What does the humanoid robotics startup landscape look like today?
In our humanoid robotics market deck, you will find everything you need to understand the market
SUMMARY
What does the humanoid robotics startup landscape look like today? It is now a real early commercial market, with China leading hardware scale, Figure AI, Agility Robotics and Apptronik forming the strongest U.S. group, and no company yet close to delivering a reliable general-purpose humanoid worker.
Leadership is fragmented because the companies are winning at different things. AgiBot and Unitree have scale, Figure has one of the strongest Western full-stack positions, Agility has unusually good commercial evidence, and Apptronik has perhaps the strongest external ecosystem.
Humanoids have moved beyond demos, but only in a narrow sense. BMW, GXO and other customers now have robots performing real repetitive work, yet nobody has shown a large fleet learning many unrelated jobs and operating through a facility with little human help.
China's manufacturing lead is already enormous. AgiBot, Unitree and UBTECH account for most of the humanoids being shipped, but shipment volume currently overstates productive adoption because research, entertainment and robot-data generation still represent a large part of demand.
Figure AI has the clearest U.S. full-stack bet, combining proprietary AI, its own factory and repeat work with BMW. Its $39 billion valuation, though, is far ahead of the commercial footprint that can currently be measured.
Agility Robotics looks almost like the inverse case. Its valuation is much lower than Figure's, while its disclosed customer operating hours, repeat deployments and Digit v5 orders give it one of the strongest bodies of commercial evidence in the sector.
Apptronik could change the U.S. ranking quickly if Apollo deployments begin producing hard customer numbers. Google DeepMind, Mercedes-Benz and Jabil give it an unusually strong setup, but partnerships still tell us less than uptime, throughput and fleet size.
The near-term industrial winner may not even look fully human. Wheeled humanoids remove much of the difficulty and energy cost of bipedal locomotion while keeping the arms, reach, sensors and manipulation abilities that factories actually need.
The hardest problem has shifted toward autonomy. Hardware is becoming cheaper and easier to manufacture while frontier robot models still fail surprisingly ordinary manipulation tasks too often for unsupervised production work.
That makes robot data valuable, but the real moat may be the learning loop rather than the dataset itself: deploy robots, capture failures, retrain, update the fleet and repeat. Cross-robot models from Google DeepMind, Skild AI and Physical Intelligence could make closed proprietary stacks less dominant than many investors assume.
Falling hardware prices will accelerate experimentation before they transform labor economics. A $15,000 humanoid that needs constant supervision can still be a much worse worker than a far more expensive machine that operates independently for years.
The industry has crossed the line from robotics spectacle into early commercial deployment. The next important milestone is pretty unglamorous: customers ordering ten robots, finding that they make economic sense, and then asking for 100 or 1,000.
This market map, featured in our humanoid robotics market deck, highlights top companies and startups in the humanoid robotics market
Why is the humanoid robotics startup landscape so hard to read right now?
The humanoid robotics startup landscape is unusually hard to read today because companies are being called “leaders” for completely different reasons.
AgiBot and Unitree can point to thousands of robots shipped. Figure AI has some of the best-documented factory work in the U.S. Agility Robotics has accumulated tens of thousands of operating hours and signed large commercial orders. Apptronik has Google DeepMind, Mercedes-Benz and Jabil around it. 1X is trying to sell robots directly to households. Meanwhile, startups such as Skild AI and Physical Intelligence are raising billions to build the AI brain without manufacturing humanoid bodies themselves.
The money makes things even noisier. Dealroom says humanoid robotics startups had already raised $8.7 billion in venture funding this year by late July, almost twice the previous full-year record. Figure was valued at $39 billion in its latest private round. NEURA Robotics announced a Series C of up to $1.4 billion. Apptronik has raised more than $935 million in its Series A. Unitree has now gone public in Shanghai.
Those numbers make the industry look mature. The operating evidence is much more uneven.
A robot can ship without doing economically useful work. A factory pilot can be real without proving that thousands of robots will be deployed. A robot can perform an extraordinary manipulation task and still fail too often for a customer to trust it for an entire shift. Even “humanoid” has become fuzzy as companies increasingly put human-like torsos and arms on wheels.
Who's actually winning the humanoid robot race today?
If we have to pick winners today, China leads manufacturing, Figure AI has the strongest Western full-stack position, Agility Robotics has some of the clearest commercial evidence, and nobody has convincingly won the general-purpose intelligence race.
A simple ranking from first to tenth would hide more than it reveals. AgiBot can manufacture at a scale Figure has yet to reach, while Figure has shown autonomous capabilities that are harder to infer from AgiBot's shipment numbers. Agility is valued far below Figure even though its public commercial disclosures are unusually detailed. Apptronik has less published throughput data, but its combination of Google DeepMind and Jabil could become very powerful if Apollo deployments start producing hard numbers.
Europe is also becoming serious. NEURA Robotics now has one of the largest war chests in the industry, while the UK startup Humanoid has Bosch helping manufacture its robots and Schaeffler planning large deployments.
| Company | What looks strongest today | What still needs proving |
|---|---|---|
| Figure AI | Western full-stack robot, proprietary AI, BMW work, growing in-house production | Whether commercial deployments can catch up with its valuation |
| Agility Robotics | Real enterprise deployments, 65,000+ operating hours, large Digit v5 orders | Whether the next generation can scale across many workflows |
| Apptronik | Google DeepMind AI, Jabil manufacturing, major industrial partners | Comparable public data on throughput, uptime and fleet size |
| AgiBot | Very high shipment volume and fast manufacturing ramp | How much volume turns into long-term productive use |
| Unitree | Low-cost hardware, global distribution, profitable broader robotics business | Whether its humanoids become serious autonomous workers |
| 1X | The clearest consumer humanoid launch so far | Whether basic home autonomy becomes genuinely useful |
| NEURA Robotics | Huge funding, broad industrial robotics stack, strong European partners | Large-scale humanoid deployments |
| Humanoid | Bosch manufacturing and large Schaeffler commitment | Moving from proof-of-concept work into sustained deployments |
If you want more recent data on this point, please see our latest humanoid robotics market report.
As this chart shows, and as featured in our humanoid robotics market deck, search interest in where to buy robots has been rising steadily
Have humanoid robots really moved beyond demos?
Yes, humanoid robots have moved beyond demos, but the commercially proven part of the market is still concentrated in a small number of repetitive factory and logistics jobs.
The cleanest example remains BMW's work with Figure AI. BMW says Figure 02 spent roughly 1,250 hours in its Spartanburg plant, worked ten-hour weekday shifts, moved more than 90,000 sheet-metal components and supported production of more than 30,000 BMW X3s. The robot repeatedly positioned parts with millimetre-level precision on a live manufacturing line.
Figure 02 was still doing one defined operation inside an already highly automated body shop. BMW later said the project led to revised safety arrangements, including additional barriers and partitions, as well as better 5G coverage.
Agility Robotics provides a different kind of proof. The company's latest public disclosures say Digit has accumulated more than 65,000 operating hours through deployment commitments across nine customer facilities. Active deployments include GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre. One GXO workflow has passed 100,000 totes moved.
We can now measure operating hours, parts handled and repeat deployments rather than relying only on videos. What we still cannot point to is a large fleet of humanoids arriving at an ordinary factory, learning many unrelated jobs and working across the facility with little human support.
Is Figure AI actually leading the U.S. humanoid robot race?
Figure AI is currently the strongest U.S. full-stack humanoid contender, although its $39 billion valuation is much further ahead than its commercial footprint.
The strongest part of Figure's case is progression. The company got Figure 02 into BMW's production environment, learned from hardware failures there, retired that generation and returned with Figure 03. BMW is now working with Figure 03 on a more complicated sequencing and logistics job involving parts and carts rather than simply repeating the original sheet-metal task.
The original BMW program already gave Figure unusually useful real-world data. The next step is more interesting: BMW chose to continue experimenting with the company after the first deployment ended.
Figure has also become much more credible on manufacturing. Its BotQ factory had produced more than 350 Figure 03 robots by late April. Figure says assembly throughput improved from roughly one robot per day to a demonstrated cycle time of one robot per hour in less than four months. End-of-line first-pass yield had climbed above 80%, while the battery line reached a 99.3% first-pass yield. The company had also manufactured more than 9,000 actuators and more than 500 battery packs.
The valuation still deserves skepticism. Figure's latest financing valued the business at $39 billion. That assumes the company becomes one of the defining platforms in physical AI, rather than merely a successful robot manufacturer. The BMW relationship and BotQ progress make that outcome credible enough to take seriously, but they do not prove anything close to a $39 billion operating business today.
This chart, featured in our humanoid robotics market deck, illustrates yearly venture capital funding for humanoid robotics startups
Is Agility Robotics further ahead commercially than people think?
Agility Robotics may be the most commercially underappreciated major humanoid company right now because its valuation is modest compared with the amount of operating evidence it has published.
Digit has already worked with GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre. Agility says its deployment commitments now span nine customer facilities, with more than 65,000 accumulated operating hours. That is a very different evidence base from a startup whose robots mainly live in its own laboratory.
The bigger development is Digit v5. Agility has disclosed more than $300 million of multi-year orders tied to 1,000 Digit v5 robots under a three-year robots-as-a-service contract. The SEC materials are specific that this figure depends on contractual milestones and should not be treated as current revenue.
The same SEC presentation gives us a rare look at how Agility thinks humanoid economics could work. Management illustrates roughly $500,000 of cumulative revenue over a five-year robot life under its RaaS model, or around $400,000 under an ownership-plus-software-and-maintenance model. Those are management assumptions, not proven economics.
Agility is also trying to solve one of the practical problems BMW exposed: working safely around people. Digit v5 is being designed around “cooperative safety,” which would allow closer operation alongside human workers without treating every humanoid like a traditional robot that needs its own fenced cell.
The company is currently pursuing a public listing through a transaction valuing Agility at $2.5 billion pre-money. Against Figure's $39 billion private valuation, the roughly fifteenfold gap looks much larger than the gap in demonstrated commercial activity.
If you want more recent data on this point, please see our latest humanoid robotics market report.
Can Apptronik catch Figure AI and Agility Robotics?
Apptronik can absolutely become a top U.S. humanoid company, but today its partnerships are stronger than its publicly disclosed deployment numbers.
The startup has assembled an unusually good group around Apollo. Google is an investor. Google DeepMind is working with Apptronik on robot intelligence. Mercedes-Benz, GXO and Jabil are customers or partners. Jabil also gives Apptronik access to a global manufacturing organization that already knows how to build complicated electronics at scale.
Apptronik has raised more than $935 million through its expanded Series A, bringing total capital raised close to $1 billion. Financing is unlikely to be the immediate constraint.
What has changed lately is the company's approach to training. Apptronik opened an expanded Robot Park in Austin and says it is building a wider network of similar environments at partner and customer sites. Fleets of Apollo 2 robots perform tasks repeatedly to generate the data needed for better AI models.
Apollo 2 also comes in both bipedal and wheeled versions. Google DeepMind then adds a serious AI layer: its newest Gemini Robotics 2 system has been demonstrated controlling Apollo 2 from feet to fingertips, including locomotion, manipulation and multi-finger hand use.
The missing piece is public operating evidence. We still do not have an Apptronik equivalent of tens of thousands of customer operating hours or a large disclosed industrial throughput figure. Mercedes-Benz, GXO and Jabil make the company highly credible, but the ranking will change only when Apollo starts producing harder customer numbers.
This chart, featured in our humanoid robotics market deck, shows how Agility Robotics is capturing share in humanoid robotics
Is China already winning the humanoid robot race?
China is already crushing the rest of the world on humanoid robot manufacturing volume, while the broader race remains open because large-scale autonomous work is still much harder than large-scale production.
Omdia's 2026 market study estimated that worldwide humanoid shipments reached 13,318 units in 2025, up almost 480% from the previous year. AgiBot alone shipped 5,168 robots according to Omdia, while Unitree was estimated at 4,200 and UBTECH at roughly 1,000. Unitree later said its actual 2025 humanoid shipments exceeded 5,500.
Whichever Unitree figure we use, the conclusion barely changes. Chinese manufacturers produced the overwhelming majority of the world's shipped humanoids.
The gap has become even more visible lately. Unitree has gone public in Shanghai, raising roughly $900 million, and investor demand was extraordinary. AgiBot has continued pushing production well beyond the thousands. Chinese startups are also filling every layer around the robot: hands, actuators, teleoperation, training centers, world models and motion-control software.
Manufacturing density gives China an advantage that will be difficult to replicate quickly. Thousands of robots expose component failures, assembly problems and cost bottlenecks faster than fleets of dozens. The same physical fleet can also generate far more training data.
Still, the most revealing comment may have come from Unitree founder Wang Xingxing himself. Speaking recently about the point at which humanoids could reliably complete most unfamiliar tasks from language instructions, Wang said the breakthrough might still take five to ten years, although he allowed for a faster two-to-three-year scenario.
| Omdia estimate for 2025 | Humanoid shipments | Approximate global share |
|---|---|---|
| AgiBot | 5,168 | 39% |
| Unitree | 4,200 | 32% |
| UBTECH | 1,000 | 7% |
| Figure AI | 150 | 1% |
| Agility Robotics | 150 | 1% |
| Global total | 13,318 | 100% |
If you want more recent data on this point, please see our latest humanoid robotics market report.
Are China's humanoid robot shipment numbers as commercial as they look?
China's humanoid shipment boom is real, but the numbers currently make productive adoption look further advanced than it actually is.
Counterpoint estimated that worldwide humanoid robot revenue passed $500 million in 2025. The interesting part is where the money came from. Entertainment and performance accounted for 26% of revenue, while data production accounted for another 22%. Intelligent manufacturing represented 17% and warehousing and logistics only 6%.
So almost half of industry revenue came from performance or producing data for AI rather than the factory and warehouse work usually used to justify the humanoid investment story.
Very recent reporting from the Financial Times adds another layer. China has built government-backed robot training centers that buy humanoids and use teleoperators to demonstrate tasks, generating physical-world data that can then be sold into the robotics ecosystem. The system helps startups build fleets and collect experience quickly, but some of the resulting robot demand is effectively investment in training infrastructure.
There are encouraging exceptions. UBTECH's 2025 annual report says full-size embodied humanoid products and services became its largest source of revenue, helping lift company gross profit sharply. AgiBot generated more than $140 million of humanoid robot revenue according to Counterpoint. Unitree reported around $250 million of company revenue in 2025 and is already profitable, although its business also includes quadruped robots and other products.
China has already created a broad early market for humanoid hardware. What remains unclear is how much of that demand persists once training centers, research labs and government-backed purchasing represent a smaller share of the market.
This chart, featured in our humanoid robotics market deck, illustrates yearly funding for humanoid robotics startups
Do humanoid robots really need two legs?
For many of the factory jobs humanoid startups are chasing today, two legs are probably unnecessary.
This is becoming obvious from what the companies themselves are building. Apptronik's Apollo 2 can use either legs or a wheeled base. The UK startup Humanoid is putting most of its near-term effort into its wheeled HMND 01 platform. BMW's Leipzig humanoid project uses Hexagon's wheeled AEON robot. Several Chinese embodied-AI companies are following the same route.
The logic is simple. Factory floors and warehouses are usually flat. Wheels consume less energy, are easier to control, carry weight efficiently and remove one of the hardest failure modes in humanoid robotics: keeping a tall machine balanced while it works.
What customers often want from a humanoid is concentrated above the waist. They need roughly human reach, arms that fit into existing workspaces, hands or grippers that can manipulate human-designed objects, cameras positioned at a useful height and software capable of understanding jobs originally designed for people.
A wheeled humanoid can deliver most of that without spending battery power and computing effort pretending to be human from the hips down.
Legs become much more valuable when stairs, curbs, cluttered terrain or homes enter the picture. A genuinely general-purpose robot will probably need them. Industrial adoption is currently happening in much friendlier environments.
This also changes how we should think about the startup landscape. The commercially strongest “humanoid robotics” company of the next few years could end up selling a machine that looks only partly human. That would be a sensible engineering decision, not a failure of the category.
If you want more recent data on this point, please see our latest humanoid robotics market report.
Is AI now the biggest bottleneck for humanoid robots, and does robot data become the moat?
Yes. AI is now the hardest problem in humanoid robotics, and the race to collect real-world robot data is becoming one of the main ways companies are trying to solve it.
Google DeepMind's newest Gemini Robotics 2 results give us a rare quantitative view of the gap. On Apptronik's Apollo 2, the model achieved roughly 68% success picking objects from a table, 46% from the floor and 76% from a shelf. With a five-finger hand, it managed 92% on unscrewing a light bulb but only 36% when screwing one in, 44% tying a trash bag and 40% closing a ziplock bag.
Those are impressive research results. They would be awful production reliability.
Figure shows the other side of the frontier. Helix 02 has autonomously completed a four-minute dishwasher task containing 61 ordered locomotion-and-manipulation actions without a reset. Figure has also shown two robots tidying a bedroom together, handling clothing, furniture, bedding, doors and other objects through a learned policy.
The problem is getting that performance to survive changes in rooms, objects, lighting and instructions. Companies are therefore building enormous data-collection systems.
Apptronik's Robot Parks repeatedly run Apollo robots through physical tasks. NEURA is building a network of NEURA Gyms where robots can train in real environments. Agility's deployed Digits generate data from actual warehouses and factories. Figure's Project Go-Big uses large-scale egocentric human video to pretrain physical behavior before transferring it onto robots. China has added government-backed training centers where teleoperators generate embodied-AI data at scale.
The interesting threat to these proprietary datasets comes from better cross-robot models. Gemini Robotics On-Device 2 can reportedly adapt to completely new robot bodies with fewer than 200 examples. Skild AI and Physical Intelligence are explicitly trying to build robot brains that work across many machines.
The stronger moat may therefore be the learning loop itself: deploy robots, capture failures, retrain, push the improvement back into the fleet and repeat faster than competitors.
This chart, featured in our humanoid robotics market deck, compares the main business model options for humanoid robot manufacturers
Are home humanoid robots actually close?
Home humanoid robots are close enough to sell today, but the first customers are buying an unfinished capability rather than a reliable robotic housekeeper.
1X has pushed further into this experiment than any other major startup. Its NEO home robot costs $20,000 for early-access ownership or $499 per month under the standard subscription plan. U.S. deliveries are scheduled to begin in 2026.
Demand has been surprisingly strong. 1X says it booked its entire first year of intended production capacity, 10,000 NEO robots, within five days of opening reservations.
We should interpret that carefully. Customers only had to put down a refundable $200 deposit, so 10,000 reservations do not equal $200 million of committed hardware revenue. They do show that thousands of people are willing to reserve a personal humanoid before anyone knows exactly how useful the final product will be.
The autonomy description on 1X's own ordering page is even more informative. NEO arrives with “basic autonomy.” When the robot encounters a difficult task, customers can schedule Expert Mode so a human operator can remotely supervise its actions and help it learn.
That is actually a clever launch model. A partially autonomous robot can deliver some value immediately, human operators cover the missing capabilities, and every intervention creates training data.
It also tells us exactly where home robotics is today. NEO is better understood as an early-access learning system that happens to live in the customer's home. Calling it an autonomous household worker would oversell what 1X itself promises.
Homes remain the hardest environment in the industry. Factories can standardize floors, objects, workflows and safety zones. A home gives the robot stairs, children, pets, cables, clothes, fragile objects and an almost infinite number of unusual situations.
The home market may eventually become much larger than factories. For now, industrial humanoids are closer to earning their keep.
Are humanoid robot prices falling fast enough to matter?
Humanoid robot prices are falling extremely quickly, but a cheap body and a cheap useful worker are still very different products.
Unitree's G1 has pushed basic humanoid hardware down to roughly $13,500. 1X is offering NEO ownership at $20,000. A few years ago, getting a sophisticated full-body robot anywhere near those prices would have sounded unrealistic.
That price compression changes the market immediately for universities, developers, AI labs and companies collecting robot data. A lab that could never justify a $200,000 machine can buy several $13,500 robots and start experimenting.
Industrial economics sit much higher.
Agility's SEC materials illustrate around $500,000 of cumulative revenue over a five-year Digit life under robots-as-a-service, or roughly $100,000 per robot per year if we simply spread the illustrative revenue evenly. The model includes the robot, deployment, software and maintenance, so it should not be compared directly with Unitree's hardware sticker price.
UBTECH provides another useful order of magnitude. Its annual report shows that full-size humanoid robots and related services became its biggest revenue category in 2025, with revenue per delivered full-size robot landing far above low-cost developer hardware once we make a simple division of category revenue by units.
The more useful question is cost per productive hour. A $15,000 robot that needs constant teleoperation can be more expensive than a $100,000 robot that works independently for years. Maintenance, battery life, supervision, utilization and failure recovery can easily matter more than the original purchase price.
Hardware prices are already low enough to accelerate experimentation. Mass labor substitution will start when autonomy and reliability make the total cost of useful work cheap.
This chart, featured in our humanoid robotics market deck, shows the revenue mix across customer segments in the humanoid robotics market
Can humanoid robotics startups really manufacture at scale?
China can already manufacture humanoid robots at meaningful scale, while most Western startups are still proving that their new factories can sustain thousands of reliable units rather than just demonstrate the capacity to do so.
AgiBot is the clearest example of how quickly the production curve can change. The company only began mass production relatively recently, yet it has already moved into five-figure cumulative robot output across its product portfolio. Unitree has separately shipped thousands of humanoids while maintaining a much broader robotics business.
UBTECH has also moved into thousand-unit full-size humanoid volumes. China now has several companies operating at a scale that the leading U.S. humanoid startups have mostly announced rather than reached.
The Western approach is catching up through purpose-built factories and manufacturing partners. Figure has already demonstrated one-robot-per-hour assembly at BotQ, although sustaining that pace over a full year is a much harder test. Agility's RoboFab is designed for up to 10,000 Digits per year and uses roughly 75% U.S.-sourced parts, according to its investor materials.
Other companies are avoiding the painful process of building everything themselves. Apptronik has Jabil. Humanoid has Bosch as its contract manufacturing partner, and Bosch has discussed a path to very large production volumes if demand develops. 1X has opened its own NEO factory in California and is preparing around 10,000 units of initial annual capacity.
This is where Europe could become more important than its current humanoid shipment numbers suggest. Bosch, Schaeffler and other industrial groups already know how to make components, run factories and certify equipment. Humanoid and NEURA can plug into that infrastructure instead of recreating it from scratch.
The next useful manufacturing milestone will be boring: thousands of nearly identical robots leaving a production line, reaching customers and coming back with low warranty and failure rates. China is closest to that reality today.
| Company | Current manufacturing evidence | The next test |
|---|---|---|
| AgiBot | Five-figure cumulative robot production across its portfolio | Sustained demand for the output |
| Unitree | Thousands of humanoids shipped plus an established broader robotics business | More industrial rather than developer or performance use |
| UBTECH | Thousand-unit full-size humanoid scale | Faster volume growth with strong unit economics |
| Figure AI | 350+ Figure 03 units and one-per-hour cycle demonstrated | Sustaining high throughput and yield |
| Agility Robotics | RoboFab designed for 10,000 units annually | Digit v5 volume production |
| 1X | NEO factory operating with 10,000 initial production slots booked | Actually delivering that fleet into homes |
| Humanoid | Bosch contract-manufacturing partnership | Moving from beta robots into repeat production |
Why are investors paying so much for humanoid robots and robot brains?
Investors are currently pricing humanoid robotics as a future computing platform, which explains valuations that look absurd beside today's robot revenue.
Dealroom counted $8.7 billion of humanoid robotics venture funding through late July, already close to twice the previous full-year record. China took roughly two-thirds of that capital, up dramatically from its share only a few years ago.
The funding is also spreading beyond companies that manufacture humanoid bodies. Skild AI raised $1.4 billion at a valuation around $14 billion to build general-purpose robot intelligence. Physical Intelligence was valued at $5.6 billion in its previous large round and has since been reported in talks for new funding at roughly twice that level. Generalist AI, another robot-brain startup, has also attracted hundreds of millions of dollars.
Those valuations show how investors expect the industry to develop. The biggest winner may own the intelligence layer used across many different machines rather than the world's largest humanoid factory.
Google DeepMind makes the possibility easier to imagine. Gemini Robotics 2 can control different robot embodiments, while its on-device version can adapt to new hardware with relatively little target-platform data. If that keeps improving, a robot manufacturer may eventually be able to plug a frontier model into its hardware instead of spending billions training an entirely proprietary brain.
Figure is betting heavily on the opposite strategy. The company owns the robot, the factory, much of the data pipeline and Helix, its own AI stack. Apptronik is taking a more mixed approach by building Apollo hardware while working closely with Google DeepMind. NEURA is also trying to combine robots, training infrastructure and a broader intelligence platform.
We do not yet know which architecture wins. Smartphones ended up with both Apple-style vertical integration and Android-style shared intelligence layers. Robotics could develop the same way.
The valuations still run miles ahead of current economics. Counterpoint put the entire global humanoid robot market at only a little above $500 million of revenue in 2025. Several individual startups are already worth tens of billions.
Investors are paying for the possibility that physical AI eventually becomes a labor market measured in trillions rather than a robot market measured in millions. Today's revenue obviously does not justify those prices on its own.
If you want more recent data on this point, please see our latest humanoid robotics market report.
This chart, featured in our humanoid robotics market deck, shows how factory humanoid robot technology has evolved over time
What does the humanoid robotics startup landscape look like today?
The humanoid robotics startup landscape today has finally become a real commercial market, but leadership is fragmented and general-purpose humanoid labor remains unsolved.
China has the clearest lead in physical scale. AgiBot, Unitree and UBTECH have already shown that humanoid robots can be manufactured and shipped in the thousands, helped by a deep electronics supply chain, strong domestic support and a rapidly growing network of embodied-AI companies.
The strongest U.S. startups currently look different from one another. Figure has the most convincing full-stack story, combining proprietary AI, its own manufacturing and repeat work with BMW. Agility has unusually strong commercial disclosure and may be further ahead operationally than its valuation suggests. Apptronik has perhaps the best external ecosystem through Google DeepMind, Jabil and major industrial customers, but we want more numbers from actual deployments before moving it higher.
1X is running a separate experiment in the home. NEURA and Humanoid have suddenly made Europe relevant again. Skild AI, Physical Intelligence and other robot-brain startups are challenging the assumption that the most valuable humanoid company even needs to manufacture humanoids.
The biggest change today is what no longer impresses us much. Walking, running, dancing and recovering from a push have become table stakes among serious competitors. We care far more about operating hours, intervention rates, task success, repeat customer deployments, manufacturing yield and cost per productive hour.
Those metrics point to the same bottleneck: hardware is improving faster than autonomy. Factories can already manufacture thousands of humanoid bodies, while even frontier AI models still fail ordinary physical tasks too often.
Our judgment is straightforward. Humanoid robotics has crossed into an early commercial industry, China leads the hardware side by a wide margin, and Figure, Agility and Apptronik currently form the strongest U.S. startup group. Nobody yet has a general-purpose humanoid worker that can be trusted across arbitrary tasks at human-level reliability.
The company that eventually wins will probably be the one whose first ten robots make enough economic sense that the customer asks for 100, then 1,000.
OUR METHODOLOGY
We broke the humanoid robotics landscape into separate dimensions rather than trying to produce a simple company ranking. The main areas were commercial deployment, manufacturing scale, autonomy and reliability, economics, funding and strategic positioning, and the broader infrastructure forming around physical AI.
Within each dimension, we focused on recent evidence that materially changed what could be said about the market. We gave more weight to robots operating in customer environments, shipment and production volumes, operating hours, task-success rates, repeat deployments, manufacturing yields, contractual commitments and pricing than to one-off demonstrations.
We kept different kinds of evidence separate before drawing conclusions. A large shipment number is strong evidence of manufacturing capability, for example, but it does not automatically prove that those robots are doing productive commercial work. In the same way, a major customer or technology partnership strengthens a startup's position without telling us how reliably its robots already perform in production.
Where possible, we relied on first-hand material: customer disclosures, company technical updates, regulatory filings, annual reports, product documentation and published model evaluations. Independent market research and tier-one reporting were used for market-wide shipment, revenue, financing and public-market comparisons that individual companies could not establish on their own.
We did not use a mechanical score. The conclusions come from looking at each dimension independently and then checking where several kinds of evidence point in the same direction. That is particularly important in humanoid robotics, where manufacturing leadership, commercial leadership and AI leadership currently belong to different companies.
Key sources include BMW Group on humanoid deployments and factory integration, Agility Robotics' SEC materials on commercial deployments and its proposed listing, Agility's investor presentation on Digit v5 orders, RoboFab and illustrative economics, Figure AI's BotQ production update, Figure AI's Series C disclosure, Apptronik's expanded Series A announcement, Apptronik's Robot Park update, Google DeepMind's Gemini Robotics 2 evaluation, Google DeepMind's Gemini Robotics On-Device 2 documentation, 1X's NEO ordering and product documentation, Unitree's 2025 shipment clarification, Omdia's 2026 embodied-intelligent-robot market study, Counterpoint Research on global humanoid commercialization, UBTECH's 2025 annual report, NEURA Robotics' Series C disclosure, NEURA's Physical AI training-network update, and Associated Press reporting on Unitree's Shanghai market debut.
In our humanoid robotics market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- The full range of business models in the humanoid robotics market
