact. A payback calculation is therefore sensitive to uptime and intervention rate, and both are the figures least often published.

Public counting is also weak. There is no audited, vendor-independent tally of humanoids shipped or in operation, and companies mix pilots, letters of intent, reservations, and delivered units in the same announcements. The IFR's statistical definition, which requires autonomous operation, would exclude a substantial part of what press coverage calls a humanoid deployment.^[1] Comparisons between a forecast unit total and a company's stated order book are therefore comparisons between two different things.

Forecasts from analysts who track shipments rather than valuations are correspondingly modest. A3 cited Omdia projections in April 2026 that global humanoid shipments could exceed 10,000 units by 2027 and reach 38,000 by 2030.^[54] For scale, the IFR reported in June 2026 that industrial robot installations in the United States alone rose 11 percent year on year to 38,000 units in 2025.^[59] A forecast that the whole world will install as many humanoids in 2030 as one country installs conventional industrial robots in a single year is not a forecast of labour replacement. Market sizing and forecast material is collected at humanoid robot market; this article treats those projections as claims about the future rather than as evidence about the present.

Safety, standards, and human interaction

Humanoid hazards include impact, crushing, trapping, sharp or hot surfaces, unstable loads, falls, unexpected motion, battery faults, and loss of communication or perception. A tall mobile machine can strike a person while walking or create a secondary hazard by dropping an object. Research on robot collisions has shown that injury risk depends on factors such as effective mass, velocity, contact geometry, and body region, so a low nominal speed alone is not a complete safety argument.^[11]

Safety is an application property, not only a robot feature. Risk assessment must cover the robot, end effector, payload, workspace, floor, tools, software, communications, people, maintenance, and foreseeable misuse. Safeguards can include physical separation, speed and force limits, monitored stops, safe zones, redundant sensing, protective coverings, emergency controls, access procedures, and operator training. A claim of collaborative operation needs evidence for the actual task and configuration. The general framework is set out at robot safety.

For a walking machine the layers are different from those of a caged arm, and no single one is sufficient. A3's 2026 account of what would be needed to work outside a separated zone lists a continuously updated environment model from depth cameras, lidar, and proximity sensors; force-limiting actuators that detect unexpected contact and shed torque before injury; active balance that can absorb a bump or stumble without redirecting the robot's mass toward a worker; and mechanical joint brakes as a last resort against a full collapse. On that account humanoid safety is "an emergent property of hardware, software, and environment working together" rather than a component that can be specified and bought.^[54]

As of August 2026, no published standard anywhere covers a legged humanoid operating among the general public. The reason is not neglect; it is that every published standard draws its scope boundary somewhere that excludes this machine.

ISO 10218-1:2025 and ISO 10218-2:2025, the third and second editions respectively, were published in February 2025 and cover industrial robots and industrial robot applications and cells. Part 1's foreword lists the main changes from the 2011 edition as additional design requirements, mode requirements, clarified functional safety, a robot classification into Class I and Class II, a test methodology for maximum force per manipulator on Class I robots, cybersecurity requirements "to the extent that it applies to industrial robot safety," and "incorporating safety requirements for industrial robots intended for use in collaborative applications (formerly, the content of ISO/TS 15066)."^[23][24][35] The introduction is precise about where that content landed and why: "Because human-robot collaboration relates to the application and not to the robot alone, most of the requirements of ISO/TS 15066 have been incorporated into ISO 10218-2:2025." It then drops the vocabulary entirely: "It is important to emphasize that the terms 'collaborative operation' and 'collaborative robot' are not used in this document. Only the application can be developed, verified and validated as a collaborative application."^[23] No machine, on that reading, is a collaborative robot in the abstract. ISO/TS 15066:2016 itself remains published, at the stage ISO marks "International Standard to be revised," and is to be replaced by a new ISO/AWI 15066-1 on biomechanical thresholds.^[53]

The scope exclusions matter more than the requirements for anyone assessing a humanoid. ISO 10218-1:2025 states that it is not applicable to "service robots, which provide a service to a person and as such where the public can have access" or to "consumer products, as this is household use to which the public can have access," and that it does not cover hazards related to "mobility when robots or manipulators are fixed to or part of mobile platforms" or situations "when the public, all ages or non-working adults have access."^[23] In other words the industrial robot standard governs the arm, not the legs, and not the crowd. A3, which publishes the United States adoption of these standards as ANSI/A3 R15.06-2025, states the consequence directly: "Neither of these standards cover the safe use of humanoids. An important distinction not covered within R15.06 or R15.08 is that a humanoid is a dynamically stable industrial mobile robot (DSIMR)."^[54]

That category, DSIMR, is where the engineering problem lives. A3 defines these as robots that "rely on dynamic management of their motions in order to maintain their stability. Put into simple terms, in the absence of power they will fall over," and draws the safety conclusion that follows: the usual machine response to an emergency stop, removing all power to the actuators, "is NOT the optimal response for a DSIMR, because that will create a new hazard due to the falling robot."^[54] A standard written for a machine that becomes safe when de-energized does not transfer to one that becomes dangerous.

The standard actually aimed at this class is ISO/CD 25785-1, "Robotics, Safety requirements for dynamically stable industrial mobile robots (legged, wheeled, or other forms of locomotion), Part 1: Robots." It defines "actively controlled stability" as a robot that "requires an active control in order to remain balanced and could become unstable in the absence of power," and covers quadrupedal, bipedal, and wheeled balancing machines. As of August 2026 it is a committee draft at ISO stage 30.60, close of comment period, and it too restricts itself to an "industrial environment," defined as "a workplace where the public is excluded or restricted from access," explicitly excluding "robots when used in non-industrial environments."^[27]

On the non-industrial side, ISO 13482:2014 covers personal-care robots in three categories, mobile servant, physical assistant, and person carrier, while excluding industrial and medical robots.^[25] Its replacement is still a draft. ISO/FDIS 13482, second edition, retitled "Robotics, Safety requirements for service robots," widens the scope to "service robots used in personal and professional/commercial applications," considers "the conditions for physical human-robot contact," and adds functional-safety material; as of August 2026 it sits at stage 50.00, final text registered for formal approval.^[26] A draft in the approval phase is not an effective international standard, and reports that a second edition has already been published are not supported by ISO's own record.

A separate ISO project, ISO/CD 26264-1, "Humanoid robot datasets, Part 1: General requirements," is at stage 30.20 and addresses dataset governance rather than physical safety: a lifecycle framework for the training, validation, and test datasets used to build humanoid policies, "regardless of their locomotion configuration."^[28] That an ISO committee is drafting dataset-governance rules before any published safety standard covers the machines themselves is a fair summary of where the field's attention has gone.

In North America, ANSI/A3 R15.06-2025 is the national adoption of ISO 10218:2025 plus a third part on the use of industrial robot cells, published in September 2025.^[54] ANSI/A3 R15.08 covers industrial mobile robots in three parts, the third, on use of IMR applications, published in April 2026.^[57] For non-industrial machines, ANSI/CAN/UL 3300 covers "Service, Communication, Information, Education and Entertainment (SCIEE) robots," with an ANSI publication date of April 2025. Its requirements "supplement the safety requirements for the intended, non-robotic function," and it excludes "Use in industrial environments, including training simulators for industrial applications" and "Use to treat, alleviate instability, or move individuals in hospitals, care facilities or in the home."^[56] It is a product-safety supplement, and nothing in its scope addresses dynamic stability or the hazard of a walking machine falling on someone.

Regulation, as distinct from voluntary standards, is moving faster on the learned-control question than on the legged one. Regulation (EU) 2023/1230 on machinery, which repeals the 2006 Machinery Directive, applies from 20 January 2027, as corrected by the corrigendum in Official Journal L 169 of 4 July 2023. It adds to Annex I, the list of categories that cannot be self-certified, both "safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions" and machinery embedding such systems, so those require third-party conformity assessment. The regulation is explicit that this applies to learning systems rather than to all software: the provision does "not apply to software incapable of learning or evolving, and programmed only to execute certain automated functions." Its essential health and safety requirements also state that control systems with self-evolving behaviour "shall not cause the machinery or related product to perform actions beyond its defined task and movement space," require the safety-related decision process to be logged, and require such machinery to communicate its planned actions, "what it is going to do and why," to operators comprehensibly.^[42] A humanoid whose behaviour comes from a policy updated after sale is squarely inside that language.

The gap is acknowledged inside the standards community. The IEEE Robotics and Automation Society convened a Study Group on Humanoid Robots whose stated purpose is to analyse "current standards that can and cannot be applied to humanoid robotics," identify the resulting gaps and the roadblocks to closing them, and produce a roadmap that standards development organizations can follow. It is a study group rather than a standards project, and it has issued no standard.^[36] Aaron Prather of ASTM International, its lead researcher, told Communications of the ACM in March 2026 why the ISO work is framed the way it is: "the title of its document doesn't even say 'humanoid' in it. Instead, it covers 'dynamically stable' robots. That's because it doesn't care if you have one leg, two legs, or four legs; if you're using dynamic stability to stay upright, you need to be aware of the safety issues under loss of power." The study group's survey of machines in development put their average mass at 83 kilograms, and his assessment was that "we're not going to have mass deployments of humanoids until they solve the dynamic stability issue."^[55]

For the home the position is starker still. A3's 2026 review concluded that "no standards have yet looked beyond industrial settings. The regulatory picture for home deployment is essentially a blank page. ISO 13482 covers personal care robots in general terms, and the EU AI Act classifies autonomous robots as high-risk AI systems that require human oversight, but neither was written with bipedal humanoids in mind."^[54]

Standardization is also becoming a policy instrument. China's Ministry of Industry and Information Technology established a Standardization Technical Committee on Humanoid Robots and Embodied Intelligence in December 2025, and in February 2026 released a "Standards System for Humanoid Robots and Embodied Intelligence (2026 Edition)" covering foundational standards, intelligence and computing, limbs and components, complete systems, applications, and safety and ethics, together with an initial list of 52 standards, according to SESEC, the European Union's standardization expert office in China.^[58]

Cybersecurity and privacy are part of physical safety. Remote control, fleet management, software updates, microphones, cameras, and stored demonstrations create attack and data-governance surfaces. A compromised command path can cause physical motion, while sensor data can reveal workers, homes, or confidential processes. The remote-operator arrangements that several vendors disclose make this concrete: a support path that lets an employee drive a robot inside a customer's home or plant is also a path an attacker would target. Access control, authenticated updates, logging, network segmentation, data minimization, and incident response must be designed for the deployment.

Human interaction adds perceptual and organizational risks. A humanlike body can make intent easier to infer, but it can also encourage users to overestimate understanding or reliability. Motion cues, lights, sound, distance, and stopping behavior should make the robot's state legible without implying capabilities it lacks. Workplace evaluation should also examine workload, training, job redesign, accessibility, and who is responsible when the robot pauses or fails.

Technical limitations

Humanoid robotics combines several hard problems whose failures interact. A perception error changes the plan, a plan changes contact forces, contact error threatens balance, and recovery consumes time and energy. A successful subsystem benchmark does not remove this coupling.

Reliability and recovery. Industrial equipment is expected to repeat tasks across long operating periods. Humanoids add moving joints, exposed limbs, variable contacts, and fall risk, and falls damage the machine as well as endangering bystanders. Describing the 2025 International Conference on Robotics and Automation, where many humanoids were demonstrated on an open show floor, Aaron Prather of ASTM International reported that within days "there were broken robots that had fallen, broken their arms or broken open their chests, everywhere."^[55] Mean time between interventions, repair time, restart behavior, and spare-part availability can matter more than peak demo performance.

Energy and thermal limits. Onboard batteries must power movement, sensing, communication, and computation. High-torque motion can heat motors and electronics, while added cooling and battery mass increase the load. The IFR's 2025 industry assessment said a battery cycle did not yet last a full working day.^[29] Manufacturer figures are consistent with that: the runtimes published for commercially listed platforms in 2026 are on the order of one to three hours, which makes battery swapping, docking, or duty cycling part of the workflow design rather than an optimization.^[31][32][33]

Manipulation and hand durability. Dexterous hands remain the weakest link between a capable body and useful work. Multi-fingered hands concentrate many small actuators, tendons, and sensors in the part of the robot that receives the most impacts and abrasion. The published success rates for multi-finger tasks in 2026, including 32 percent on sweeping with a dustpan and 36 percent on screwing in a bulb from one frontier laboratory's own report, indicate that fine manipulation outside structured settings is a research problem and not a deployment-ready capability.^[39] The same assessment is common among people who track the hardware. Prather's summary in March 2026 was that "human-level dexterity in robots is still not there," and that until manufacturers "figure out the manipulation issue and really get as close as possible to human-like dexterity, they do not have a viable product."^[55]

Speed, precision, and payload tradeoffs. Higher speed and payload increase forces and energy demand. Compliance that improves contact safety can reduce positioning bandwidth. Highly dexterous hands can be less robust than simple grippers. A robot optimized for gentle interaction may not match a guarded industrial machine on cycle time.

Generalization. Policies can fail when objects, lighting, floors, tools, or timing differ from training. Language instructions add ambiguity rather than removing physical constraints. Generalization claims require testing across predefined variations and unseen conditions, not a collection of selected demonstrations.

Data. Learned control is limited by the quantity, diversity, and documentation of robot experience. Unlike text and images, robot trajectories are not lying on the internet: they must be produced by teleoperators, scripted collection, or simulation, each of which imprints its own bias on the resulting policy. Human video is abundant but lacks joint states, forces, and the robot's own morphology, so its use requires an explicit retargeting or representation step.

Human support. Remote operators, safety staff, integration engineers, and maintenance teams may remain essential even when nominal task execution is automatic. Reported autonomy should disclose interventions, resets, offboard computing, environmental preparation, and whether a human chose grasps or recovery actions.

Economics and fit. A human-compatible body may reduce facility changes, but it may also carry mechanical complexity that a purpose-built system avoids. The correct comparison is task-specific and includes throughput, integration, supervision, downtime, safety, and lifecycle costs. Market forecasts cannot substitute for measured operational evidence.

The field is advancing in actuation, control, simulation, data collection, and learned policies. Its durable measure of progress is not visual similarity to a person or performance in one video. It is reproducible, safe, and maintainable operation on clearly defined tasks, with the limits of autonomy and evidence stated explicitly.

References

1. ^International Federation of Robotics. "Humanoid Robot Survey." IFR Statistical Department. ifr.org/...IFR_Questionnaire_on_Humaniods.pdf
2. ^Gu, Zhaoyuan, et al. "Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning." 2025. arxiv.org/...2501.02116
3. ^International Federation of Robotics. "Humanoid Robots: Vision and Reality Paper Published by IFR." August 14, 2025. ifr.org/...sion-and-reality-paper-published-by-ifr
4. ^Waseda University. "The Robots of Waseda: A 50-Year Journey in Humanoid Innovation." May 8, 2026. waseda.jp/...88080
5. ^Honda Motor Co. "History of Robotics Development." global.honda/...history
6. ^National Institute of Advanced Industrial Science and Technology. "AIST Successfully Develops a Laboring Humanoid Robot That Can Lie Down and Get Up." September 19, 2002. aist.go.jp/...20020919
7. ^Metta, Giorgio, et al. "The iCub Humanoid Robot: An Open-Systems Platform for Research in Cognitive Development." Neural Networks 23, no. 8-9, 2010. pubmed.ncbi.nlm.nih.gov/20864311
8. ^National Aeronautics and Space Administration. "Robonaut 2." nasa.gov/robonaut2
9. ^Defense Advanced Research Projects Agency. "Twenty-five Teams From around the World to Participate in DARPA Robotics Challenge Finals." March 5, 2015. darpa.mil/...robotics-challenge-finals
10. ^Pratt, Gill A., and Matthew M. Williamson. "Series Elastic Actuators." IEEE/RSJ International Conference on Intelligent Robots and Systems, 1995. doi.org/...IROS.1995.525827
11. ^Haddadin, Sami, Alin Albu-Schaffer, and Gerd Hirzinger. "Requirements for Safe Robots: Measurements, Analysis and New Insights." International Journal of Robotics Research 28, no. 11-12, 2009. journals.sagepub.com/...0278364909343970
12. ^Mittendorfer, Philipp, Eiichi Yoshida, and Gordon Cheng. "Realizing Whole-Body Tactile Interactions with a Self-Organizing, Multi-Modal Artificial Skin on a Humanoid Robot." Advanced Robotics 29, no. 1, 2015. doi.org/...01691864.2014.952493
13. ^Vukobratovic, Miomir, and Branislav Borovac. "Zero-Moment Point: Thirty Five Years of Its Life." International Journal of Humanoid Robotics 1, no. 1, 2004. doi.org/...S0219843604000083
14. ^Kajita, Shuuji, et al. "Biped Walking Pattern Generation by Using Preview Control of Zero-Moment Point." IEEE International Conference on Robotics and Automation, 2003. doi.org/...ROBOT.2003.1241826
15. ^Kuindersma, Scott, et al. "Optimization-Based Locomotion Planning, Estimation, and Control Design for the Atlas Humanoid Robot." Autonomous Robots 40, 2016. doi.org/...s10514-015-9479-3
16. ^Sentis, Luis, and Oussama Khatib. "Synthesis of Whole-Body Behaviors through Hierarchical Control of Behavioral Primitives." International Journal of Humanoid Robotics 2, no. 4, 2005. doi.org/...S0219843605000594
17. ^Sferrazza, Carmelo, et al. "HumanoidBench: Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation." Robotics: Science and Systems, 2024. roboticsproceedings.org/...p061
18. ^He, Tairan, et al. "Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation." IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024. arxiv.org/...2403.04436
19. ^Fu, Zipeng, et al. "HumanPlus: Humanoid Shadowing and Imitation from Humans." Conference on Robot Learning, 2024. arxiv.org/...2406.10454
20. ^NVIDIA GEAR Team. "NVIDIA Isaac GR00T N1: An Open Foundation Model for Humanoid Robots." March 17, 2025. research.nvidia.com/...ation-model-humanoid-robots
21. ^BMW Group. "BMW Group Advances the Use of Physical AI in Production with Figure 03 Project in Spartanburg." June 25, 2026. press.bmwgroup.com/...re-03-project-in-spartanburg
22. ^GXO Logistics. "GXO Signs Industry-First Multi-Year Agreement with Agility Robotics." June 27, 2024. investors.gxo.com/...ar-agreement-agility-robotics
23. ^International Organization for Standardization. "ISO 10218-1:2025, Robotics, Safety Requirements, Part 1: Industrial Robots." February 2025. iso.org/...73933
24. ^International Organization for Standardization. "ISO 10218-2:2025, Robotics, Safety Requirements, Part 2: Industrial Robot Applications and Robot Cells." February 2025. iso.org/...73934
25. ^International Organization for Standardization. "ISO 13482:2014, Robots and Robotic Devices, Safety Requirements for Personal Care Robots." February 2014. iso.org/...53820
26. ^International Organization for Standardization. "ISO/FDIS 13482, Robotics, Safety Requirements for Service Robots." iso.org/...83498
27. ^International Organization for Standardization. "ISO/CD 25785-1, Robotics, Safety Requirements for Dynamically Stable Industrial Mobile Robots, Part 1: Robots." iso.org/...91469
28. ^International Organization for Standardization. "ISO/CD 26264-1, Humanoid Robot Datasets, Part 1: General Requirements." iso.org/...93011
29. ^International Federation of Robotics. "Humanoid Robots: Vision and Reality." 2025. ifr.org/...Humanoids_Position_Infograph_2025.pdf
30. ^International Federation of Robotics. "China Makes AI-powered Robots Core of National Strategy." May 5, 2026. ifr.org/...owered-robots-core-of-national-strategy
31. ^Unitree Robotics. "Unitree G1 Humanoid Agent." Product page, accessed August 2026. unitree.com/g1
32. ^Unitree Robotics. "Unitree H2." Product page, accessed August 2026. unitree.com/H2
33. ^Unitree Robotics. "Unitree R1." Product page, accessed August 2026. unitree.com/R1
34. ^1X Technologies. "NEO." Product page, accessed August 2026. 1x.tech/neo
35. ^The Robot Report. "ISO 10218 Industrial Robot Safety Standard Receives Major Overhaul." February 18, 2025. therobotreport.com/...dard-receives-major-overhaul
36. ^IEEE Robotics and Automation Society. "Study Group: Humanoid Robots." ieee-ras.org/...study-group-humanoid-robots
37. ^Physical Intelligence. "pi-0.5: A Vision-Language-Action Model with Open-World Generalization." April 22, 2025. arxiv.org/...2504.16054
38. ^Gemini Robotics Team, Google DeepMind. "Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer." October 2, 2025. arxiv.org/...2510.03342
39. ^Google DeepMind. "Gemini Robotics 2 Brings Whole-Body Intelligence to Robots." July 30, 2026. deepmind.google/...ole-body-intelligence-to-robots
40. ^NVIDIA. "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots." March 18, 2025. arxiv.org/...2503.14734
41. ^Wensing, Patrick M., Albert Wang, Sangok Seok, David Otten, Jeffrey Lang, and Sangbae Kim. "Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth Physical Interaction for Dynamic Legged Robots." IEEE Transactions on Robotics 33, no. 3, 2017, pp. 509-522. doi.org/...TRO.2016.2640183
42. ^European Parliament and Council. "Regulation (EU) 2023/1230 of 14 June 2023 on Machinery and Repealing Directive 2006/42/EC." Official Journal L 165, June 29, 2023. eur-lex.europa.eu/...HTML
43. ^Open X-Embodiment Collaboration. "Open X-Embodiment: Robotic Learning Datasets and RT-X Models." 2023. arxiv.org/...2310.08864
44. ^van Marum, Bart, Aayam Shrestha, Helei Duan, Pranay Dugar, Jeremy Dao, and Alan Fern. "Revisiting Reward Design and Evaluation for Robust Humanoid Standing and Walking." IEEE/RSJ International Conference on Intelligent Robots and Systems, 2024. arxiv.org/...2404.19173
45. ^Rudin, Nikita, David Hoeller, Philipp Reist, and Marco Hutter. "Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning." Conference on Robot Learning, 2021. arxiv.org/...2109.11978
46. ^Radosavovic, Ilija, Tete Xiao, Bike Zhang, Trevor Darrell, Jitendra Malik, and Koushil Sreenath. "Real-World Humanoid Locomotion with Reinforcement Learning." 2023. arxiv.org/...2303.03381
47. ^Jiang, Tianchong, Xiangshan Tan, Samuel Wheeler, Luzhe Sun, Tewodros W. Ayalew, and Matthew Walter. "What Are We Actually Benchmarking in Robot Manipulation?" June 2, 2026. arxiv.org/...2606.04233
48. ^Arkhangelskiy, Sergey. "PhAIL: A Real-Robot VLA Benchmark and Distributional Methodology." May 28, 2026. arxiv.org/...2605.29710
49. ^Yeke, Doguhan, Elif Su Temirel, Ananth Shreekumar, Brandon Lee, Dongyan Xu, and Z. Berkay Celik. "The Yes-Man Syndrome: Benchmarking Abstention in Embodied Robotic Agents." May 19, 2026. arxiv.org/...2605.20544
50. ^He, Tairan, et al. "ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills." February 3, 2025. arxiv.org/...2502.01143
51. ^AgiBot-World Contributors. "AgiBot World Colosseo: A Large-Scale Manipulation Platform for Scalable and Intelligent Embodied Systems." March 9, 2025. arxiv.org/...2503.06669
52. ^Gemini Robotics Team, Google DeepMind. "Gemini Robotics: Bringing AI into the Physical World." March 25, 2025. arxiv.org/...2503.20020
53. ^International Organization for Standardization. "ISO/TS 15066:2016, Robots and Robotic Devices, Collaborative Robots." February 2016. iso.org/...62996
54. ^McCarthy, Dan. "Safety by Design: How Humanoid Robots Must Evolve to Depart the Walled Garden." Association for Advancing Automation (A3), April 23, 2026. automate.org/...evolve-to-depart-the-walled-garden
55. ^Marks, Paul. "Setting Standards for Humanoid Robots." Communications of the ACM, March 23, 2026. cacm.acm.org/...ting-standards-for-humanoid-robots
56. ^UL Standards and Engagement. "ANSI/CAN/UL 3300:2024, Service, Communication, Information, Education and Entertainment Robots (SCIEE Robots)." webstore.ansi.org/...ansiul33002024
57. ^American National Standards Institute. "ANSI/A3 R15.08-3-2026, Industrial Mobile Robots, Safety Requirements, Part 3: Use of IMR Applications." April 2026. webstore.ansi.org/...ansia3r15082026
58. ^Seconded European Standardization Expert in China (SESEC). "China's First Standards System for Humanoid Robots and Embodied Intelligence." 2026. sesec.eu/...anoid-robots-and-embodied-intelligence
59. ^International Federation of Robotics. "US Robot Industry Returns to Double Digit Growth." June 18, 2026. ifr.org/...industry-returns-to-double-digit-growth