Humanoid robot

A humanoid robot is a robot designed to resemble and move like a human being, typically featuring a head, torso, two arms, and two legs arranged in a bipedal configuration. The defining characteristic of a humanoid is not just visual resemblance but functional mimicry: these robots are built to operate in environments designed for people, using tools made for human hands, navigating doorways and staircases sized for human bodies, and interacting with humans in intuitive ways. The term covers a wide range of machines, from full-sized autonomous bipeds to upper-body platforms mounted on wheeled bases.

Humanoid robots have been a fixture of science fiction for over a century, from Karel Capek's 1920 play R.U.R. (which coined the word "robot") to the androids of modern film and television. For most of that time, real humanoid robots remained impractical curiosities, limited by inadequate actuators, insufficient computing power, and the sheer difficulty of bipedal balance. That changed in the 2020s. Breakthroughs in artificial intelligence, cheaper hardware, and billions of dollars in venture capital have pushed humanoid robots from lab demonstrations to factory floors. As of early 2026, multiple companies are manufacturing humanoid robots for commercial customers, and Goldman Sachs projects the global humanoid robot market could reach $38 billion by 2035 [1].

taxonomy and definitions

The term humanoid robot covers a broad family of machines that share an anthropomorphic body plan but differ substantially in form and purpose. Roboticists generally distinguish among several categories:

Category	Description	Examples
Full humanoid	Two arms, two legs, bipedal locomotion, head with sensors	Atlas, ASIMO, Tesla Optimus, Figure 02, Apollo
Semi-humanoid	Upper body on a wheeled or tracked base, no legs	Pepper, Galbot G1, Reachy, Robonaut 1
Android	Humanoid with realistic skin, hair, and facial features intended to closely resemble a person	HRP-4C, Geminoid (Hiroshi Ishiguro Lab), Ameca
Gynoid	Android with female appearance	HRP-4C "Miim", Sophia
Mini humanoid	Small (typically under 1 m) bipedal platforms used for research and education	NAO, Robotis OP3, Unitree G1

The term android has a narrower meaning in robotics than in everyday speech: it refers specifically to humanoids that imitate human appearance, while a robot like Atlas, with no skin and clearly mechanical features, is a humanoid but not an android. Some researchers reserve the word humanoid for robots with bipedal locomotion and treat fixed-base or wheeled "upper body humanoids" as a separate category.

history

early concepts and automata

Human-shaped machines predate electronics by centuries. Eighteenth-century clockmakers such as Pierre Jaquet-Droz built elaborate mechanical automata that could write, draw, and play music, although they had no autonomy in the modern sense. The word "robot" itself entered English from the 1920 Czech-language play R.U.R. (Rossum's Universal Robots) by Karel Capek, derived from the Czech robota, meaning forced labor or drudgery. Capek credited his brother Josef with suggesting the term, replacing the Latin-derived labori he had originally considered [14]. The play imagined factory-built artificial workers and gave the science fiction tradition the vocabulary it has used ever since.

The modern engineering effort to build walking, sensing, human-shaped machines began in the 1960s and 1970s, primarily in Japan and the United States. Hydraulic walking research at Ohio State University and General Electric in the late 1960s produced legged machines that could carry a human operator, but no practical autonomous bipeds.

WABOT and Japanese pioneers

Waseda University's WABOT-1, unveiled in 1973 under the direction of Professor Ichiro Kato, is generally considered the first full-scale anthropomorphic robot. The machine stood roughly 1.6 meters tall, weighed about 100 kg, and was built from aluminum covered with a plastic shell. It had six degrees of freedom per arm, a vision system that could measure distances, an artificial voice that spoke simple Japanese, and a quasi-dynamic walking system. WABOT-1's first steps were extremely slow: a single step took about 45 seconds and covered roughly 10 cm of ground [2][15]. The robot demonstrated for the first time that the integration of locomotion, manipulation, vision, and speech in a single human-shaped machine was feasible.

WABOT-2, introduced in 1984, could read sheet music with a camera, recognize the notation, and play an electronic organ with both hands and feet. The same year, Kato's laboratory demonstrated the WL-10RD, the first dynamically balanced biped, which served as the first practical demonstration of the zero moment point (ZMP) concept that had been introduced by the Yugoslav engineer Miomir Vukobratovic and Davor Juricic in 1968. The ZMP gives a precise condition for dynamic stability of a walking robot and remained the dominant theoretical framework for biped control for the next three decades [16][17].

In the 1990s, the Japanese Humanoid Robotics Project (HRP), funded by the Ministry of Economy, Trade and Industry (METI) and led by Kawada Industries with the National Institute of Advanced Industrial Science and Technology (AIST) and Kawasaki Heavy Industries, advanced bipedal research toward practical platforms. HRP-1 was based on three Honda P3 robots licensed by the project. Subsequent models added new capabilities: HRP-2 "Promet" (2002) could stand up after lying flat on the floor; HRP-3 (2006) was sealed against dust and rain; HRP-4 (2010) reduced weight to 39 kg with a more slender form factor; HRP-4C "Miim" (2009) was a feminine-looking humanoid built for entertainment, with eight motors driving facial expressions and the ability to sing and dance alongside human performers; HRP-5P (2018) demonstrated drywall installation, showing that humanoids could perform construction tasks [18][19].

Honda P-series and ASIMO

Honda began its humanoid research program in 1986 with the E-series of experimental walking robots. The E0 through E6 prototypes (1986 to 1993) explored static, then quasi-dynamic, then dynamic bipedal locomotion. The first humanoid prototype with arms and torso, P1, appeared in 1993. P2, unveiled in December 1996, became the first self-contained, wireless, autonomous bipedal walking robot. It stood 1.82 m tall, weighed 210 kg, and could walk, climb stairs, push carts, and turn while walking. P3 (1997) was smaller (1.6 m, 130 kg) and improved walking smoothness. The P-series program led directly to ASIMO [3][20].

ASIMO (Advanced Step in Innovative Mobility) debuted in 2000 and became the world's most recognizable humanoid robot. Standing 130 cm tall and weighing 54 kg, ASIMO could walk, run (up to 9 km/h), climb stairs, recognize faces and gestures, and carry objects with both hands. An updated version in 2011 added the ability to run backward, hop on one foot, and pour drinks [3].

ASIMO traveled the world as a technology demonstrator, rang the opening bell at the New York Stock Exchange in 2002, and conducted the Detroit Symphony Orchestra. Despite its fame, ASIMO was never a commercial product. Honda discontinued development in 2018 and retired ASIMO in March 2022, citing a shift toward avatar-style telepresence robotics. ASIMO's real legacy was proving that bipedal humanoid locomotion was achievable with 2000s-era technology and inspiring an entire generation of robotics researchers [3].

Boston Dynamics and dynamic locomotion

Boston Dynamics, founded in 1992 as a spin-off from MIT, took a different approach to humanoid robotics. While ASIMO focused on controlled, flat-surface walking, Boston Dynamics pursued dynamic locomotion: robots that could handle rough terrain, recover from pushes, and perform acrobatic maneuvers.

The company's Atlas robot, first unveiled in 2013 as a hydraulic research platform for the DARPA Robotics Challenge, became famous through YouTube videos showing it performing backflips, parkour, and dance routines. The hydraulic Atlas demonstrated capabilities that no other humanoid could match in terms of agility and dynamic balance. However, it was extremely expensive, heavy (89 kg), and not designed for commercial use [4].

Boston Dynamics was acquired by Google's parent company Alphabet in 2013, then sold to SoftBank in 2017, and finally acquired by Hyundai Motor Group in 2021 for roughly $1.1 billion. Under Hyundai's ownership, the company shifted toward commercial viability [4].

DARPA Robotics Challenge

The DARPA Robotics Challenge (DRC), held from 2012 to 2015 in response to the Fukushima Daiichi nuclear disaster, accelerated humanoid robotics research more than any other single event of the decade. Teams competed to build robots that could enter a degraded human environment, drive a vehicle, open doors, climb a ladder, cut through a wall, and complete other emergency response tasks. The 2015 DRC Finals at Fairplex in Pomona, California saw Team KAIST from South Korea win with their DRC-HUBO robot, completing all eight tasks in 44 minutes 28 seconds and claiming the $2 million grand prize. DRC-HUBO's distinctive feature was a transformer-style ability to kneel onto wheels mounted at the knees, switching between walking and rolling locomotion [21]. Many of the engineers and ideas that emerged from the DRC went on to found the commercial humanoid wave of the 2020s.

service and educational platforms

A parallel strand of humanoid robotics has focused on smaller, lower-cost platforms for research, education, and consumer service. The NAO robot, designed by the French company Aldebaran Robotics from 2004 onward, became the most widely deployed research humanoid in history: as of 2024, more than 13,000 NAO units were in use across more than 70 countries, primarily in universities and schools. NAO stands 58 cm tall, weighs 5.6 kg, and has 25 degrees of freedom [22]. Aldebaran's larger Pepper robot, introduced in 2014 after SoftBank's acquisition, was a 120 cm tall semi-humanoid on a wheeled base, designed for retail and reception roles. SoftBank produced roughly 27,000 Pepper units before pausing production in June 2021 due to weak demand [23].

The iCub, developed by the RobotCub Consortium and built by the Italian Institute of Technology (IIT) starting in 2008, is a 104 cm child-sized humanoid intended as an open-source platform for cognitive science research. It has 53 degrees of freedom, tendon-driven hand and shoulder joints, and capacitive tactile skin. More than 40 iCub units have been delivered to laboratories worldwide, each costing around 250,000 euros [24].

NASA's Robonaut program produced two generations of humanoid for space operations. Robonaut 2, developed jointly with General Motors, became the first humanoid robot in space when it was launched on Space Shuttle Discovery (STS-133) in February 2011 and installed in the International Space Station. R2 was deployed initially as a torso-only unit fixed to a stanchion and was later upgraded with climbing legs. A hardware fault forced its return to Earth in 2018 for repairs [25].

the humanoid form factor: why build robots that look like people?

A common question about humanoid robots is: why give them a human shape at all? Wheeled robots are more stable. Robotic arms on fixed bases are more precise. The answer lies in the built environment and the economics of deployment.

Human civilization has spent millennia designing buildings, tools, vehicles, and infrastructure around the human body. Doorways are sized for humans. Stairs are built for bipeds. Tools have handles shaped for human hands. Factory workstations are arranged at human heights. Retrofitting all of this infrastructure for a different robot form factor would be enormously expensive.

A humanoid robot can, in principle, drop into any workspace or home designed for people without modification. It can use the same tools, operate the same equipment, and navigate the same spaces. This is a massive economic advantage, because it means the robot adapts to existing infrastructure rather than requiring the infrastructure to adapt to the robot [5].

Additional arguments for the humanoid form include:

Human interaction. People find it more natural to interact with, instruct, and collaborate alongside a robot that has a recognizable human shape. Gestures, eye contact, and body language transfer intuitively.
Versatility. A humanoid body plan is a general-purpose platform. Two arms with dexterous hands can perform an enormous range of tasks, from lifting boxes to threading needles.
Data leverage. Billions of hours of video of humans performing tasks exist on the internet. A robot with a similar body plan can learn from this data more directly than a robot with a fundamentally different morphology.

Critics counter that bipedal locomotion is inherently less stable and energy-efficient than wheeled or tracked motion, and that the complexity of a humanoid body adds cost and failure points. The market is effectively running the experiment: if humanoids can deliver enough versatility to offset their complexity, they will succeed; if not, more specialized form factors will win.

technical components

actuators

The joints of a humanoid robot are driven by actuators that convert electrical or hydraulic energy into rotational or linear motion. Three families dominate modern designs:

Type	Principle	Strengths	Weaknesses	Examples
Hydraulic	Pressurized fluid drives pistons	Extremely high power density, fast response	Heavy, leaky, requires pumps and reservoirs, energy inefficient	Atlas (2013 hydraulic), Sanctuary AI Phoenix hands
Series elastic actuator (SEA)	Motor and gearbox in series with a compliant spring	Force sensing via spring deflection, passive shock absorption, safe contact	Lower control bandwidth, mechanical complexity	Older Boston Dynamics legs, NASA Valkyrie
Quasi-direct drive (QDD)	High-torque motor with low-ratio gearbox (typically 6:1 to 30:1)	Low reflected inertia, backdrivable, impact-tolerant	Larger motors needed, higher current draw	Unitree H1, MIT Mini Cheetah lineage, Tesla Optimus

QDD actuators have become the dominant choice for the 2020s wave of humanoids because the low gear ratio reduces reflected motor inertia by the square of that ratio, making the joint compliant under impact and easier to control with learned policies. The combination of high-pole-count brushless motors and low-ratio cycloidal or planetary gearboxes lets the joint act almost like a torque-controlled muscle [26].

Tesla Optimus uses a mix of harmonic and planetary drive systems. Boston Dynamics' electric Atlas uses fully electric direct-drive actuators with industry-leading torque density of around 220 Nm/kg [27]. Sanctuary AI's Phoenix retains a hydraulic strategy, but only in its 21-degree-of-freedom hands, where miniaturized hydraulic valves provide high power density in a compact volume.

sensors

A humanoid integrates a fusion of sensors:

Vision. Multiple cameras, often including stereo pairs, depth (RGB-D) sensors, and event cameras. Boston Dynamics Atlas uses NVIDIA Jetson Thor compute and an array of stereo cameras for perception. Tesla Optimus uses cameras only, with no LiDAR, mirroring the philosophy of Tesla's Full Self-Driving system.
3D LiDAR. Used by Unitree H1 and G1 (LIVOX MID-360) and several Chinese platforms for mapping and obstacle detection.
Inertial measurement units (IMUs). Provide angular velocity and linear acceleration for balance control. Typically located in the torso and sometimes in feet.
Force/torque sensors. Mounted at the wrists and ankles to measure interaction forces with the environment, essential for both walking and manipulation.
Tactile sensors. Distributed across fingertips and palms. Sanctuary AI's Phoenix hands include sensors that detect pressure as low as 5 millinewtons, comparable to human fingertip sensitivity. Figure 03's fingertip sensors detect forces as small as 3 grams [8][28].
Audio. Multi-microphone arrays for speech recognition and sound localization.

compute

Onboard compute has expanded dramatically with the shift to neural network policies. The original ASIMO ran on a Pentium-class processor with kilowatts of external support; the Boston Dynamics electric Atlas runs on the NVIDIA Jetson Thor platform, delivering roughly 800 teraflops of AI performance. The Xpeng Iron robot is powered by three proprietary Turing AI chips delivering a combined 2,250 TOPS, while the Unitree H2 uses a 2,070-TOPS onboard chip [29][30][31]. This compute is used for vision-language-action models, model predictive control, sensor fusion, and motion planning, all running at hundreds of hertz.

batteries and power

Most current humanoids are powered by lithium-ion battery packs ranging from roughly 0.8 to 2.5 kWh. Operating times vary widely with workload: a humanoid standing still and computing draws a few hundred watts, while one walking quickly and lifting heavy loads can pull more than a kilowatt. 1X NEO's 842 Wh battery delivers about four hours of operation; Tesla Optimus Gen 3 carries a 2.3 kWh pack rated for 8 to 10 hours; Boston Dynamics Atlas uses dual swappable batteries for around four hours of runtime; Apptronik Apollo offers a five-minute battery hot-swap for continuous operation [27][32][33]. Xpeng has announced plans to use full solid-state batteries in its Iron robot, which would substantially improve safety and energy density if delivered at production scale [29].

locomotion control

Bipedal walking is, in mathematical terms, the controlled fall of an inverted pendulum with non-trivial mass distribution and intermittent contact. The dominant frameworks for stabilizing this fall include:

Zero moment point (ZMP). A point on the ground at which the net horizontal moment of inertia and gravity forces equals zero. Walking is dynamically stable as long as the ZMP remains within the convex hull of the support polygon defined by the foot or feet on the ground. The ZMP framework was introduced by Vukobratovic in 1968 and remains the basis for many classical biped controllers, including ASIMO and the HRP series [16][17].
Linear inverted pendulum model (LIPM). A reduced-order model that approximates the robot as a point mass on a massless leg, used to plan center-of-mass trajectories that satisfy ZMP constraints. Widely used for footstep planning.
Whole-body control (WBC). A hierarchical optimization approach that solves a quadratic program at each control cycle (typically 1 kHz) to compute joint torques that satisfy multiple weighted or strictly prioritized objectives, such as tracking a desired hand pose, maintaining balance, and respecting contact constraints. Atlas, Apollo, and other modern humanoids use WBC layered with model predictive control [34].
Model predictive control (MPC). Plans contact sequences and center-of-mass trajectories over a receding horizon (typically 0.5 to 2 seconds). The combination of MPC over a simplified dynamics model with WBC over the full dynamics has become the model-based standard for advanced humanoids.
Reinforcement learning (RL) policies. Train a neural network to map states to joint targets through millions of simulated trials, often with massive domain randomization. RL policies have largely replaced classical controllers for low-level locomotion on platforms such as the Unitree H1 and the electric Atlas.

end-effectors and dexterity

The robotic hand is the most challenging mechanical subsystem. Modern end-effector designs converge on between 11 and 22 degrees of freedom per hand, driven by tendon, geared, or hydraulic actuators:

Robot	DOF per hand	Drive	Notable feature
Tesla Optimus Gen 3	22	Tendon	25 actuators per forearm/hand
1X NEO	22	Tendon	95% backdrivable, soft enclosure
Sanctuary AI Phoenix	21	Hydraulic	5 mN tactile sensitivity
Xpeng Iron	22	Tendon	Solid-state battery, 22 DOF
Figure 03	16	Tendon	3-gram fingertip force resolution
Fourier GR-2	12	Tendon	Six tactile sensors per fingertip

Even these advanced hands fall short of human dexterity for tasks like tying shoelaces, peeling soft fruit, or threading a needle. The Shadow Robot Company's Dexterous Hand (DEX-EE), the most sensorized commercial robotic hand, has hundreds of tactile sensors but is too heavy and expensive for integration into a full humanoid.

current humanoid robots

As of early 2026, more than two dozen companies are actively developing humanoid robots. The following table summarizes the most prominent:

Company	Robot	Height / Weight	DOF	Key features	Status (early 2026)
Boston Dynamics	Atlas (Electric)	~150 cm / ~90 kg	56	50 kg lift capacity; dual swappable batteries (4 hr runtime); NVIDIA Jetson Thor compute	Production version launched at CES 2026; commercial deployments at Hyundai and Google DeepMind
Figure AI	Figure 02 / Figure 03	167 cm / 60 kg	Not disclosed	Full-stack AI; robot-as-a-service model ($1,000/month)	Figure 02 deployed at BMW (90,000+ parts loaded); Figure 03 unveiled Oct 2025 for home environments
Tesla	Optimus Gen 3	173 cm / 57 kg	50+ (hands: 22 DOF, 25 actuators per hand)	End-to-end neural network control; self-supervised learning from factory data	Gen 3 production beginning summer 2026 at Fremont; currently used for internal data collection
1X Technologies	NEO	165 cm / 30 kg	22 DOF hands	Soft-body design (3D lattice polymer); 150 lb lift capacity; 22 dB noise level	Pre-orders at $20,000; U.S. early access delivery starting 2026
Unitree Robotics	H1 / G1 / H2	H1: 180 cm / 47 kg; G1: 127 cm / 35 kg; H2: 182 cm / 70 kg	H1: 19; G1: 23-43; H2: 31	H1 running speed 3.3 m/s; G1 starting at ~$16,000; H2 at $29,900	Commercially available; $1.3B unicorn valuation (Jun 2025)
Agility Robotics	Digit	175 cm / 65 kg	16+	Purpose-built for logistics; RoboFab manufacturing facility in Salem, Oregon	Deployed at Amazon and GXO; 100,000+ totes moved at GXO facility
Apptronik	Apollo	172 cm / 73 kg	Not disclosed	Force-control emphasis; 5-minute battery swap; NASA heritage (UT Austin)	Factory pilots with Mercedes-Benz at Berlin-Marienfelde; targeting sub-$50,000 at scale
Sanctuary AI	Phoenix (Gen 8)	170 cm / 70 kg	21 DOF hydraulic hands	Carbon AI control system; tactile sensitivity to 5 millinewtons	Partnership with Magna International for automotive tasks
Xpeng	Iron (next-gen)	178 cm / 70 kg	82 (22 per hand)	Three Turing AI chips (2,250 TOPS); solid-state battery; 1st-gen Physical World Large Model	Mass production base construction starting Q1 2026 in Guangzhou, target end-2026 mass production
UBTech	Walker S1	172 cm / 76 kg	Not disclosed	Industrial focus; multi-robot coordination demos	Deployed at BYD, Zeekr, Foxconn, Audi-FAW; 500+ orders
Fourier	GR-2	175 cm / 63 kg	53 (12 per hand)	Six tactile sensors per fingertip; 380 N·m peak torque actuators (FSA 2.0); detachable battery	Commercially available
AgiBot	A2 / GO-1	Not disclosed	Not disclosed	Open-source GO-1 visual-language model	5,100+ shipments in 2025; 39% global market share
Galbot	G1	Not disclosed	Not disclosed	Wheeled lower body, dual arms; mobile pick-and-place focus	$300M+ funding; ~$3B valuation
Kepler	K2	Not disclosed	Not disclosed	30 kg dual-arm payload, 8-hour work cycle on 1-hour charge	Logistics deployments; ~$30,000 base price target
Xiaomi	CyberOne	177 cm / 52 kg	21	Demonstrator unveiled August 2022	Not yet commercial

Not every entrant in the table has shipped meaningful volumes. Galbot's G1 in particular is a wheeled-base humanoid rather than a full biped, and Kepler's K2 is described primarily as a logistics platform.

AI integration

What separates modern humanoid robots from their predecessors is the depth of AI integration. Earlier humanoids relied primarily on pre-programmed behaviors and classical control algorithms. Today's systems use neural networks for nearly every aspect of operation.

LLM-powered reasoning and task planning

Large language models give humanoid robots the ability to understand natural language instructions and decompose them into sequences of physical actions. Figure AI demonstrated this in early 2024 when it showed its Figure 01 robot holding a conversation with a person, understanding a request to hand over something edible from items on a table, correctly selecting an apple (rather than, say, a cup), and explaining its reasoning afterward. The system used OpenAI's models for language understanding and planning, with Figure's own vision and action models handling the physical execution [6].

Tesla's Optimus uses what the company calls an "end-to-end neural network" that takes sensor inputs and produces motor commands, similar in philosophy to the approach Tesla uses for its Full Self-Driving system for cars. The robot learns tasks by watching human demonstrations (teleoperation) and then practices in simulation before deploying on physical hardware [7].

vision-language-action models

A major shift in 2023 to 2026 has been the rise of vision-language-action (VLA) models: large neural networks that take camera images and natural-language instructions as input and directly output robot actions. VLA models inherit semantic knowledge and visual understanding from pretraining on internet-scale text and image data, which gives them better generalization to new objects and instructions than older robot-only training pipelines.

Model	Developer	Year	Parameters	Key idea
RT-2	Google DeepMind	2023	Up to 55B	Co-fine-tuning a vision-language model on robot trajectories with actions tokenized as text
RT-X / Open X-Embodiment	Google DeepMind + 33 labs	2023	Various	Pooled dataset of 22 robot embodiments with 970k trajectories
OpenVLA	Stanford / Google / TRI / others	2024	7B	Open-weight VLA based on Llama 2; outperforms RT-2-X with 7x fewer parameters
Pi0 (π0)	Physical Intelligence	2024	Not disclosed	Flow matching for continuous action generation at 50 Hz; trained on 10k+ hours of robot data; open-sourced 2025
Helix	Figure AI	2025	7B (System 2) + 80M (System 1)	Dual-system architecture with slow VLM reasoning and fast 200 Hz control
GR00T N1 / N1.7	NVIDIA	2025 to 2026	3B	Open foundation VLA for humanoids; N1.7 trained on 20,854 hours of human egocentric video
Gemini Robotics / Gemini Robotics-ER	Google DeepMind	2025	Not disclosed	Robotics fine-tuned variants of Gemini 2.0; deployed by Apptronik and Mercedes
VLT	Xpeng	2025	Not disclosed	Vision-Language-Task model serving as Iron robot's reasoning engine

NVIDIA's GR00T (Generalist Robot 00 Technology) project, announced in 2024, has become an important reference design. GR00T N1, released in early 2025, was the first open foundation model for humanoid reasoning, taking multimodal input and producing manipulation actions. GR00T N1.7, an early-access reasoning VLA released in April 2026, was a 3B-parameter model trained on EgoScale, a dataset of 20,854 hours of human egocentric video covering more than 20 task categories. NVIDIA reported the first scaling law for robot dexterity, showing that increasing egocentric data from 1,000 to 20,000 hours more than doubled average task completion rates [35][36].

Physical Intelligence's Pi0 (π0), released in late 2024, took a different approach: a flow-matching policy that generates smooth action trajectories at 50 Hz, trained on more than 10,000 hours of robot data across eight distinct robot platforms. Pi0 demonstrated tasks no prior robot learning system had performed successfully, including folding laundry from a hamper and assembling cardboard boxes. Physical Intelligence open-sourced Pi0 weights in 2025 through its openpi repository [37].

Figure AI's Helix model, introduced with Figure 03 in October 2025, uses a dual-system architecture inspired by Daniel Kahneman's Thinking, Fast and Slow: System 2 is a 7-billion-parameter VLM that reasons about goals and context at low frequency, while System 1 is an 80-million-parameter network that produces precise joint commands at 200 Hz. The system was trained on 500 hours of teleoperation data [38].

multimodal perception

Modern humanoids fuse data from multiple sensor types: cameras for vision, microphones for speech, force/torque sensors for contact, IMUs for balance, and in some cases tactile sensors for fine manipulation. Vision transformers and multimodal models process this combined sensory stream to build a representation of the robot's environment and the objects in it.

Sanctuary AI's Phoenix is notable for its tactile capability. Its hydraulic hands include sensors that can detect pressure as low as 5 millinewtons, giving it sensitivity comparable to human fingertips. This enables tasks like sorting small mechanical parts and handling flexible materials like wiring harnesses, tasks that require feeling as much as seeing [8].

reinforcement learning for locomotion

Reinforcement learning has become the dominant approach for teaching humanoid robots to walk, run, and maintain balance. The typical pipeline involves training a locomotion policy in a physics simulator with massive domain randomization (varying floor friction, robot weight, push disturbances, etc.) and then deploying the learned policy on the real robot with no additional real-world training.

Unitree's H1 used this approach to achieve its record-setting 3.3 m/s running speed, trained entirely in simulation using NVIDIA Isaac Lab. Boston Dynamics has also adopted RL-based locomotion for the electric Atlas, complementing its decades of classical control expertise with learned policies that can handle a wider range of terrain and disturbances [9].

teleoperation and imitation learning

Most current humanoid programs depend on large-scale teleoperation to bootstrap their learning systems. A human operator wearing a virtual reality headset and motion-tracked gloves controls the robot in real time, and the resulting trajectories are recorded as demonstrations. The robot's policy is then trained to imitate the demonstrations through behavior cloning or hybrid RL-plus-behavior-cloning approaches. 1X's NEO ships with an explicit "Expert Mode" in which a remote operator pilots the robot via VR while the on-board AI watches and learns. Frameworks such as Stanford's HumanPlus, Carnegie Mellon's H2O (Human to Humanoid), and TWIST (Teleoperated Whole-Body Imitation System) push this further, mapping human motion-capture data directly onto whole-body humanoid policies [39][40].

applications

manufacturing

Manufacturing is the first major market for humanoid robots. The appeal is straightforward: factories have structured layouts, repetitive tasks, and clear economic incentives to reduce labor costs. Figure AI's deployment at BMW and Agility Robotics' work with Amazon and GXO represent the leading edge of this trend.

Typical manufacturing tasks for humanoids include:

Task	Description	Companies active
Parts loading/unloading	Picking parts from bins and placing them on assembly lines or in carriers	Figure AI (BMW), Agility (Amazon, GXO)
Machine tending	Loading and unloading CNC machines, presses, or injection molding equipment	Apptronik (Mercedes pilot)
Inspection	Using cameras and sensors to check parts for defects	UBTech (Audi-FAW air-conditioning leak detection)
Assembly sub-tasks	Inserting components, fastening, or routing cables	Sanctuary AI (Magna), UBTech (Foxconn)
Multi-robot coordination	Multiple humanoids cooperating on tasks	UBTech (Zeekr 5G smart factory)

logistics and warehousing

Warehouses present a compelling use case because they are designed for human workers but involve highly repetitive physical tasks. Agility Robotics' Digit has moved over 100,000 totes at GXO's facility in Flowery Branch, Georgia, and has been tested at Amazon's fulfillment centers near Seattle, achieving a 98% task success rate. The economics are favorable: Digit operates at an estimated $10-12 per hour in total cost compared to roughly $30 per hour for human labor in similar roles [10].

household

1X Technologies' NEO is positioned as the first humanoid robot designed primarily for the home. At 30 kg (66 pounds), it is lightweight enough to be safe around people, with a soft-body design using custom 3D lattice polymer structures that absorb impact. Target tasks include dishwashing, cleaning, and organizing items. However, the initial product ships with a significant caveat: the robot learns by being teleoperated by human operators, meaning its autonomy builds over time through collected data rather than working fully independently from day one [11].

Figure 03, unveiled in October 2025 with planned deliveries in late 2026, was Figure AI's first robot designed primarily for residential use, with soft textiles, wireless charging, an upgraded audio system, and the on-board Helix model handling autonomous operation [38].

healthcare

Healthcare applications are still mostly in the research phase. Potential uses include patient lifting and transfer, fetching medical supplies, and assisting with rehabilitation exercises. The humanoid form is useful here because hospital environments are designed for people, and patients may respond better to a human-shaped assistant than to an abstract machine.

Fourier Intelligence, originally a rehabilitation-robotics company, has launched the Nexus and Care platforms targeting eldercare and emotional support for patients, integrating its GR-series humanoid bodies with conversational AI. The company describes Nexus as the first humanoid care robot.

space and hazardous environments

NASA's Robonaut 2, launched to the ISS in 2011, is the canonical example of a humanoid in space. Beyond Robonaut, NASA's Valkyrie (officially designated R5) was developed at Johnson Space Center for DARPA-funded research into Mars exploration. Humanoid form factors are being studied for tasks where machines must operate inside spacecraft, lunar habitats, or planetary surface installations designed for human occupancy. On Earth, humanoids are also being explored for nuclear decommissioning and disaster response, the original mission of the DARPA Robotics Challenge.

service and hospitality

Semi-humanoid platforms such as Pepper found early traction in retail, banking, and hospitality, primarily as greeters or interactive information kiosks. Despite around 27,000 Pepper units being produced, the limited utility of stationary upper-body humanoids led SoftBank to pause production in 2021 [23]. The next wave of service humanoids will likely be full bipeds capable of doing more than greeting guests, although consumer-facing deployments lag manufacturing by several years.

major commercial deployments

The industry's transition from demos to deployments accelerated sharply in 2024 and 2025. Notable commercial relationships as of early 2026 include:

Customer	Robot	Site	Application	Status
BMW	Figure 02	Spartanburg, South Carolina	Body-shop parts insertion; 10-hour shifts	Operational; 90,000+ parts loaded as of late 2024
GXO Logistics	Agility Digit	Flowery Branch, Georgia	Tote moving (Robots-as-a-Service contract)	Operational; 100,000+ totes moved by November 2025
Amazon	Agility Digit	Seattle-area fulfillment centers	Empty-tote moving	Pilot expanded 2024 to 2025
Mercedes-Benz	Apptronik Apollo	Berlin-Marienfelde, Germany	Logistics, kitting	Pilot ongoing; backed by Mercedes equity investment
Hyundai Motor Group	Boston Dynamics Atlas	Metaplant America (Savannah, Georgia)	Manufacturing, parts handling	Initial fleet 2026; goal of 30,000 robots/year by 2028
Google DeepMind	Boston Dynamics Atlas	Various research sites	AI research and Gemini Robotics	Initial fleet 2026
Magna International	Sanctuary AI Phoenix	North America	Auto parts manufacturing	Strategic partnership and equity stake
Zeekr (Geely)	UBTech Walker S1	Ningbo 5G smart factory	Multi-robot assembly coordination	First multi-humanoid factory deployment
BYD	UBTech Walker S1	Shenzhen-area plant	Manual production tasks	Operational since 2024
Foxconn	UBTech Walker S1	Longhua, Shenzhen	Logistics validation	Operational
Audi-FAW	UBTech Walker S1	Changchun, Jilin	Air-conditioning leak detection	Pilot

Figure AI has also confirmed deployments with a U.S. logistics customer in addition to BMW, although the partner has not been publicly named [12].

the Chinese ecosystem

China has emerged as a major center of humanoid robot manufacturing in parallel with the United States. The Chinese government identified humanoid robots as a strategic technology in 2023 and 2024, and a wave of new entrants reached production in 2025 to 2026. Unitree achieved unicorn status at a $1.3 billion valuation in June 2025 and reported shipping more than 5,500 units of its G1 platform alone. AgiBot reported about 5,100 shipments in 2025, equating to roughly a 39% share of global humanoid shipments. Xpeng broke ground on a 110,000 square meter humanoid factory in Guangzhou in Q1 2026, with mass production targeted for the end of 2026 [13][29][41]. Other Chinese entrants of note include Galbot, Kepler, UBTech (which floated on the Hong Kong Stock Exchange in 2023), Xiaomi (with the CyberOne demonstrator unveiled in 2022), Leju Robotics, EX Robots, Robotera, and Engineai. Component supply chains for harmonic reducers, frameless motors, and force sensors remain partially Japanese-controlled (notably Harmonic Drive Systems and Nidec), although Chinese suppliers are expanding rapidly.

major research labs

Alongside commercial activity, university and government laboratories continue to drive long-term humanoid research:

Lab	Affiliation	Focus	Notable platforms
JSK Lab	University of Tokyo	Whole-body sensing and control	H6, H7, JAXON, Kengoro
AIST Humanoid Research Group	Japan AIST	Industrial humanoids and standards	HRP series (HRP-2 through HRP-5P, HRP-4C)
MIT Biomimetic Robotics Lab	MIT	QDD legged robots, locomotion	MIT Cheetah, MIT Humanoid
AMBER Lab	Caltech	Hybrid zero dynamics and walking control	DURUS, ATALANTE, Cassie research
Italian Institute of Technology	IIT (Genoa)	Cognitive humanoids, aerial humanoids	iCub, iCub3, iRonCub
Toyota Research Institute (TRI)	Toyota	Manipulation and policy learning	Punyo, large behavior models
Disney Research	Disney	Expressive humanoid characters	BD-X droids, Project Kiwi
Tohoku University, Tokyo Tech	Japan	Disaster response, planetary exploration	Various
Carnegie Mellon Robotics Institute	CMU	Humanoid teleoperation, manipulation	H2O framework, HumanPlus
Stanford IRIS Lab and ILIAD	Stanford	VLA models, behavior cloning	OpenVLA, ALOHA platforms

Many of the engineers who founded the current commercial humanoid wave came from these labs or from corporate research divisions such as Honda Research, Schaft (acquired by Google in 2013), and Boston Dynamics itself.

market projections

Multiple financial institutions have issued forecasts for the humanoid robot market, though estimates vary widely:

Source	Forecast	Key assumptions
Goldman Sachs (2024, updated)	$38 billion TAM by 2035; 250,000+ shipments by 2030	AI progress (robotic LLMs); 40% decline in material costs; industrial focus
Morgan Stanley (2024)	$5 trillion by 2050	Long-term inclusion of consumer and service markets
MarketsandMarkets (2025)	$21 billion by 2030	CAGR of ~46% from 2025
TrendForce (2026)	Rapid scale-up; China and U.S. in production race	Component supply chains maturing; Japan controls key components

Goldman Sachs notably revised its 2035 projection upward by more than sixfold (from $6 billion to $38 billion) between 2022 and 2024, citing faster-than-expected progress in AI as the primary reason [1].

challenges

dexterity

Human hands remain far more capable than any robotic equivalent. Tasks that humans find trivial, like tying shoelaces, peeling a banana, or threading a needle, remain extremely difficult for robots. Even the most advanced robotic hands (such as the Shadow DEX-EE with its hundreds of tactile sensors) fall short of human dexterity, particularly for tasks requiring fine force control and rapid finger repositioning. Progress is being made, but dexterous manipulation remains one of the field's hardest open problems.

balance and locomotion

Bipedal walking is inherently unstable. A biped is essentially a controlled fall, constantly correcting its balance with every step. This makes humanoid robots vulnerable to unexpected disturbances (being bumped, stepping on uneven ground, carrying asymmetric loads) in ways that wheeled robots are not. While RL-trained locomotion policies have improved significantly, humanoids still cannot match humans in their ability to walk confidently across arbitrary terrain.

battery life

Most current humanoid robots operate for one to four hours on a single charge. Boston Dynamics' Atlas leads with approximately four hours thanks to dual battery packs, while many others are in the one-to-two-hour range. Tesla Optimus Gen 3 targets 8 to 10 hours from a 2.3 kWh pack. This limits their usefulness for full work shifts. Battery swapping (used by Atlas and Apptronik's Apollo) is one solution, but it requires charging infrastructure and interrupts work. Advances in battery energy density, including the solid-state batteries planned for Xpeng Iron, will gradually improve this, but it remains a near-term constraint.

cost

Prices for commercially available humanoid robots span a wide range:

Robot	Approximate price	Target market
Unitree G1	$16,000-$74,000 (depending on configuration)	Research, education, development
1X NEO	$20,000 (early access) / $499/month subscription	Consumer household
Apptronik Apollo	Sub-$50,000 target at scale	Industrial
Unitree H1	~$90,000	Research, development
Unitree H2	$29,900	Research, industrial
Sanctuary AI Phoenix	$100,000-$250,000 (estimated)	Enterprise
Figure 02	~$1,000/month (Robot-as-a-Service)	Enterprise manufacturing
Boston Dynamics Atlas	$140,000+ (estimated)	Enterprise, industrial
Tesla Optimus (target)	$20,000-$30,000 (target at production scale)	Enterprise, eventually consumer
Kepler K2 (target)	~$30,000	Industrial logistics

For mass adoption, prices need to drop significantly. The robot-as-a-service model (used by Figure AI at approximately $1,000 per month) attempts to make the economics work by spreading costs over time, similar to how cloud computing replaced upfront server purchases.

safety and regulation

A full-sized humanoid robot weighing 50 to 90 kg, moving at walking speed, with arms capable of lifting 25 to 50 kg, is inherently capable of injuring a person. Two main international standards are relevant: ISO 10218 (industrial robots, in revision since 2011) and ISO/TS 15066 (collaborative robots, published 2016). Both were written for fixed-base industrial arms operating with engineering controls such as safety fences, light curtains, or speed-and-separation monitoring. Neither was designed for autonomous bipeds navigating shared workspaces.

The most relevant existing standard for human-shaped service robots is ISO 13482:2014, which covers personal care robots in non-industrial environments and addresses three categories: mobile servant robots, physical assistant robots, and person-carrier robots. ISO 13482:2014 has been criticized as insufficient for the next generation of autonomous humanoids, and a revised standard, ISO/FDIS 13482, is in development under the broader title "Robotics: Safety requirements for service robots" [42][43]. Manufacturers and integrators must currently combine ISO 13482, ISO 10218, ISO/TS 15066, and country-specific machinery directives (such as the EU Machinery Regulation 2023/1230) to construct a defensible safety case. The absence of a single, universally accepted humanoid standard creates uncertainty for companies planning large-scale deployments.

Large-scale humanoid deployment raises concerns about labor displacement, particularly in repetitive manufacturing, warehouse, and service jobs. Goldman Sachs's analysis emphasizes industrial use cases as the first market, where the substitution is for tasks already considered ergonomically taxing. Consumer privacy is a separate concern: home humanoids carry cameras and microphones throughout the household, and current platforms typically rely on cloud-based AI services for at least part of their reasoning. Companies including 1X have committed to local processing for some perception tasks, but a comprehensive privacy and data-handling framework for in-home humanoids has not yet been established.

current state (2025-2026)

The period from 2024 to early 2026 has seen humanoid robotics shift from a speculative technology to an emerging industry:

Boston Dynamics launched the production Atlas at CES 2026 and announced commercial deployments at Hyundai's Metaplant America (Savannah, Georgia) and with Google DeepMind. Hyundai plans to build a manufacturing facility capable of producing 30,000 Atlas units per year by 2028 [4].
Figure AI's valuation grew from $2.6 billion (February 2024 Series B) to $39 billion (September 2025 Series C), fueled by demonstrated real-world deployment at BMW and partnerships with NVIDIA and OpenAI [12].
Tesla began slow production of Optimus Gen 3 at its Fremont factory, though Elon Musk confirmed on the Q4 2025 earnings call that no robots are performing useful work yet; they are being used for learning and data collection. Musk projects volume production in 2027 and has suggested an eventual price of $20,000-$30,000 [7].
1X Technologies opened pre-orders for NEO at $20,000 and is seeking funding at a valuation of up to $10 billion, a more than 12x increase from the previous year [11].
China emerged as a major force in humanoid robotics, with Unitree achieving unicorn status at a $1.3 billion valuation in June 2025 and the Chinese government making humanoid robots a strategic priority. TrendForce described 2026 as the year of a "humanoid boom" with both China and the United States racing for production scale [13].
Agility Robotics opened RoboFab, a 70,000 square foot manufacturing facility in Salem, Oregon, with plans to eventually produce over 10,000 Digit robots per year [10].
NVIDIA released Isaac GR00T N1.7 in early access in April 2026, a 3-billion-parameter open reasoning vision-language-action model trained on more than 20,000 hours of human egocentric video and licensed for commercial deployment [36].
Mercedes-Benz announced an overhaul of its Berlin-Marienfelde production site centered on AI and humanoid robotics, deploying Apptronik Apollo robots backed by Google DeepMind technology, including Gemini Robotics models [44].
Xpeng announced the start of construction on a 110,000 square meter humanoid mass production base in Guangzhou in Q1 2026, with mass production targeted for the end of the year [29].

The consensus among industry analysts is that humanoid robots will first achieve scale in manufacturing and logistics, where the environments are semi-structured and the return on investment is most easily calculated. Consumer and household deployment, while technically possible (as 1X's NEO demonstrates), will take longer due to the much greater variability of home environments and the higher bar for safety and reliability that consumer products must clear.

references

[1] Goldman Sachs. (2024). "The global market for humanoid robots could reach $38 billion by 2035." https://www.goldmansachs.com/insights/articles/the-global-market-for-robots-could-reach-38-billion-by-2035

[2] Waseda University. "Humanoid History - WABOT." https://www.humanoid.waseda.ac.jp/booklet/kato_2.html

[3] Wikipedia. "ASIMO." https://en.wikipedia.org/wiki/ASIMO

[4] Boston Dynamics. (2026). "Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry." https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/

[5] Articsledge. (2026). "AI Humanoid Robots 2026: Technology, Builders & Future." https://www.articsledge.com/post/ai-humanoid-robots

[6] Figure AI. (2024). "Figure 01 + OpenAI demo." Referenced in: https://www.figure.ai/

[7] BotInfo. (2026). "Tesla Optimus: Complete Analysis of AI, Specs & Future Outlook." https://botinfo.ai/articles/tesla-optimus

[8] Sanctuary AI. "Sanctuary AI Unveils Phoenix, a Humanoid General-Purpose Robot Designed for Work." https://www.sanctuary.ai/blog/sanctuary-ai-unveils-phoenix-a-humanoid-general-purpose-robot-designed-for-work

[9] NVIDIA. (2025). "Closing the Sim-to-Real Gap: Training Spot Quadruped Locomotion with NVIDIA Isaac Lab." https://developer.nvidia.com/blog/closing-the-sim-to-real-gap-training-spot-quadruped-locomotion-with-nvidia-isaac-lab/

[10] Agility Robotics. "Digit Moves Over 100,000 Totes in Commercial Deployment." https://www.agilityrobotics.com/content/digit-moves-over-100k-totes

[11] 1X Technologies. "NEO Home Robot." https://www.1x.tech/discover/neo-home-robot

[12] Figure AI. (2025). "Figure Exceeds $1B in Series C Funding at $39B Post-Money Valuation." https://www.figure.ai/news/series-c

[13] Humanoid Horizon. (2026). "The 2026 Humanoid Boom: China and U.S. Race for Scale While Japan Controls the Components." https://humanoid-horizon.com/humanoid-robot-2026-trendforce-forecast/

[14] Wikipedia. "R.U.R." https://en.wikipedia.org/wiki/R.U.R.

[15] Nippon.com. "Leading the World in Humanoid Robotics." https://www.nippon.com/en/views/b00901/

[16] Vukobratovic, M. and Borovac, B. (2004). "Zero-Moment Point: Thirty Five Years of Its Life." International Journal of Humanoid Robotics, 1(1), 157-173. https://www.worldscientific.com/doi/10.1142/S0219843604000083

[17] Wikipedia. "Zero moment point." https://en.wikipedia.org/wiki/Zero_moment_point

[18] Wikipedia. "Humanoid Robotics Project." https://en.wikipedia.org/wiki/Humanoid_Robotics_Project

[19] Wikipedia. "HRP-4C." https://en.wikipedia.org/wiki/HRP-4C

[20] Wikipedia. "Honda P series." https://en.wikipedia.org/wiki/Honda_P_series

[21] IEEE Spectrum. (2015). "How South Korea's DRC-HUBO Robot Won the DARPA Robotics Challenge." https://spectrum.ieee.org/how-kaist-drc-hubo-won-darpa-robotics-challenge

[22] Wikipedia. "Nao (robot)." https://en.wikipedia.org/wiki/Nao_(robot)

[23] IEEE Spectrum. (2021). "SoftBank Stops Making Pepper Robots, Will Cut 165 Robotics Jobs in France." https://spectrum.ieee.org/softbank-stops-making-pepper-robots-will-cut-165-robotics-jobs-in-france

[24] Wikipedia. "iCub." https://en.wikipedia.org/wiki/ICub

[25] NASA. "Robonaut 2." https://www.nasa.gov/robonaut2/

[26] PatSnap Eureka. "SEA vs Quasi-Direct Drive for Legged Robots." https://www.patsnap.com/resources/blog/rd-blog/sea-vs-quasi-direct-drive-for-legged-robots-patsnap-eureka/

[27] Boston Dynamics. (2024). "An Electric New Era for Atlas." https://bostondynamics.com/blog/electric-new-era-for-atlas/

[28] Mike Kalil. (2025). "Figure 03: The Most Advanced AI Robot for Home?" https://mikekalil.com/blog/figure-03-robot-for-home/

[29] CnEVPost. (2026). "Xpeng to break ground on humanoid robot factory in Q1, with mass delivery targeted in 2026." https://cnevpost.com/2026/02/25/xpeng-to-break-ground-on-humanoid-robot-factory-q1/

[30] Robozaps Blog. (2026). "Unitree H2 Review: Price ($29,900) & Specs [2026]." https://blog.robozaps.com/b/unitree-h2-review

[31] Humanoid.guide. "Tesla Optimus Gen 3 - Specs, Price & Release Date." https://humanoid.guide/product/optimus-gen-3/

[32] HumanoidsDaily. (2025). "1X Details NEO's 'Human-in-the-Loop' Strategy and Hardware as $20,000 Pre-Orders Go Live." https://www.humanoidsdaily.com/news/1x-details-neo-human-in-the-loop-strategy-and-hardware-as-pre-orders-go-live

[33] WardsAuto. "Mercedes-Benz to pilot humanoid robots in its manufacturing facilities." https://www.wardsauto.com/news/archive-auto-mercedes-benz-apptronik-humanoid-robots-apollo-manufacturing/710570/

[34] Kuindersma, S. et al. (2016). "Optimization-based locomotion planning, estimation, and control design for the Atlas humanoid robot." Autonomous Robots. https://link.springer.com/article/10.1007/s10514-015-9479-3

[35] NVIDIA. (2025). "NVIDIA Announces Isaac GR00T N1, the World's First Open Humanoid Robot Foundation Model." https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks

[36] NVIDIA Hugging Face Blog. (2026). "NVIDIA Isaac GR00T N1.7: Open Reasoning VLA Model for Humanoid Robots." https://huggingface.co/blog/nvidia/gr00t-n1-7

[37] Physical Intelligence. (2024). "π0: Our First Generalist Policy." https://www.pi.website/blog/pi0

[38] HumanoidsDaily. (2025). "Figure Unveils Figure 03, a Humanoid Robot Designed for Mass Production and the Home." https://www.humanoidsdaily.com/feed/figure-unveils-figure-03-a-humanoid-robot-designed-for-mass-production-and-the-home

[39] Stanford / He, et al. (2024). "HumanPlus: Humanoid Shadowing and Imitation from Humans." https://humanoid-ai.github.io/HumanPlus.pdf

[40] Cheng, et al. (2024). "Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation (H2O)." https://human2humanoid.com/

[41] South China Morning Post. (2025). "China's humanoid robot sector enters mass production with Unitree, AgiBot among pack." https://www.scmp.com/tech/tech-trends/article/3307356/chinas-humanoid-robot-sector-enters-mass-production-unitree-agibot-among-pack

[42] ISO. "ISO 13482:2014 Robots and robotic devices - Safety requirements for personal care robots." https://www.iso.org/standard/53820.html

[43] ISO. "ISO/FDIS 13482 Robotics - Safety requirements for service robots." https://www.iso.org/standard/83498.html

[44] WardsAuto. "Mercedes Uses AI and Humanoid Robotics in Production Overhaul." https://www.wardsauto.com/news/mercedes-uses-ai-and-humanoid-robotics-in-production-overhaul/799085/

taxonomy and definitions

history

early concepts and automata

WABOT and Japanese pioneers

Honda P-series and ASIMO

Boston Dynamics and dynamic locomotion

DARPA Robotics Challenge

service and educational platforms

the humanoid form factor: why build robots that look like people?

technical components

actuators

sensors

compute

batteries and power

locomotion control

end-effectors and dexterity

current humanoid robots

AI integration

LLM-powered reasoning and task planning

vision-language-action models

multimodal perception

reinforcement learning for locomotion

teleoperation and imitation learning

applications

manufacturing

logistics and warehousing

household

healthcare

space and hazardous environments

service and hospitality

major commercial deployments

the Chinese ecosystem

major research labs

market projections

challenges

dexterity

balance and locomotion

battery life

cost

safety and regulation

social and economic impact

current state (2025-2026)

see also

references

Improve this article

Related Articles

Machine learning terms/Fairness

ARC-AGI 2

DeepSeek 3.0

Open-source AI

AI search

Tesla Optimus

taxonomy and definitions

history

early concepts and automata

WABOT and Japanese pioneers

Honda P-series and ASIMO

Boston Dynamics and dynamic locomotion

DARPA Robotics Challenge

service and educational platforms

the humanoid form factor: why build robots that look like people?

technical components

actuators

sensors

compute

batteries and power

locomotion control

end-effectors and dexterity

current humanoid robots

AI integration

LLM-powered reasoning and task planning

vision-language-action models

multimodal perception

reinforcement learning for locomotion

teleoperation and imitation learning

applications

manufacturing

logistics and warehousing

household

healthcare