What Is a Humanoid Robot?

A working definition

Wikipedia’s entry on humanoid robots defines one as a robot “resembling the human body in shape.” That is a fair start, but it hides the reason anyone builds them. The world is shaped for human bodies: stairs, door handles, shelves at shoulder height, carts, hand tools, car seats. A robot with roughly human proportions can, in principle, use all of that without the building being redesigned around it. Every serious humanoid program, from Tesla to AgiBot, makes the same pitch: one machine that slots into existing workplaces and homes.

We track 66 humanoid models from 45 companies in our robot database. The definition we use for inclusion is practical. The robot needs a human-like upper body (torso, two arms, a head or sensor mast) and is meant to do physical work or interact with people. Legs are optional.

What makes a robot humanoid: the form factor spectrum

Humanoids are not one shape. They sit on a spectrum, and the trade-offs between the shapes explain a lot about price and readiness.

Form factor	What it looks like	Examples we track	Strengths	Weaknesses
Full biped	Two legs, two arms, head	Figure 03, Unitree G1, Boston Dynamics Atlas, Tesla Optimus	Stairs, steps, tight spaces, human reach	Hardest to balance, falls are a safety risk, higher cost
Wheeled-base humanoid	Human upper body on a mobile base	Galbot G1, Humanoid HMND 01, Dexmate Vega	Stable, longer runtime, can stop safely mid-task	No stairs or uneven ground
Upper-body or tabletop	Torso and arms, fixed or on a stand	Pollen Reachy 2, Hugging Face Reachy Mini	Cheap, good for research and data collection	Cannot move around a site
Android	Human face and skin-like covering	Engineered Arts Ameca, Hanson Sophia	Expressive interaction	Little or no useful physical work

Wheeled humanoids are more common in real deployments than the viral biped videos suggest. Several of the Chinese robots with commercial deployment evidence in our data, including Galbot G1 and AgiBot G2, roll rather than walk. Legs matter when a site has stairs or clutter; in a flat warehouse, wheels win on uptime.

Humanoid robots vs other robots

Most robots at work today are not humanoids, and for many single jobs a humanoid is the wrong tool.

Robot type	Typical job	Where it beats a humanoid	Where a humanoid beats it
Industrial arm	Welding, painting, assembly on a fixed cell	Speed, precision, uptime, price per task	Moving between stations, working without a custom cell
Autonomous mobile robot (AMR)	Carrying shelves and carts in warehouses	Cheaper, simpler, never falls	Picking items, opening doors, using stairs
Quadruped	Site inspection, patrol	Rough terrain, stability	Handling objects with two arms
Cobot on a cart	Machine tending, lab work	Cost, safety certification	Reach, lifting from the floor, tasks across a site

The case for the humanoid is generality. A factory can install one robot type that moves totes in the morning and loads fixtures in the afternoon, without rebuilding the line. That is the pitch every general-purpose humanoid maker makes: fill labor gaps across many stations, without trying to beat an arm at any single one. Whether that generality is real yet is the main thing our evidence grades try to measure.

Types of humanoid robots by use

• Industrial and logistics. Built for payload, shift length and battery swaps: Agility Digit, Atlas, UBTECH Walker S2, Apptronik Apollo, AgiBot G2. Most paid humanoid work happens here.
• Home. Built to be lighter, quieter and soft to the touch: 1X NEO (66 lb, per 1X) is the clearest example on preorder. Figure describes Figure 03 as designed for homes too, but home use has only been partner or internal testing.
• Research and education. Affordable, open SDKs, easy to repair: Unitree G1 and R1, Booster K1 and T1, Aldebaran NAO. This is where most units ship by volume.
• Social and entertainment. Expressive faces and conversation rather than physical work: Engineered Arts Ameca, Hanson Sophia, Aldebaran Pepper.

The main parts inside a humanoid

Actuators

An actuator is the “muscle” of each joint: an electric motor, a reducer (gearbox) and an encoder that reports position. Most current humanoids use electric rotary actuators with harmonic or planetary reducers, plus linear actuators (a motor driving a screw) in the legs of some designs. A full-size humanoid has 20 to 50 or more actuated joints. Boston Dynamics’ Atlas spec sheet lists a 50 kg payload; Figure lists 20 kg for Figure 03. Actuators are also the biggest cost line, which is why the supply chain for them matters so much (more on that in our China vs US comparison).

Hands

Hands are where humanoids differ most. Some robots ship with simple one-motor grippers (1X EVE has a 1-DoF gripper per hand), while others carry five-finger hands with 20 or more joints. 1X says the NEO hand on its NEO home robot has 22 actuated degrees of freedom plus three at the wrist, driven by tendons from motors in the forearm. Figure says each Figure 03 fingertip senses about three grams of pressure. Our robot hands guide compares them in detail.

Sensors

The standard sensor kit is several RGB cameras in the head, often stereo or depth cameras, an IMU (inertial measurement unit) for balance, encoders in every joint, and force or torque sensors at wrists and ankles. Newer designs add cameras in the palms and tactile skins on the fingers. Lidar shows up on some industrial models but many teams skip it to save cost.

Compute

Most humanoids carry two layers of computing. A real-time controller runs balance and joint control at hundreds to thousands of cycles per second. A higher-level computer, frequently an NVIDIA Jetson module, runs perception and the AI model. 1X lists NVIDIA Jetson Thor in NEO’s specs. Heavy reasoning can also be pushed to the cloud.

Batteries

Battery packs sit in the torso and typically give two to five hours of work. Figure lists 5 hours for Figure 03 and charges it wirelessly through coils in its feet. NEO lists an 842 Wh pack and 4 hours. Industrial robots such as Atlas and UBTECH’s Walker S2 swap batteries so the robot does not leave the line to charge. Our battery life ranking sorts tracked models by listed runtime.

AI models

The newest layer is the “robot foundation model,” usually a vision-language-action model (VLA). It takes camera images and a text instruction (“put the cups in the dishwasher”) and outputs motor commands. Google DeepMind introduced RT-2 as a VLA in July 2023. Figure’s Helix, announced in February 2025, splits the job in two: a 7-billion-parameter model that understands the scene at 7 to 9 Hz, and an 80-million-parameter policy that outputs actions at 200 Hz across a 35-DoF upper body. NVIDIA released Isaac GR00T N1 in March 2025 as an open model that other robot makers can fine-tune.

How humanoid robots are controlled

There are three modes, and most robots mix them.

1. Teleoperation. A human wearing a VR headset or motion-capture suit drives the robot in real time. This is how most training data is collected, and how many demo videos are made. Figure says Helix was trained on about 500 hours of teleoperated data.
2. Scripted or programmed autonomy. Engineers define a task, such as picking totes off a conveyor, and the robot repeats it with classic motion planning. Most paid industrial work today runs this way or close to it.
3. Learned autonomy. A neural network policy, often a VLA, decides what to do from camera input. This is what the industry is betting on, because it promises one robot that learns many tasks.

Balance and walking are automatic on every modern humanoid, even when the arms are teleoperated. The open question is task autonomy. 1X openly discloses that NEO’s owners can book a remote “1X Expert” to guide the robot through chores it cannot do yet, per its order page. That is honest, and it also means some of the household work a buyer sees is done by a person somewhere else. When you watch a demo, look for an “autonomous” or “1x speed” label. Our methodology scores teleoperated evidence lower than autonomous evidence for exactly this reason.

What humanoid robots do today

Paid work is narrow and mostly logistics. These are the deployments with the strongest public evidence:

Robot	Customer or site	Task	Evidence
Agility Digit	GXO Logistics	Moving totes in warehouses	GXO announced a multi-year robots-as-a-service agreement in June 2024 (GXO); Agility reports Digit at 9 customer sites and over 65,000 operating hours (The Robot Report)
Figure 02	BMW Spartanburg	Placing sheet-metal parts for welding	BMW says the robots supported production of more than 30,000 X3 vehicles over about 1,250 operating hours (BMW Group)
Figure 03	BMW logistics	Sorting parts into sequencing trolleys	Announced June 2026 (The Robot Report); no output metrics yet
Atlas	Hyundai Metaplant, Google DeepMind	Part sequencing, research	Boston Dynamics says 2026 deployments are committed (Boston Dynamics)
UBTECH Walker S2	Chinese auto and electronics plants	Material handling	UBTECH announced mass delivery with orders above 800 million yuan (PR Newswire)
Unitree G1	Universities, labs, developers	Research and education	Sold openly from $13,500 in Unitree’s shop

By volume, the market is mostly Chinese research and education robots. Omdia counted about 13,000 humanoids shipped worldwide in 2025, with AgiBot (5,168 units), Unitree (4,200) and UBTECH (1,000) at the top (TechNode). Tesla is building Optimus but has not shown it doing useful factory work: in January 2026 Musk said no Optimus robots were yet doing useful work at Tesla (Electrek), and its Q3 2026 update says production started on Fremont lines with outside deliveries expected around the end of 2027 (Technology Org).

A short history of humanoid robots

Year	Milestone
Early 1970s	Waseda University builds WABOT-1, usually called the first full-scale humanoid robot (Wikipedia)
1986-1997	Honda builds its E series (1986-1993) and P series (1993-1997) of experimental bipeds
2000	Honda introduces ASIMO; development stops in 2018 and ASIMO retires in March 2022 (Wikipedia)
July 2013	Boston Dynamics reveals the hydraulic Atlas with DARPA funding (Wikipedia)
August 2021	Tesla announces Optimus at AI Day; first prototypes appear in 2022 (Wikipedia)
March 2023	Figure AI reveals Figure 01
August 2023	Unitree announces H1, a full-size biped sold to researchers
April 2024	Hydraulic Atlas retires; electric Atlas is unveiled the next day
May 2024	Unitree launches the compact G1, now listed at $13,500 in its official shop
June 2024	GXO signs the first multi-year humanoid RaaS deal, for Agility Digit
February 2025	Figure introduces the Helix VLA model
September 2025	Figure raises over $1 billion at a $39 billion valuation (Figure)
October 2025	Figure 03 is unveiled; 1X opens $20,000 NEO preorders for US delivery in 2026
January 2026	Product version of Atlas shown at CES
July 2026	The FCC bans imports of new foreign-made humanoids into the US (PBS)
October 2026	Tesla says Optimus production has started in Fremont

Key humanoid robot companies by country

From our company directory, grouped by home country:

• China (20 companies, 35 models): AgiBot, Unitree, UBTECH, Galbot, Fourier, EngineAI, Kepler, Leju, Booster, XPeng, Xiaomi and the Beijing Humanoid Robot Innovation Center (Tiangong).
• United States (9 companies, 12 models): Figure AI, Tesla, Agility Robotics, Boston Dynamics, Apptronik, Dexmate, Foundation and Persona AI. 1X (NEO, EVE) operates from Norway and the United States.
• Europe: Humanoid and Engineered Arts (UK), NEURA Robotics (Germany), PAL Robotics (Spain), Aldebaran and Pollen Robotics (France), IIT (Italy).
• Rest of world: Sanctuary AI (Canada), Mentee Robotics (Israel), Honda and Kawasaki (Japan), Rainbow Robotics (South Korea).

China leads on the number of models and units shipped; the US leads on funding and holds the top two spots on our leaderboard. The China vs US guide goes through prices, policy and supply chains.

Limits and open problems

• Reliability. Few companies publish uptime, mean time between failures or intervention rates. BMW’s Figure 02 numbers are among the only customer-published metrics.
• Generalization. VLA models handle the tasks they were trained on far better than new ones. A robot that folds towels in the lab may fail in a different kitchen.
• Hands. Dexterous hands are fragile, expensive and hard to control. Many deployed robots use simple grippers instead.
• Safety. A 60-90 kg biped that falls can hurt someone, and a balancing robot cannot simply freeze in place the way a wheeled one can. That is one reason wheeled upper bodies keep showing up in early deployments.
• Economics. Agility reported $1.8 million in 2025 revenue against a $140 million operating loss (The Robot Report). The business case still rests on future volume.
• Teleoperation gap. Demos are often remote-controlled. Until autonomy is measured independently, treat capability claims as company-reported.

How to read a humanoid robot demo

Most of what the public sees of humanoids comes from company videos. Before you take one at face value, check:

• Is it labeled autonomous? If the video does not say, assume a person may be driving it. Teleoperation is a normal way to collect training data. It just does not show what the robot can do alone.
• What speed is it played at? Sped-up manipulation clips are common. Look for a “1x” label on screen.
• How many cuts? A single unbroken take is stronger evidence than a montage.
• Is the setting staged? A clean lab table with one object is far easier than a cluttered real kitchen.
• Is a customer named, and did the customer say anything? BMW publishing Figure 02 numbers is worth more than any company video.
• Are there metrics? Hours, units handled, success rate and intervention rate. Most companies publish none.
• Can you buy it? A robot with a price and a delivery date has passed a different test from one shown only on stage.

We apply roughly this checklist to every piece of evidence we log, and weight a customer-confirmed deployment far above a polished video.

Where to compare humanoid robots

We built Humanoid Arena to answer “which robot is actually better” with evidence instead of press releases. The Arena leaderboard ranks every tracked robot by an Elo-style rating built from graded evidence. The robots list has specs, status and sources for each model, and the robot finder filters by budget, form factor and use case. Prices live on the humanoid robot price page. Our methodology explains how evidence is graded, and the full dataset is on open data if you want to check our work.