History of humanoid robots
Last reviewed
Sources
37 citations
Review status
Source-backed
Revision
v5 ยท 7,474 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
37 citations
Review status
Source-backed
Revision
v5 ยท 7,474 words
Add missing citations, update stale details, or suggest a clearer explanation.
The history of humanoid robots runs from pre-modern automata, such as the programmable musicians described by the Muslim engineer Al-Jazari around 1206 and Leonardo da Vinci's mechanical knight of about 1495, through the first full-scale walking humanoid, WABOT-1, built at Waseda University in 1973, to Honda's ASIMO in 2000 and the modern, AI-driven commercial boom of the 2020s. Humanoid robots are machines designed to resemble and replicate the structure and movement of the human body, typically featuring a head, torso, two arms, and two legs capable of bipedal locomotion. The field has evolved through distinct developmental eras, progressing from basic walking experiments to robots that can perform backflips, navigate complex terrain, and interact with people using large language models. As of 2022, more than 17,488 academic papers on humanoid robotics had been published worldwide (covering 1900 to 2022), with annual output exceeding 1,000 papers per year since 2014. [1] After decades as a research curiosity, the field turned commercial in the mid-2020s: Goldman Sachs raised its estimate of the total addressable market for humanoid robots to $38 billion by 2035, up sixfold from a prior $6 billion projection, while Morgan Stanley has put the broader humanoid ecosystem at roughly $5 trillion by 2050. [33][34]
Long before electronics, engineers built mechanical figures that mimicked human and animal motion, establishing the cultural and technical ambition that modern humanoid robotics inherited. These automata used cams, gears, levers, weights, and hydraulics rather than motors and software, but several anticipated the core idea of a machine that performs a programmed sequence of human-like actions.
In the early thirteenth century the Muslim polymath Ibn al-Razzaz al-Jazari, working in the Artuqid court in Upper Mesopotamia, completed his Book of Knowledge of Ingenious Mechanical Devices (Kitab fi ma'rifat al-hiyal al-handasiyya) in 1206, the year of his death. It described and illustrated more than 100 devices, including a boat carrying four automatic musicians (two drummers, a harpist, and a flautist) that floated on a lake to entertain guests at royal drinking parties. [35] Roboticist Noel Sharkey has argued that the drummer was driven by a pegged, rotating drum whose movable cams struck levers in a chosen order, which he characterizes as an early programmable mechanism, leading some historians to credit al-Jazari with the first recorded design of a programmable automaton. [35]
Around 1495, while under the patronage of Ludovico Sforza, Duke of Milan, Leonardo da Vinci designed a mechanical knight: an armored humanoid figure operated by cables and pulleys that could, by design, sit up, move its arms, turn its head, and open its jaw. The drawings went largely unnoticed until the robotics historian Carlo Pedretti identified Leonardo's notes on the knight around the 1950s. In 2002, American robotics engineer Mark Rosheim built a working full-scale reconstruction from the surviving sketches, filmed by the BBC, demonstrating that the design was mechanically feasible; Leonardo's knight is often cited as the first humanoid robot of Western civilization. [36]
The Enlightenment produced the most lifelike automata of the pre-industrial age. In 1738 the French inventor Jacques de Vaucanson exhibited The Flute Player (Le joueur de flute traversiere) at the Academie Royale des Sciences in Paris, a life-size figure whose moving lips, tongue, and three sets of bellows let it actually play a real flute with a repertoire of about twelve tunes. [37] Vaucanson followed it with the Digesting Duck (Canard digerateur), an automaton with hundreds of moving parts that appeared to eat, drink, and defecate; the digestion was a clever illusion, but the duck became the most famous automaton of the eighteenth century and an emblem of the era's fascination with mechanical life. [37] The word "robot" itself did not appear until Karel Capek's 1920 play R.U.R. (Rossum's Universal Robots), and the modern engineering discipline of humanoid robotics began only with the bipedal walking experiments of the late 1960s described below.
The history of humanoid robot development can be divided into three broad eras, each defined by its primary research focus, the level of intelligence achieved, and the enabling technologies that drove progress. [1]
The Foundation Era was characterized by a focus on solving the fundamental challenge of bipedal walking. Intelligence levels were low; robots in this period could execute pre-programmed movements but had limited sensory feedback or environmental awareness. The primary research goal was proving that a machine could stand upright and walk on two legs without falling over.
Waseda University in Tokyo served as the global epicenter of humanoid robotics research during this period, pioneering three major robot families that laid the groundwork for the entire field. [3]
WAP series (Waseda Automatic Pedipulator). Beginning in 1969, Professor Ichiro Kato and his team at Waseda developed the WAP series, which represented the earliest systematic attempts at mechanical bipedal locomotion. The WAP-1 (1969) was a simple planar walking mechanism. Subsequent iterations improved stability and gait control, establishing the basic principles of static walking that would inform all future research. [3]
WL series (Waseda Leg). Building on the WAP work, the WL series focused on developing more capable lower-body locomotion. The WL-10RD (1985) demonstrated quasi-dynamic walking, where the robot could shift its center of gravity during locomotion rather than relying on purely static balance. This was a critical step toward natural walking patterns. [3]
WABOT-1 (1973). The Waseda Robot 1 was a landmark achievement: the world's first full-scale anthropomorphic robot with integrated limb control, vision, and conversation abilities, built under Ichiro Kato, who became known as the father of Japanese robotics research. WABOT-1 stood approximately 1.6 meters tall and featured a limb control system that allowed it to walk with its lower limbs, grip and carry objects with its hands, and communicate in Japanese using an artificial mouth and ears. [5] Its vision system could measure distances and directions to objects. While its walking was slow and required a carefully controlled environment, WABOT-1 proved that a single robot could integrate locomotion, manipulation, and perception. [5]
WABOT-2 (1984). The second generation Waseda Robot was a specialized humanoid robot designed to play keyboard instruments. WABOT-2 could read a musical score with its camera-based vision system and play an organ or electronic keyboard using its ten fingers and two feet on the pedals. It demonstrated dexterity and coordination that went well beyond simple walking, showing that humanoid platforms could perform complex, skilled tasks. [3]
WABIAN (1996). The WAseda BIpedal humANoid represented a return to full-body humanoid design. Standing at approximately 1.66 meters and weighing about 107 kilograms, WABIAN demonstrated improved walking capabilities compared to earlier systems. It featured 35 degrees of freedom and could carry out various walking patterns, including forward walking, sideways stepping, and turning. WABIAN established a platform that would be continuously refined over the next two decades. [3]
Beyond Waseda, other institutions also contributed during this era. Honda began its secretive humanoid research program in 1986 with E0, an experimental machine that walked in a straight line on two legs but took about five seconds per step. Over more than a decade the company built a series of prototype walking robots (E0 through E6, then P1, P2, and P3) before publicly revealing any results. [6] The Honda P2, unveiled in December 1996, stunned the robotics community as the first self-contained humanoid robot capable of independent walking, standing at 1.82 meters and weighing 210 kilograms. [6]
The Integration Era marked a shift from pure locomotion research to the integration of sensing, intelligent control, and human interaction capabilities. Robots in this period achieved medium levels of intelligence: they could perceive their environment using multiple sensors, make simple decisions, and interact with people. The key technological advance was the combination of improved computer vision, force and torque sensing, and more sophisticated control algorithms that allowed robots to anticipate movements and dynamically adjust their center of gravity.
ASIMO (2000). Honda's Advanced Step in Innovative Mobility became the first globally recognizable humanoid robot. Standing 1.30 meters tall and weighing 54 kilograms in its initial version, ASIMO could walk at 1.6 km/h, climb stairs, and recognize faces and voices. [2] Honda had invested over $100 million and 14 years of research before ASIMO's public debut. The 2005 update enabled ASIMO to run at 6 km/h, and by 2011 the robot had 57 degrees of freedom, weighed 48 kilograms, could run at 9 km/h, hop on one leg, pour drinks, and sign with its hands. [2] ASIMO became a cultural icon, performing demonstrations worldwide and inspiring a generation of roboticists. Honda retired the ASIMO program in 2022 to focus on more practical robotic applications.
QRIO (2003). Sony's Quest for cuRIOsity was a compact humanoid standing just 0.58 meters tall and weighing 7.3 kilograms. Despite its small size, QRIO made history as the first humanoid robot capable of bipedal running, briefly achieving a state where both feet left the ground simultaneously. QRIO could also recognize faces, respond to voice commands, and perform coordinated dance routines. Sony discontinued the project in 2006 as part of a broader corporate restructuring.
HRP-2 (2004). Developed by Japan's National Institute of Advanced Industrial Science and Technology (AIST) in collaboration with Kawada Industries, HRP-2 (Humanoid Robotics Project 2) stood 1.54 meters tall, weighed 58 kilograms, and featured 30 degrees of freedom. Its design was notable for its slim, almost anime-inspired aesthetic, created by robot designer Yutaka Izubuchi. HRP-2 was specifically designed as a research platform and became one of the most widely used humanoid robots in academic laboratories worldwide. It demonstrated capabilities including lying down and getting back up, walking on uneven terrain, and performing cooperative tasks with human operators.
HUBO (2004). Developed at the Korea Advanced Institute of Science and Technology (KAIST) under Professor Jun-Ho Oh, HUBO (short for "humanoid robot") stood 1.25 meters tall and weighed 55 kilograms. It was South Korea's first full-sized humanoid and featured 41 degrees of freedom. HUBO could walk, gesture, and interact with people, and its development established KAIST as a leading humanoid research center. [10]
Other notable developments. During this era, several other significant humanoid projects emerged. Toyota unveiled its Partner Robots in 2004, including a humanoid that could play a trumpet by controlling its lips and fingers to produce musical notes. In 2006, Aldebaran Robotics (later acquired by SoftBank) introduced NAO, a small (0.58 m) programmable humanoid that became the standard platform for the RoboCup Standard Platform League and the most widely used humanoid in education and research. NASA and DARPA funded various humanoid projects aimed at robots that could operate in environments designed for humans. The iCub, introduced in 2004 by the Italian Institute of Technology as an open-source platform for research in cognitive development, became one of the most influential research humanoids of the era. [15]
The current era is defined by the convergence of highly dynamic motion capabilities and artificial intelligence, particularly deep learning, computer vision, and large language models. Robots in this period demonstrate high levels of intelligence: they can perceive complex, unstructured environments in real time; make sophisticated decisions about locomotion, manipulation, and interaction; and learn new skills through reinforcement learning and imitation learning.
Boston Dynamics Atlas (2013 onwards). The original hydraulic Atlas, unveiled in 2013, stood 1.65 meters tall, weighed 82 kilograms, and featured 28 degrees of freedom. Designed initially for the DARPA Robotics Challenge, Atlas demonstrated human-like perception and decision-making combined with extraordinary dynamic abilities. [4] Over the following years, Boston Dynamics released videos showing Atlas navigating obstacle courses, performing backflips (2017), executing parkour sequences (2018), and performing synchronized dance routines (2020). These demonstrations fundamentally changed public perceptions of what humanoid robots could achieve. In April 2024, Boston Dynamics retired the hydraulic Atlas and unveiled the all-electric Atlas, which features a fully electric design with higher strength-to-weight ratio and an expanded range of motion that exceeds human capabilities, including joints that can rotate 360 degrees. [11]
Tesla Optimus (2022 onwards). Elon Musk unveiled the Tesla Bot concept at the company's AI Day on 19 August 2021, and Tesla demonstrated a working prototype (Optimus Gen 1) at its second AI Day on 30 September 2022, where the robot walked across the stage at about 0.3 meters per second and waved to the audience. Standing 1.73 meters tall and weighing approximately 73 kilograms, Optimus aims to be a general-purpose humanoid robot manufactured at scale using Tesla's expertise in AI, batteries, motors, and mass manufacturing. The Optimus Gen 2, shown in December 2023, featured improved hands with 11 degrees of freedom per hand and smoother walking. [13] Tesla has stated its goal is to produce millions of units at a target price below $20,000, which would revolutionize the economics of humanoid robotics.
Figure 01 and 02 (2024). Figure AI, founded in 2022 by Brett Adcock, attracted significant attention and investment (over $750 million by early 2024) for its approach to building commercially viable humanoid robots. Figure 01 was designed as a 1.68-meter, 60-kilogram general-purpose humanoid. In early 2024, Figure demonstrated Figure 01 integrated with OpenAI's vision-language models, enabling the robot to hold natural conversations, identify objects, and reason about tasks in real time. Figure 02, unveiled in August 2024, featured fourth-generation hands with 16 degrees of freedom each, onboard vision-language models running on custom compute, and the ability to perform tasks in BMW manufacturing facilities. [12] Figure 03 was announced in March 2025 with significant improvements in form factor and capabilities.
Other significant developments. Agility Robotics began deploying its Digit robot in Amazon warehouses for material handling tasks in 2024. Unitree released the H1 and G1 humanoids at price points starting around $16,000, making humanoid platforms accessible to a much wider range of researchers and developers. UBTECH continued developing its Walker series, with the Walker S line targeting industrial and commercial applications. Fourier Intelligence released the GR-1 and GR-2 for rehabilitation and general-purpose applications. China's rapid investment in humanoid robotics led to dozens of new humanoid platforms from companies including Agibot, XPeng Robotics, and many others.
Between early 2025 and mid-2026, the AI and Dynamics Era was reshaped by the arrival of robot foundation models, sometimes called vision-language-action (VLA) models or large behavior models, which replaced much of the hand-engineered, task-specific control software that had previously governed humanoid behavior. These models map camera images and natural-language instructions directly to continuous motor commands, and they are typically pretrained on a heterogeneous mixture of real-robot teleoperation data, human video, and simulation. [16][17]
On 20 February 2025, Figure AI released Helix, a generalist VLA model that the company described as the first to output high-rate continuous control of an entire humanoid upper body, including wrists, torso, head, and individual fingers. [16] Helix uses a dual-system design: System 2 is an internet-scale vision-language model for scene understanding and language comprehension, while System 1 is a visuomotor policy that converts System 2's latent representations into full upper-body control at 200 Hz. [16] Figure also reported that Helix was the first VLA to run simultaneously on two robots collaborating on a shared, long-horizon manipulation task. [16]
On 18 March 2025, NVIDIA announced Isaac GR00T N1, which it described as the world's first open, fully customizable foundation model for generalist humanoid reasoning and skills. [17][18] GR00T N1 is a dual-system VLA model: a vision-language module (System 2) based on NVIDIA's Eagle-2 backbone interprets the scene and instruction, and a diffusion-transformer action module (System 1) generates motor actions using a flow-matching objective. [18] The publicly released GR00T-N1-2B checkpoint has about 2.2 billion parameters, of which roughly 1.34 billion are in the vision-language model. [18] NVIDIA evaluated the model on the Fourier Intelligence GR-1 humanoid and reported strong data efficiency on language-conditioned bimanual manipulation, and it listed 1X Technologies, Agility Robotics, Boston Dynamics, Mentee Robotics, and Neura Robotics among early-access developers. [17] An upgraded GR00T N1.5 followed on 19 May 2025. [17]
Google DeepMind also entered humanoid control in 2025. It released Gemini Robotics On-Device on 24 June 2025, a VLA optimized to run locally on a robot and adaptable to new tasks with as few as 50 demonstrations, and on 25 September 2025 it announced Gemini Robotics 1.5 together with an embodied-reasoning model, Gemini Robotics-ER 1.5, that plans tasks (including by calling tools such as web search) before handing execution to the action model. [19][20] Boston Dynamics and the Toyota Research Institute (TRI), under a research partnership formed in October 2024, demonstrated on 20 August 2025 a single large behavior model controlling the electric Atlas end to end, treating its hands and feet within one policy rather than separating locomotion and manipulation. [21][22] The Atlas policy used a roughly 450-million-parameter diffusion-transformer architecture with a flow-matching objective, ran at 30 Hz, and could chain tasks such as picking parts, tying rope, and spreading a tablecloth without task-specific code. [22]
The same period saw new home-oriented humanoids. On 28 October 2025, 1X Technologies opened pre-orders for NEO, a 30-kilogram (about 66-pound) consumer humanoid for household chores, priced at $20,000 outright or $499 per month with a $200 refundable deposit, with deliveries planned for 2026. [23] On 9 October 2025, Figure AI unveiled Figure 03, a home- and commercial-oriented humanoid that runs Helix, weighs about 9 percent less than Figure 02, and adds palm cameras and tactile fingertip sensors able to detect forces as small as three grams; Figure said it would build Figure 03 at a new facility called BotQ with an initial capacity of up to 12,000 units per year. [24]
The following table summarizes key specifications of historically significant humanoid robots across different eras and institutions.
| Robot | Year | Institution | DOF | Mass (kg) | Height (m) | Notable achievement |
|---|---|---|---|---|---|---|
| WABOT-1 | 1973 | Waseda University | N/A | N/A | ~1.60 | First full-scale humanoid with limbs, vision, and speech |
| WABIAN-2 | 2005 | Waseda University | 41 | 64.5 | 1.53 | Advanced bipedal walking with emotion expression |
| HRP-2 | 2004 | AIST | 30 | 58 | 1.54 | Widely used research platform with fall recovery |
| iCub | 2004 | IIT (Italy) | 53 | 25 | 1.04 | Open-source cognitive humanoid for developmental robotics |
| ASIMO (2011 ver.) | 2011 | Honda | 57 | 48 | 1.30 | First humanoid to run, hop, and pour drinks autonomously |
| Atlas (hydraulic) | 2016 | Boston Dynamics | 28 | 82 | 1.65 | Backflips, parkour, dynamic obstacle navigation |
| Valkyrie (R5) | 2013 | NASA | 44 | 44 | 1.90 | Designed for space and disaster response missions |
| DRC-HUBO | 2015 | KAIST | 33 | 80 | 1.75 | Winner of DARPA Robotics Challenge Finals |
| Digit | 2020 | Agility Robotics | 16+ | 65 | 1.75 | First humanoid commercially deployed in warehouses |
| Optimus Gen 2 | 2023 | Tesla | 28+ | ~73 | 1.73 | Mass production target below $20,000 |
| Figure 02 | 2024 | Figure AI | 41+ | ~60 | 1.68 | Integrated vision-language models for real-time reasoning |
| Figure 03 | 2025 | Figure AI | 44 | ~61 | ~1.68 | Home-oriented humanoid running Helix VLA; BotQ mass production |
| Tiangong Ultra | 2025 | Beijing Humanoid Robot Innovation Center | N/A | 52 | 1.80 | Won world's first humanoid half-marathon (2:40:42) |
| NEO | 2025 | 1X Technologies | 22 (hands) | ~30 | ~1.65 | Consumer home humanoid offered at $20,000 |
Humanoid robotics research and development is concentrated in five major geographic regions, each with distinct strengths, institutional traditions, and commercial ecosystems.
Japan holds the longest continuous history in humanoid robotics, spanning over 50 years beginning with Waseda University's pioneering work in the late 1960s. The Japanese approach has historically emphasized morphological simulation (replicating the physical form and movement patterns of the human body) and service robotics (designing robots for direct interaction with people in everyday settings). [3]
Key institutions: Waseda University, AIST (National Institute of Advanced Industrial Science and Technology), University of Tokyo, Osaka University.
Key companies: Honda (ASIMO, now retired), Sony (QRIO, discontinued), Toyota (T-HR3, Partner Robots), SoftBank Robotics (Pepper, NAO; originally Aldebaran Robotics of France, acquired by SoftBank in 2012).
Japan's cultural affinity for robots, influenced by manga and anime traditions portraying robots as helpful companions, has contributed to sustained public and government support for humanoid research. The Japanese government has identified robotics as a strategic national priority, with the New Robot Strategy (2015) and subsequent programs directing billions of yen toward humanoid and service robot development.
The United States has focused on understanding brain mechanisms and achieving dynamic control, often driven by defense and space applications. Substantial funding from DARPA (Defense Advanced Research Projects Agency) and NASA has shaped the direction of American humanoid research, emphasizing performance in extreme or hazardous environments.
Key institutions: MIT (Massachusetts Institute of Technology), Carnegie Mellon University, NASA Johnson Space Center, IHMC (Institute for Human and Machine Cognition), Stanford University, UC Berkeley.
Key companies: Boston Dynamics (Atlas), Tesla (Optimus), Agility Robotics (Digit), Figure AI (Figure 01/02/03), Apptronik (Apollo), Sanctuary AI (Phoenix), 1X Technologies (NEO; Norwegian-American).
The DARPA Robotics Challenge (2012 to 2015) was a watershed event for American humanoid robotics. Spurred by the Fukushima Daiichi nuclear disaster of 2011, the challenge tasked teams with building robots capable of driving vehicles, opening doors, climbing ladders, and performing other tasks in disaster scenarios. It ran in phases: a software-only Virtual Robotics Challenge in June 2013, the physical Trials at Homestead, Florida, on 20-21 December 2013, and the Finals at the Fairplex in Pomona, California, on 5-6 June 2015. [10][8] Twenty-five teams from countries including the United States, South Korea, Japan, Germany, Italy, and Hong Kong competed in the Finals. Team KAIST of South Korea won the $2 million first prize with DRC-HUBO, the only entrant to complete all eight tasks, finishing in 44 minutes and 28 seconds; American teams from IHMC, Tartan Rescue (CMU/NREC), and MIT placed in the top ranks. [10] The competition is widely credited with accelerating humanoid development by several years and seeding many of the companies and researchers that later drove the commercial boom.
Since 2020, Silicon Valley venture capital has poured billions of dollars into humanoid robotics startups. Figure AI raised over $750 million, Agility Robotics secured over $175 million, and numerous other companies have attracted significant funding. This investment wave has been driven by advances in AI (particularly transformer models and reinforcement learning) that make general-purpose humanoid robots appear commercially viable for the first time. Goldman Sachs attributed its sixfold upward revision of the humanoid market (to $38 billion by 2035, with 1.4 million units shipped) largely to advances in AI, including robotic large language models, alongside cheaper components and improved manufacturing. [33]
The funding wave accelerated sharply in 2025 and 2026. In February 2026, Apptronik raised a $520 million extension to its Series A at a reported $5 billion valuation, bringing the round to more than $935 million, with backers including Google and Mercedes-Benz; the company said Apollo was being tested by Mercedes-Benz and GXO Logistics. [25][26] Agility Robotics and the logistics provider GXO signed what both described as the industry's first multi-year, robots-as-a-service humanoid deployment, with Digit moving totes in commercial operation, and Agility took a strategic investment from the German motion-technology company Schaeffler in 2025. [27] Industry trackers reported that humanoid companies raised roughly $2.37 billion in equity in the first quarter of 2026, a large increase over the same period of 2025, though such market-size and growth figures vary widely between sources and should be treated as estimates. [28]
China represents the fastest-growing market for humanoid robotics, with rapid development driven by massive government investment, a large manufacturing base, and an ambitious national strategy to lead in robotics by 2030.
Key institutions: Beijing Institute of Technology, National University of Defense Technology (NUDT), Zhejiang University, Tsinghua University, Shanghai Jiao Tong University.
Key companies: UBTECH (Walker series), Xiaomi (CyberOne), Unitree Robotics (H1, G1, H2), Fourier Intelligence (GR-1, GR-2), Agibot (A2 series), XPeng Robotics (Iron, PX5), Deep Robotics (DR01), LimX Dynamics, Galaxea Dynamics, Booster Robotics.
China's State Council released the "Humanoid Robot Innovation and Development Guidance" in November 2023, setting a target for China to establish a "preliminary humanoid robot innovation system" by 2025 and achieve mass production of humanoid robots by 2027. [9] Multiple Chinese cities, including Beijing, Shanghai, and Shenzhen, have established dedicated humanoid robotics industrial zones and offered subsidies to attract companies. By 2025, China had more humanoid robot companies than any other country.
Several 2025 events made Chinese humanoids globally visible. On 28 January 2025, sixteen Unitree Robotics H1 humanoids performed a synchronized Yangko folk dance, directed by filmmaker Zhang Yimou, during the CCTV Spring Festival Gala, an event watched by a reported one billion viewers; Unitree said the robots had rehearsed for about three months and used AI to keep time with the music, and the company reported a surge in orders afterward. [29] On 19 April 2025, Beijing's E-Town district hosted what organizers called the world's first humanoid robot half-marathon, run alongside human participants; Tiangong Ultra, a 1.80-meter, 52-kilogram robot from the Beijing Humanoid Robot Innovation Center, was the only entrant to finish within the human cutoff, completing the 21.1-kilometer course in 2 hours, 40 minutes, and 42 seconds with several battery swaps. [30] On 5 November 2025, XPeng Robotics unveiled its next-generation Iron humanoid, whose lifelike gait prompted widespread online speculation that a person was inside the suit; the following day XPeng chief executive He Xiaopeng cut open the robot's leg covering on stage to expose its mechanical frame and dispel the claim. [31] By 2026, several Chinese makers had moved toward public listings: Unitree filed for an IPO on the Shanghai STAR Market on 20 March 2026, seeking to raise about 4.2 billion yuan, after reporting that humanoid robots had become its largest revenue line in 2025 with shipments it said topped 5,500 units. [32]
Europe's humanoid robotics tradition emphasizes cognitive robotics and human-robot interaction, with a strong focus on understanding how robots can learn, reason, and collaborate with humans.
Key institutions: Italian Institute of Technology (IIT), German Aerospace Center (DLR), INRIA (France), CNRS-LAAS (France), Technical University of Munich, University of Edinburgh.
Key companies: SoftBank Robotics/Aldebaran (France; NAO, Pepper), PAL Robotics (Spain; TALOS, Kangaroo), Engineered Arts (UK; Ameca), Pollen Robotics (France; Reachy 2), Agile Robots (Germany/China), Oversonic Robotics (Italy; RoBee).
The European Union has funded major humanoid-related research programs, including the iCub project at IIT (one of the most influential open-source humanoid platforms in the world, with 53 degrees of freedom and copies operating in over 30 laboratories globally), and the WALK-MAN project for disaster response. [15] The European approach tends to prioritize safety, ethics, and human-centered design, reflecting the EU's broader regulatory philosophy toward AI and robotics.
South Korea has demonstrated world-class capabilities in humanoid robotics, most dramatically through KAIST's DRC-HUBO winning the DARPA Robotics Challenge Finals in 2015 with a $2 million prize. South Korean research emphasizes advanced mobility and task performance. [10]
Key institutions: KAIST (Korea Advanced Institute of Science and Technology), Seoul National University, KIST (Korea Institute of Science and Technology).
Key companies: Rainbow Robotics (RB-Y1; a KAIST spinoff), Samsung (humanoid research division), Naver Labs (AMBIDEX), Robotis.
Rainbow Robotics, founded by Professor Jun-Ho Oh (the creator of HUBO), went public on the Korean stock exchange in 2023, signaling growing commercial interest in humanoid robotics within South Korea. The South Korean government has included humanoid robots in its national robotics roadmap and allocated substantial research funding through programs administered by the Korea Institute of Robot and Convergence.
The control systems used in humanoid robots have evolved significantly across the three developmental eras, progressing from simple pre-programmed trajectories to sophisticated learning-based approaches. [1]
The earliest and most foundational control method for humanoid walking is Zero Moment Point (ZMP) control, introduced by Miomir Vukobratovic in 1968. The ZMP criterion provides a mathematical condition for dynamic balance: as long as the zero moment point (the point on the ground where the sum of horizontal inertial and gravity forces produces zero net moment) remains within the support polygon formed by the robot's feet, the robot will not tip over. [14] ZMP-based control was used in nearly all humanoid robots from the 1970s through the 2000s, including WABOT, ASIMO, HRP-2, and HUBO. [14]
Dynamic model-based approaches extend ZMP control by incorporating full-body dynamics models. These methods use the robot's complete dynamic equations (often modeled as a multi-link rigid body system) to compute joint torques that achieve desired motions while maintaining balance. The inverted pendulum model, which simplifies the humanoid's dynamics to a point mass atop a massless leg, became a widely used abstraction for gait planning. A related modern refinement, the divergent component of motion (also called the capture point), provides a way to plan three-dimensional bipedal walking by controlling the unstable part of the center-of-mass dynamics. [7]
As computational power increased, optimization-based methods became practical for real-time humanoid control.
Model Predictive Control (MPC) solves an optimization problem at each control step, planning the robot's trajectory over a finite time horizon while respecting constraints on balance, joint limits, and contact forces. MPC allows humanoid robots to anticipate future states and plan accordingly, resulting in smoother and more robust locomotion.
Whole-Body Control (WBC) treats the humanoid as a single integrated system rather than controlling individual joints or limbs independently. WBC formulates the control problem as a quadratic program that simultaneously manages locomotion, balance, manipulation, and other tasks, prioritizing objectives through a task hierarchy. Optimization-based whole-body control of this kind was used by the teams that competed in the DARPA Robotics Challenge, including for the Atlas robot. [8][4]
Trajectory optimization computes optimal motion plans offline or in near-real-time by minimizing cost functions (such as energy consumption or deviation from desired trajectories) subject to dynamic and kinematic constraints.
Central Pattern Generators (CPGs) are neural circuit models inspired by biological locomotion systems found in vertebrates. CPGs produce rhythmic motor patterns without requiring continuous sensory feedback, generating stable walking gaits through coupled oscillator networks. This approach is computationally efficient and naturally produces smooth, rhythmic movements.
Cerebellar Model Articulation Controller (CMAC) is a type of neural network inspired by the cerebellum's role in motor control. CMAC networks provide fast, local learning for motor skill acquisition and have been used for adaptive control in humanoid systems.
The most recent and rapidly evolving category of humanoid control methods leverages machine learning, particularly deep learning.
Reinforcement learning (RL) trains control policies through trial and error in simulated environments. The robot (agent) learns to maximize cumulative reward by exploring different actions and observing their outcomes. Sim-to-real transfer techniques, where policies trained in physics simulators are deployed on real hardware, have become increasingly effective. Companies including Boston Dynamics, Agility Robotics, and Unitree have demonstrated RL-based locomotion and manipulation policies on real humanoid hardware.
Imitation learning trains robots by having them observe and replicate human demonstrations. This can involve teleoperation (a human operator controls the robot, and the robot learns from the recorded motions), motion capture (tracking human movements and mapping them to robot joints), or video-based learning (extracting motion information from video recordings of humans).
Behavior cloning is a specific form of imitation learning that treats the problem as supervised learning: given a dataset of state-action pairs from expert demonstrations, the robot learns a direct mapping from observations to actions.
Inverse reinforcement learning (IRL) infers the reward function that an expert (typically a human demonstrator) is implicitly optimizing, then uses that recovered reward function to train a policy via standard RL. This approach can generalize better than behavior cloning because it captures the underlying objectives rather than memorizing specific trajectories.
Foundation-model control (2023 onwards). The newest learning-based approach treats the entire control problem as a single large neural network, often a vision-language-action (VLA) model or large behavior model, pretrained on a mixture of teleoperation data, human video, and simulation, and then deployed to map images and language directly to motor commands. Examples demonstrated on real humanoids include Figure AI's Helix, NVIDIA's Isaac GR00T N1, Google DeepMind's Gemini Robotics, and the Boston Dynamics and Toyota Research Institute large behavior model on the electric Atlas, several of which combine a slow vision-language reasoning module with a fast action module in a dual-system design. [16][17][22]
Actuators are the "muscles" of humanoid robots, converting energy into physical motion. The choice of actuator technology profoundly affects a humanoid robot's capabilities, including its strength, speed, precision, efficiency, and safety characteristics. [1]
| Actuator type | Operating principle | Advantages | Disadvantages | Example robots |
|---|---|---|---|---|
| Electric (DC/BLDC motors) | Electromagnetic rotation converted via gearboxes | Precise position and velocity control; compact; energy-efficient; low maintenance | Limited force output relative to size; gear backlash reduces accuracy; overheating at high loads | ASIMO, NAO, Optimus, Digit, Figure 02 |
| Hydraulic | Pressurized fluid drives pistons or rotary actuators | Very high force output; excellent for dynamic movements; high power density | Heavy supporting infrastructure (pumps, valves, reservoirs); fluid leaks; noisy; difficult to miniaturize | Atlas (hydraulic version), HRP-2 |
| Pneumatic | Compressed air drives actuators (including McKibben artificial muscles) | Lightweight; inherent compliance; high force-to-weight ratio; natural shock absorption | Difficult to control precisely; compressibility of air causes position uncertainty; requires air supply | Shadow Robot hand, various research platforms |
| Series Elastic Actuators (SEAs) | Motor drives a spring in series with the load | Inherent safety through compliance; accurate force sensing via spring deflection; energy storage and return | Lower bandwidth than rigid actuators; added mechanical complexity; spring can limit maximum stiffness | Valkyrie, DRC-HUBO, various research platforms |
| Shape Memory Alloys (SMAs) | Material deformation through temperature-induced phase change | Very lightweight; noiseless; compact; no gears needed | Slow response time; limited force; low energy efficiency; difficult thermal management | Experimental hand and finger designs |
Despite decades of progress, humanoid robotics faces six fundamental technological challenges that limit current capabilities and define the frontier of research. [1]
The human body contains over 200 bones and more than 600 muscles, working together through an intricate system of tendons, ligaments, and joints to produce fluid, efficient, and highly adaptable movement. Replicating this biological complexity in an engineered system remains an enormous challenge. Current humanoid robots typically have 28 to 57 degrees of freedom, compared to the human body's estimated 244 degrees of freedom. The human hand alone has 27 degrees of freedom and 34 muscles, while even the most advanced robotic hands (such as Figure 02's fourth-generation hand with 16 DOF) remain far less dexterous. Understanding and replicating the passive dynamics, variable stiffness, and energy efficiency of biological musculoskeletal systems is an active area of research. [1]
Humanoid robots must perceive their environment using multiple sensor modalities: vision (cameras, depth sensors, LiDAR), touch (force/torque sensors, tactile arrays), proprioception (joint encoders, IMUs), and sometimes auditory input (microphones). The challenge lies not only in the individual sensors but in multi-modal real-time sensor fusion, combining data from all sensors into a coherent, low-latency representation of the world that the control system can act upon. Humans process sensory information with remarkable speed and accuracy; achieving comparable performance in robots requires solving fundamental problems in real-time computer vision, object detection, contact estimation, and state estimation.
Designing the physical structure of a humanoid robot requires balancing competing requirements: sufficient degrees of freedom for versatile movement, low weight for energy efficiency and safety, and high structural strength to withstand impacts and carry loads. Each additional degree of freedom adds weight (actuator, gearbox, wiring, structural support), increases control complexity, and consumes more energy. Engineers must also design for manufacturability, maintainability, and cost. The optimal balance between these factors depends heavily on the intended application, which is why different humanoid robots vary dramatically in their DOF count, weight, and capabilities.
Humanoid robots need materials that are simultaneously lightweight, strong, durable, and (in many cases) compliant or flexible. Current robots use combinations of aluminum alloys, carbon fiber composites, titanium, and various engineering plastics. However, no existing material fully matches the properties of biological materials: bone is remarkably strong relative to its weight, skin is self-healing and highly sensitive, and muscle tissue can produce force while being inherently compliant. Research into advanced materials, including metamaterials, soft robotics materials, and bio-inspired composites, aims to close this gap.
Controlling a humanoid robot in real time requires solving high-dimensional optimization problems at frequencies of hundreds to thousands of hertz. A robot with 40 degrees of freedom must compute coordinated joint commands for all 40 actuators simultaneously, while respecting balance constraints, contact conditions, joint limits, and task objectives. This computational burden is compounded by model uncertainties (the robot's actual dynamics never perfectly match the mathematical model) and environmental disturbances (unexpected pushes, uneven terrain, slippery surfaces). Achieving robust, real-time whole-body control that can handle the full range of situations a humanoid might encounter remains an open problem.
Battery life is one of the most significant practical limitations of current humanoid robots. Most battery-powered humanoids can operate for only 1 to 4 hours on a single charge, depending on their activity level. ASIMO's battery lasted approximately 1 hour, Atlas's hydraulic system consumed energy rapidly, and even newer electric humanoids like Unitree H1 have limited operational duration. The fundamental challenge is energy density: current lithium-ion batteries provide roughly 250 Wh/kg, while the human body's metabolic energy system achieves an effective energy density many times higher. Thermal management is a related challenge, as electric motors and batteries generate substantial heat during operation, requiring cooling systems that add weight and complexity.
The humanoid robotics field is converging on several key research directions that will likely define progress over the coming decade. [1]
Brain-like intelligence and cognitive architectures. Integrating large language models, vision-language models, and other foundation AI models into humanoid control systems represents the most active area of current research. The goal is to give humanoid robots the ability to understand natural language instructions, reason about tasks, plan multi-step actions, and adapt to novel situations without explicit programming. Companies like Figure AI have already demonstrated early versions of this integration, with robots that can hold conversations while performing physical tasks.
Advanced multi-sensor fusion. Future humanoid robots will require more sophisticated sensor fusion systems that combine visual, tactile, proprioceptive, auditory, and potentially olfactory data into unified world models. Event-based cameras (neuromorphic sensors), high-resolution tactile skins, and novel proprioceptive sensors are being developed to provide richer sensory input.
Bio-mechanical-electrical integration. Tighter integration between mechanical design, electrical systems, and biological inspiration is expected to produce more capable and efficient humanoids. This includes research into artificial muscles (electroactive polymers, hydraulically amplified self-healing electrostatic actuators), tendon-driven mechanisms, and variable-stiffness joints that more closely replicate biological motor systems.
Self-adaptive and smart materials. Materials that can change their properties in response to environmental conditions (self-healing polymers, variable-stiffness materials, phase-change materials for thermal management) could dramatically improve humanoid robot durability and versatility.
High-density energy storage. Advances in battery technology (solid-state batteries, lithium-sulfur, lithium-air) and alternative power sources (micro fuel cells, supercapacitors, wireless power transfer) are critical for extending humanoid operational time from the current 1 to 4 hours toward full working day endurance.
Sim-to-real transfer and world models. Training humanoid robots in increasingly realistic physics simulations and transferring learned skills to real hardware is becoming more effective. Combined with learned world models that allow robots to predict the consequences of their actions, this approach could dramatically accelerate the acquisition of new skills. Open robot-foundation-model releases such as NVIDIA's Isaac GR00T N1, together with simulation and synthetic-data tooling, are intended to make this sim-to-real pipeline available to a broad range of developers. [17][18]