What Is a Humanoid Robot? Inside the Race to Build a General-Purpose Robot Body

What exactly is a humanoid robot?

A humanoid robot is a machine built to move and manipulate the world the way a person does: two legs for walking, a torso and head carrying cameras and sensors, and two arms ending in grasping hands. That body plan isn’t decoration. Cameras, inertial measurement units, and force-torque sensors at the joints feed a control stack that has to solve balance, locomotion, and manipulation at the same time, usually across more than twenty independent degrees of freedom. The whole point of the shape is compatibility — a robot that matches human proportions can walk through a doorway, climb a staircase, pull open a truck door, or reach a shelf that was never designed with robots in mind. That’s the core promise behind a general-purpose robot: instead of retooling a factory or a warehouse for a machine, you drop a humanoid robot into the built environment as it already exists.

That definition matters because it’s easy to conflate “humanoid” with “any robot that does useful work.” A pick-and-place arm bolted to a conveyor is not a humanoid robot, no matter how dexterous its gripper. A wheeled warehouse robot that shuttles shelves is not a humanoid robot either, even if it’s fully autonomous. The term is specifically about the robot body — bipedal, torso-and-head, two-armed — because that’s the configuration that lets one machine, in theory, do the wide variety of physical tasks a human employee does across a shift: walk to a station, pick up a tote, carry it, place a part, open a cabinet, climb a step. Whether that theory holds up under real operating conditions is the harder question, and it’s the one the rest of this guide tries to answer honestly.

Why build a robot shaped like a person at all?

The short answer is infrastructure reuse. Warehouses, factories, hospitals, and retail stores were built around human dimensions — door widths, shelf heights, stair rise, tool handles, vehicle cabins. A wheeled or tracked robot can be faster and cheaper for a single repetitive task, but it usually needs the environment modified around it: ramps instead of stairs, dedicated lanes, custom fixtures. A bipedal machine with human-scale reach is a bet that it’s cheaper, in aggregate, to build one flexible body than to re-engineer every workspace for a specialized one.

The tradeoff is that bipedal locomotion and human-grade dexterity are both extraordinarily hard control problems, which is why most of the companies in this space are, in effect, AI companies as much as hardware companies. The mechanical body has existed in rough form for decades; what’s changed recently is the software — large-scale learned policies and vision-language-action models that let a robot generalize from data instead of being hand-programmed for every task. That shift is part of a broader move often known as physical AI, where machine learning models control real bodies in real environments rather than just generating text or images.

What actually makes a robot body hard to build?

It helps to separate the two engineering problems that get lumped together under “humanoid robot,” because they have very different maturity levels. The first is mechanical: actuators, joints, batteries, and structural design that let a machine stand, balance, and walk on two legs without falling over or overheating. This problem is largely solved at a research level — bipedal walking, running, and even some dynamic recovery from stumbles have been demonstrated across multiple platforms for several years now, including the running-speed record set by Unitree’s H1. The second problem is the control software: given a camera feed and a spoken or written instruction, deciding what the arms and hands should actually do next, in a world full of objects the robot has never seen in that exact position before. That second problem — generalizable manipulation — is the one still being actively solved, and it’s the reason most of the companies building humanoid robots today spend as much or more on AI research as they do on mechanical engineering. A robot that can walk perfectly but can’t reliably pick up an unfamiliar box is not yet a general-purpose robot in any meaningful sense, no matter how impressive its gait looks in a demo video.

Who is actually building humanoid robots right now?

Three broad camps are worth tracking: a well-funded American startup racing to prove commercial deployments, the world’s most-watched EV maker trying to turn its car-manufacturing scale into robot-manufacturing scale, and a cluster of Chinese manufacturers competing hard on price rather than payload.

Figure AI

Figure AI is currently the best-funded pure-play humanoid robotics startup. In September 2025 the company closed a Series C round that exceeded $1 billion, pushing its post-money valuation to $39 billion — a jump that reflects how much capital investors are now willing to commit to the humanoid robot category specifically, not just robotics broadly. Figure’s current-generation machine, Figure 03, was introduced in October 2025. It weighs roughly 60 to 61 kilograms, can lift about 20 kilograms, and runs for around five hours per charge on a swappable battery pack that recharges wirelessly through induction pads built into the robot’s feet. (Sources disagree slightly on the robot’s height — Figure’s own materials and most secondary coverage cite about 5’8“/173 cm, while IEEE Spectrum’s reporting put it closer to 1.6 m/160 cm — but the weight and payload figures are consistent across reporting.)

Tesla Optimus

Tesla’s Optimus program is the highest-profile attempt to apply automotive manufacturing scale to a humanoid robot. Elon Musk has stated a target of producing 50,000 to 100,000 Optimus units in 2026, with longer-term ambitions of scaling to as much as 1 million units of annual capacity at the Fremont factory and potentially 10 million a year at Giga Texas. In January 2026, on Tesla’s Q4 earnings call, Musk announced the company would end production of the Model S and Model X by the end of Q2 2026 and convert the freed-up Fremont lines to Optimus production — a striking signal of how central the robot program has become to Tesla’s manufacturing strategy.

That said, the gap between the stated ambition and the operational reality is worth naming plainly. On the Q1 2026 earnings call, Musk confirmed that Optimus production at Fremont would not actually start ramping until late July or August 2026, and he described the pace of that ramp as “literally impossible to predict,” citing roughly 10,000 unique parts on an entirely new production line. That caution lines up with Tesla’s track record here: an earlier goal of building around 10,000 Optimus units in 2025 fell dramatically short, with actual output reportedly in the low hundreds. Anyone tracking Optimus should treat the 2026 unit targets as a stated aspiration, not a confirmed manufacturing schedule — a distinction worth keeping in mind when weighing Tesla’s roadmap against Figure’s, something covered in more depth in this Figure vs. Optimus breakdown.

The Chinese humanoid manufacturers

While Figure and Tesla dominate Western headlines, a separate and arguably faster-moving competition is playing out among Chinese manufacturers, led by Unitree. Unitree’s G1 — unveiled in May 2024 with mass production beginning that August — carries a base price of just $16,000, making it by far the cheapest mass-produced humanoid robot on the market; higher-spec EDU, Pro, and Ultimate configurations range from about $43,900 up to $73,900. Unitree also sells a larger, full-size machine, the H1 (about 180 cm tall), priced around $90,000, which at one point held the world record for bipedal robot running speed at 3.3 meters per second (roughly 7.4 mph). The strategic logic here is different from Figure’s or Tesla’s: rather than chasing a single flagship deployment, Unitree and its domestic competitors are pushing price down aggressively, treating the humanoid robot as a platform that research labs, universities, and increasingly small manufacturers can simply buy off the shelf. The wider competitive landscape among these manufacturers, including where China’s push into Unitree and UBTech’s humanoid lineups is winning share and what it still lacks compared with Western rivals, is its own story.

A side-by-side comparison

Robot	Maker	Price	Weight	Payload	Runtime	Status (mid-2026)
Figure 03	Figure AI	Not publicly listed (enterprise deals)	~60–61 kg	~20 kg	~5 hours	Deployed in a pilot at BMW’s Spartanburg plant
Optimus	Tesla	Not yet sold commercially	Not disclosed	Not disclosed	Not disclosed	Pre-production; Fremont ramp targeted for late July/August 2026
Unitree G1	Unitree	$16,000 base; $43,900–$73,900 for EDU/Pro/Ultimate variants	Compact frame	Lower than full-size rivals	Not specified	Commercially available, mass production since August 2024
Unitree H1	Unitree	~$90,000	Full human scale (~180 cm)	Not specified	Not specified	Commercially available; held bipedal running speed record (3.3 m/s)

How much does a humanoid robot actually cost?

Cost is where the market splits sharply into two philosophies. On one end, Unitree’s G1 starts at $16,000 for a base configuration, undercutting almost every other humanoid platform by an order of magnitude, with fuller-featured EDU, Pro, and Ultimate variants running $43,900 to $73,900. The full-size H1 sits around $90,000. These are list prices for purchasable hardware, aimed at research institutions, universities, and increasingly commercial pilots. On the other end, Figure and Tesla have not published retail prices at all — Figure’s deployments so far are structured as enterprise partnerships (like the one with BMW), and Tesla has not sold a single Optimus commercially as of mid-2026. That’s not a minor gap in disclosure; it reflects a real difference in business model. Unitree is selling hardware. Figure and Tesla are, for now, selling pilots and demonstrating capability while the economics of mass deployment — including the total cost of ownership once you factor in integration, maintenance, and fleet management — are still being worked out. That total-cost picture, beyond the sticker price, is where the unit economics really get tested.

Are humanoid robots actually working anywhere today, or is this still a lab demo?

This is the question that separates hype from evidence, and the strongest real-world data point currently available comes from BMW. BMW Group ran a ten-month pilot of Figure’s earlier-generation robot, Figure 02, at its Spartanburg, South Carolina plant. Over that pilot, the robots supported production of more than 30,000 X3 vehicles, moving over 90,000 components across roughly 1,250 hours of operation. That’s not a demo reel — it’s a sustained, multi-month operational run inside a real automotive production line, with volume and hours logged.

Following that pilot, BMW moved to deploy the newer Figure 03 at the same Spartanburg facility for logistics and parts-sorting tasks, with the deployment confirmed around late June 2026. BMW has also announced plans to bring humanoid robots into production in Germany for the first time, extending the program beyond its original U.S. pilot site. Together, this is currently the single clearest example in the industry of a humanoid robot moving from pilot to a second-generation, expanded deployment with a named enterprise customer — which is precisely why it gets cited so often as evidence that the category is past pure research and into early industrial use.

Automotive isn’t the only setting where a humanoid-shaped machine has moved past the demo stage — warehouse operators running Agility Robotics’ Digit have been putting a similar bipedal design through tote-handling work on the logistics side, which is a useful comparison point for how “deployed” is defined outside a car plant.

It’s worth being precise about what this does and doesn’t prove. A ten-month pilot moving parts in one automotive plant is meaningful evidence of reliability and integration, but it is not the same as “humanoid robots are now common in factories.” One customer, even a large one running an expanded second deployment, is a single data point. What it does establish is that the failure mode isn’t “the robot fell over constantly” or “it couldn’t finish a shift” — the operational numbers (1,250+ hours, 90,000+ components, 30,000+ vehicles supported) suggest a level of dependability that would have been hard to credit even two or three years ago.

How close are we to genuinely general-purpose robot labor?

Close enough that a major automaker has run a robot for over a thousand hours on a real production line, and nowhere near close enough that you should expect a humanoid robot doing varied tasks across a typical warehouse or retail floor by next year. Both of those statements are true at once, and holding them together is the honest answer.

The case for “closer than it looks”: funding has scaled dramatically (Figure’s $39 billion valuation is a serious institutional bet), a real manufacturer has moved from pilot to expanded deployment with measurable throughput, and hardware costs have started to fall fast enough that a base-model humanoid robot now costs about what a well-equipped pickup truck costs. The software layer — vision-language-action models that let a robot follow a natural-language instruction and generalize across variations of a task — has also improved quickly, which is the real bottleneck behind “general-purpose” in a way that leg strength or battery life never was.

The case for “further than the marketing suggests”: every deployment we can point to with hard numbers is narrow in scope — logistics and parts-sorting, not the full breadth of tasks a human floor worker handles across a shift. Tesla, the company making the boldest production claims, has itself walked back its own timeline twice: a 2025 target of roughly 10,000 units landed at a small fraction of that, and the 2026 ramp that was supposed to already be underway was pushed to late summer, with Musk himself calling the pace “impossible to predict.” Battery runtime around five hours means today’s humanoid robots need charging breaks during any real shift. And “general-purpose” as a phrase is doing a lot of marketing work — a robot proven reliable at sorting parts in one plant hasn’t yet been proven equally reliable at, say, stocking irregular retail shelves or navigating a crowded hospital corridor.

The realistic read: the industry is past the “can it walk without falling” phase and into the “can it be trusted with narrow, repetitive physical tasks at industrial scale” phase, with BMW’s Spartanburg program as the clearest evidence of that. True general-purpose labor — one robot fluidly switching between many unrelated physical jobs the way a human worker does — is still a multi-year proposition, gated less by mechanical design than by how far current AI models can generalize manipulation skills across novel objects and environments they weren’t explicitly trained on.

What’s actually holding humanoid robots back?

A few concrete constraints show up across every serious deployment discussion:

Battery life. Five hours per charge, as reported for Figure 03, is workable for a shift with rotation or charging windows but is a real operational constraint compared with a human worker’s full shift — which is why swappable batteries and induction charging through the feet are being engineered in rather than treated as an afterthought.

Manufacturing scale. Tesla’s own admission that Optimus production timing is “impossible to predict,” on a brand-new line with roughly 10,000 unique parts, illustrates that going from a working prototype to hundreds of thousands of reliable units a year is an entirely different engineering problem than the robot’s locomotion or manipulation software.

Generalization. The AI models steering these robots need to handle objects, layouts, and instructions they weren’t specifically trained on. Task-specific reliability (sorting known parts in a known plant layout) is demonstrably achievable today; open-ended generalization across unfamiliar tasks is the harder unsolved layer.

Unit economics. A $16,000 Unitree G1 and an enterprise-only Figure deployment sit at opposite ends of a cost spectrum, and neither price point tells you the true cost of running a fleet — maintenance, downtime, integration engineering, and fleet-level software support all add up in ways that a sticker price hides. That’s the layer covered in a closer look at what a humanoid robot really costs to run, which is where the real return-on-investment math gets done.

How big is the humanoid robot market, and can you trust the forecasts?

Honestly, not very — at least not the precision that market-research headlines imply. Estimates for the global humanoid robot market by 2030 vary enormously depending on which analyst firm you read: Grand View Research puts it at roughly $4.04 billion, ABI Research at about $6.5 billion, Roots Analysis at around $8.18 billion, MarketsandMarkets at $15.26 billion, and Knowledge Sourcing as high as $18.9 billion. That’s roughly a 4-5x spread between the most conservative and most bullish forecasts, which tells you the analysts don’t agree on how fast adoption will actually happen, not that any one of them is simply wrong. Even near-term forecasts for 2026 alone diverge sharply — Fortune Business Insights estimates $6.24 billion for that year, while Future Market Insights puts the figure at $10.69 billion.

The practical takeaway: treat any single market-size figure you see cited in a headline as one analyst’s model, not a settled fact. The underlying signal worth trusting more is the pattern across sources — funding is scaling fast (Figure’s valuation jump), real deployments are growing (BMW’s move from pilot to expanded rollout), and hardware prices are falling (Unitree’s aggressive pricing) — even while the exact dollar size of the eventual market stays genuinely unresolved. That kind of market uncertainty alongside real operational traction is a pattern that shows up across the physical AI economy more broadly, not just in humanoid robots specifically.

What should you actually watch to track real progress?

Given how much noise surrounds this category, a few signals are more reliable than press-release headlines:

• Deployment hours and throughput, not press releases. BMW’s 1,250+ hours and 90,000+ components moved during the Figure 02 pilot are the kind of concrete operational numbers worth weighing far more heavily than an announcement video.
• Whether a pilot gets renewed or expanded. BMW moving from Figure 02 to Figure 03, and extending deployment plans to a German plant, is a stronger signal than the original pilot announcement — customers don’t usually expand programs that quietly failed.
• Manufacturers’ own hedging language. When a company like Tesla shifts from a firm unit target to language like “impossible to predict,” that’s more informative than the original headline number — it tells you where the honest uncertainty actually sits.
• Price movement at the low end. Unitree’s sub-$17,000 base model matters because it signals the hardware itself is becoming commoditized, which shifts the competitive battle toward software, reliability, and fleet operations rather than raw mechanical engineering.
• Runtime and charging design. Battery life and charging mechanics (like induction charging through the feet) are unglamorous details, but they’re a direct proxy for how close a robot is to fitting into a real, continuous work shift.

For a broader view of how developments like these fit together, the Humanoids: The Race for the Robot Body hub tracks ongoing coverage across every major program mentioned here — Figure, Tesla, Unitree, and beyond. And if any of the terminology in this piece was unfamiliar, the glossary is a good next stop.