What's still open in humanoid robotics?
In our humanoid robotics market deck, you will find everything you need to understand the market
SUMMARY
Humanoid robotics still has a lot of room for new startups today, but the best opportunities have shifted away from building another generic humanoid and toward the technologies and applications required to make robots dependable workers.
The market looks more mature from the funding headlines than it does on factory floors. Figure, Skild AI, Generalist and major Chinese manufacturers have attracted enormous amounts of capital, yet most humanoids operating commercially are still performing a narrow set of material-handling and production tasks.
That creates an important split in the market. Full-stack humanoid manufacturing and horizontal robotics foundation models are becoming crowded, while deployment, manipulation, data, safety, maintenance and application-specific software remain much less settled.
Factories and warehouses are the clearest near-term market. BMW, Agility Robotics customers, Mercedes-Benz, Jabil and others are already putting robots into real workflows, but those deployments also reveal how many awkward jobs between traditional automation systems remain unsolved.
Legs are becoming less central to the opportunity. AEON, Apollo 2 and Genesis AI's Eno all show that a human-scale upper body on wheels can reach human workstations while avoiding much of the complexity of balancing and walking.
Manipulation may ultimately matter more than locomotion. Dexterous hands, tactile sensing, force control and durable end effectors are still immature enough that better components could improve almost every humanoid platform rather than competing for a single robot customer's success.
The data opportunity is also changing. The bottleneck is increasingly not collecting millions of routine robot actions, but finding rare failures, useful recoveries, good demonstrations and data that transfers between different hands, arms and robot bodies.
Teleoperation, fleet software and maintenance look more attractive as deployment grows. Human intervention can recover failed tasks and create training data, while customers increasingly need software that measures uptime, completed tasks, maintenance, intervention rates and actual cost per unit of work.
Safety is moving from a research problem toward infrastructure. New standards, NVIDIA's Halos stack and the arrival of learned robot policies create room for independent testing, monitoring, certification and tools that can verify what changed after a robot receives a software or model update.
Home humanoids remain one of the largest untouched markets, but also one of the most dangerous startup bets. Household environments demand far more adaptability than controlled factories, while consumers will expect a heavy autonomous machine to be safe, affordable and reliable around children, pets and unpredictable objects.
The strongest startup thesis is increasingly to own an expensive physical workflow rather than every layer of the robot. A company that becomes exceptionally good at one factory, warehouse or service task can benefit from cheaper third-party hardware and stronger foundation models instead of financing an entire humanoid platform itself.
The least attractive opening is another undifferentiated full-stack humanoid OEM. The more interesting market is everything required to turn increasingly capable robot bodies into machines customers can deploy, trust, maintain and pay for based on useful work.
This market map, featured in our humanoid robotics market deck, highlights top companies and startups in the humanoid robotics market
Is humanoid robotics already too crowded for a new startup?
Humanoid robotics still has room for new startups today, although another generic full-stack robot maker would be entering one of the most crowded parts of the market.
The amount of money already committed to the category is extraordinary. Figure raised more than $1 billion at a $39 billion valuation and is building Figure 03 at its BotQ factory. Skild AI raised $1.4 billion at a valuation above $14 billion for its general-purpose robot intelligence. Generalist reportedly raised another $200 million this week, only two months after a $400 million round. In China, Unitree's public-market debut valued the company at around $50 billion after raising roughly $900 million.
Those numbers make the market look almost finished. The commercial evidence says otherwise. The Financial Times reported this week that Unitree sold 5,215 humanoids in 2025, with most going into research rather than mass industrial deployment; its much larger robot-dog business still accounted for 28,000 units. Figure, Agility, Apptronik and others have real customer projects, but the number of jobs proven over long periods remains small compared with the enormous range of physical work humanoids are supposed to automate.
So we would draw the line quite aggressively. Starting another company whose pitch is essentially “we are building a general-purpose humanoid” looks difficult now. Building something that makes humanoids cheaper, safer, more dexterous, easier to deploy or genuinely useful for one type of work still leaves a lot of room.
If you want more recent data on this point, please see our latest humanoid robotics market report.
Are humanoid robots actually useful today?
Humanoid robots are already useful for narrow factory and logistics jobs today, while general-purpose autonomous labor remains unproven.
BMW gives us one of the best real-world tests. Figure 02 worked roughly 1,250 hours at the Spartanburg plant, handled more than 90,000 sheet-metal parts and contributed to production of more than 30,000 BMW X3s. The robot worked ten-hour weekday shifts over a ten-month period. That is enough operating time to take the result seriously.
Agility has gone further on total fleet experience. In materials released for its planned public listing, the company said Digit had accumulated more than 65,000 operating hours across nine customer facilities. Customers include GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre. We are well past the stage where every useful humanoid example can be dismissed as a conference demo.
The limit becomes obvious when we look at what those robots actually do. They move totes, position components, feed production processes and handle defined logistics workflows inside controlled sites. Unitree CEO Wang Xingxing made the gap unusually explicit this week: he said robots that can autonomously complete most tasks in unfamiliar environments from natural-language instructions may still be five to ten years away, although he thinks it could happen faster.
Humanoids can do useful work now. We still do not have a robot that can walk into an unfamiliar workplace, understand whatever job needs doing and reliably get on with it.
As this chart shows, and as featured in our humanoid robotics market deck, search interest in where to buy robots has been rising steadily
Where are humanoid robots really working now?
Humanoid robots are working mainly in factories, warehouses and distribution centers right now, where companies can define the job and measure whether the robot completes it reliably.
The pattern has become much clearer lately. Agility has commercial deployments with manufacturers and logistics companies. Figure has moved from BMW's body shop into a more complex sequencing project and recently signed a commercial agreement with Catalyst Brands for distribution operations. Apptronik is training Apollo around industrial work with Mercedes-Benz and Jabil. BMW is also testing AEON in battery-module assembly and component manufacturing at its Leipzig plant.
Notice how concentrated that list is. We are not seeing comparable operating records across restaurants, construction sites, hospitals, hotels and millions of homes. The first real humanoid market is being built around controlled industrial environments.
That tells us where customers currently see enough value to tolerate immature technology. A plant can redesign a small work area, improve wireless coverage, create safety zones and measure output thousands of times. A household or small business usually cannot.
| Company | Real-world work being targeted | What the deployment tells us |
|---|---|---|
| Figure | BMW parts handling and sequencing; Catalyst Brands distribution work | Industrial logistics is moving from simple repetition toward more variable manipulation |
| Agility Robotics | Tote movement and material handling across logistics and manufacturing sites | Repetitive material movement has reached genuine commercial deployment |
| Apptronik | Intralogistics, kitting, sorting, inspection and line-side work | Customers are testing several jobs before committing to broader fleets |
| BMW / Hexagon | Battery-module assembly and component manufacturing with AEON | Wheeled humanoid-style robots can fit many factory jobs without copying human locomotion |
Which part of humanoid robotics is already crowded?
Building another generic humanoid body is currently the least attractive open space in humanoid robotics.
Figure already combines proprietary AI, manufacturing, customer deployments and more than $1 billion of recent funding. Apptronik combines Apollo hardware with Google DeepMind models, Mercedes-Benz deployments, Jabil manufacturing expertise and dedicated Robot Parks. Agility now has more than $300 million of contracted multi-year Digit v5 orders, subject to agreed milestones, and is pursuing a transaction valuing the company at roughly $2.5 billion before new capital.
China makes the hardware race even tougher. Unitree's online store now spans several humanoid models, with the R1-D starting around $4,290, the R1 around $4,900 and the G1 at $13,500. Those prices do not mean a $5,000 robot can replace a worker, but they show how quickly basic humanoid hardware is moving down the cost curve.
A new OEM can still break through, but “better humanoid” is too weak a thesis these days. We would want to see something hard for the established players to copy: a radically cheaper architecture, unusually good manipulation, privileged access to a major customer base, a supply-chain advantage or a robot designed around one class of work that others handle badly.
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, illustrates yearly venture capital funding for humanoid robotics startups
Does a humanoid startup really need to build legs?
Many new humanoid startups can skip legs today because customers care more about reaching, grasping and moving safely through a site than copying the human body exactly.
BMW's Leipzig deployment makes this easy to see. AEON has a humanoid upper body but moves on wheels. BMW plans to use the robot around high-voltage battery assembly and component manufacturing, which are exactly the sort of environments often cited as natural markets for humanoids.
Apptronik has reached a similar conclusion from another direction. Apollo 2 now comes in both bipedal and wheeled-base configurations. The company explicitly positions the wheeled version around stability and efficiency in high-throughput environments. Genesis AI's new Eno robot also uses a wheeled base underneath a human-scale manipulation system.
Three separate teams have therefore converged on the same practical idea. A robot can use human-height shelves, tools, carts and workstations without having human legs. Wheels remove a large part of the balancing problem and can make more sense on flat factory floors.
We would define the open market around mobile manipulation rather than strict humanoid anatomy. Legs become valuable where stairs, uneven ground or human-style mobility genuinely matter. Elsewhere, they can add engineering work without adding much customer value.
Is factory humanoid robotics still open?
Factory humanoid robotics is still very open today, especially for startups that become exceptionally good at one difficult workflow.
The latest Figure project at BMW shows why. Figure 03 is now being tested on parts sequencing, where components arrive with small changes in position and orientation and the robot must pick them, reposition its body and place them correctly. That is a meaningful step up from repeatedly loading the same sheet-metal part.
Other customers are pointing toward a much wider task set. Jabil has discussed inspection, sorting, kitting, line-side delivery, fixture placement and sub-assembly with Apptronik. BMW's AEON project adds battery assembly and component work. Mercedes-Benz has been training Apollo around intralogistics and quality-control tasks.
Most factories contain hundreds of these awkward gaps between existing automation systems. Traditional industrial arms are excellent when a movement can be fixed inside a cell. Human workers still handle many of the variable transitions between cells: presenting parts, moving racks, loading fixtures, recovering exceptions and dealing with items that do not arrive perfectly.
That creates a strong opening for companies built around the job rather than the robot. A startup could use third-party humanoids, own the integration with one production process, collect the best data for that process and charge for completed work. That looks like a more credible entry point now than spending years building another body from scratch.
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 Agility Robotics is capturing share in humanoid robotics
Is warehouse humanoid robotics still open?
Warehouse humanoid robotics remains open, but simply moving totes is already becoming an established use case rather than an untouched opportunity.
Agility has the strongest proof here. Its Digit robots have moved beyond pilot work into commercial agreements, and the company says it has secured more than $300 million in multi-year orders for Digit v5. Mercado Libre plans to start with fulfillment work in Texas, while Toyota Motor Manufacturing Canada converted its Digit pilot into a commercial Robots-as-a-Service agreement earlier this year.
Figure is attacking another part of the warehouse. Catalyst Brands plans to introduce Figure robots at its Reno distribution center around a Joey Pouch sorting system, initially targeting repetitive sorting and packing work. That broadens the evidence beyond tote transport.
We found much less public proof of humanoids running at scale in trailer unloading, mixed-item depalletization, returns processing, damaged-goods handling, exception recovery and other jobs where the objects and circumstances change constantly. Those are harder problems, but they are also where flexible manipulation starts to justify itself.
The warehouse opportunity therefore looks increasingly application-specific. A company that solves one ugly workflow across hundreds of facilities could build a serious business even if the actual robot underneath comes from someone else.
Is home humanoid robotics still wide open?
Home humanoid robotics is still wide open because nobody has shown mass-market household autonomy with the reliability and price ordinary consumers will expect.
1X is currently running the most interesting commercial experiment. NEO can be ordered for $20,000 or through a $499 monthly subscription, with US deliveries beginning in 2026. Yet 1X is unusually transparent about the limitations: NEO arrives with basic autonomy, and owners can schedule an Expert Mode session where a human operator remotely supervises difficult chores the robot does not know.
Figure has shown Helix 02 cleaning bedrooms and tidying living rooms. Genesis AI says its wheeled Eno robot is being designed for environments including homes and plans targeted customer deployments by the end of 2026. Those demonstrations show real progress in manipulation, but we still have nothing close to a large household fleet independently proving that autonomous robots can handle everyday mess for months.
Homes are vicious environments for robotics. Clothes deform, cupboards differ, objects move constantly, floors change, children and pets appear unexpectedly and a task that sounds simple can involve dozens of tiny judgments. Reliability expectations are also much harsher when a heavy autonomous machine is operating around a family.
That combination makes home robotics one of the biggest remaining prizes in humanoids and one of the riskiest places to start a company. The market is open because the problem is still unsolved.
This chart, featured in our humanoid robotics market deck, illustrates yearly funding for humanoid robotics startups
Are dexterous robot hands still worth building?
Dexterous robot hands remain one of the strongest hardware opportunities in humanoid robotics, although LinkerBot has already built a serious lead at the low-cost end.
The economics are starting to become visible. Wired reported that LinkerBot shipped about 10,000 dexterous hands in 2025 and says it represented roughly 80% of global high-degree-of-freedom demand. Its five-finger hands start around $600, and the company has been raising capital at a valuation reportedly around $6 billion.
That level of activity tells us the hand is becoming a real component market. It does not mean the design is settled. McKinsey's recent humanoid supply-chain work found that no dominant approach has emerged in tactile sensing, while force sensing and high-dexterity manipulation remain relatively immature supplier categories. Genesis AI went as far as building its own human-scale hand to collect data and transfer human skills into robots.
The open question is less about adding fingers and more about getting useful manipulation at industrial reliability. A beautiful hand that can thread a needle once is less valuable than a cheaper hand that survives months of gripping parts, detects slip, knows how much force it is applying and can be replaced quickly when something breaks.
| Humanoid hand opportunity | Why we still see room |
|---|---|
| Tactile sensing | The industry has no dominant sensing architecture yet |
| Durable industrial hands | High dexterity still has to survive huge numbers of physical cycles |
| Low-cost force-controlled hands | LinkerBot proves prices can fall quickly, which should expand demand |
| Task-specific end effectors | Many customers need reliable manipulation rather than five perfectly human fingers |
| Hand data and control software | Better hardware still needs policies that understand contact, force and slip |
Are humanoid robot actuators still worth building?
Humanoid robot actuators are still worth building, especially outside China, although a generic motor supplier will struggle to stand out.
The reason is simple: actuators still dominate the cost of the robot. McKinsey estimates actuation accounts for roughly 40% to 60% of a humanoid's bill of materials. Motors, gearboxes, bearings, encoders, drives and joint assemblies therefore have an outsized effect on price, weight, reliability and performance.
The geographic gap is even more interesting. J.P. Morgan recently published Morgan Stanley estimates putting the 2025 bill of materials for an Optimus-style robot at about $45,500 using a Chinese supply chain and $131,800 without one. Within that estimate, actuators cost roughly $22,000 in China versus $58,000 outside China. Dexterous hands show a similarly large gap.
McKinsey's supply-chain analysis helps explain it. China accounts for around 90% of permanent-magnet processing capacity and has large shares of precision bearings, encoders and power electronics. Those components benefit from China's much larger EV and electromechanical manufacturing base.
We therefore see a real opening for non-Chinese suppliers that can deliver high torque density, long life, safe backdriving, integrated force sensing or a credible domestic supply chain. Competing head-on with Chinese suppliers on a standard motor is far less appealing.
This chart, featured in our humanoid robotics market deck, compares the main business model options for humanoid robot manufacturers
Is humanoid robot training data still wide open?
Humanoid robot training data is still wide open because every serious robotics company is currently building its own way to turn scarce physical experience into something a model can learn from.
Apptronik's answer is Robot Park. The company now operates fleets of Apollo 2 robots continuously collecting real-world task data across dedicated facilities and customer sites. Figure has pursued large-scale human-video collection and direct human-to-robot transfer through Project Go-Big. Genesis AI built a data engine mixing real robot experience, simulation and synthetic data around its GENE foundation model.
The interesting problem is shifting from raw collection toward data quality. A million repetitive grasps are less useful than the right examples of failures, recoveries, unusual object positions and contact-rich actions. Companies also need to know whether data collected on one hand, arm or robot body transfers cleanly to another.
There is still no obvious standard provider for cleaning physical-action datasets, finding rare failures, scoring demonstrations, converting behavior across embodiments or measuring whether a new model is genuinely better in the real world.
That makes the data layer attractive because demand should grow even if robot hardware becomes more standardized. Every improvement in base models makes good physical data more valuable, since the remaining mistakes become increasingly specific.
Is it too late to start another robotics foundation-model company?
Starting another horizontal robotics foundation-model company looks much harder today because capital and research talent have piled into the category extremely fast.
Skild AI raised $1.4 billion at a valuation above $14 billion earlier this year. Its newest S1 model can learn tasks from a single video demonstration in context. Generalist has moved just as aggressively: Axios reported this week that the company raised another $200 million after a $400 million round only two months earlier, and its GEN-1.5 work claims one-shot adaptation from very short demonstrations.
Google DeepMind raises the bar further. Gemini Robotics 2 now covers whole-body control and cross-robot adaptation, while its On-Device 2 model can adapt to a new robot embodiment with fewer than 200 examples and only a few hours of training. Google can bring frontier multimodal models, enormous compute and existing relationships with companies such as Apptronik into the same stack.
We would still back an exceptional new research team here, but the easy story around “building the foundation model for robots” has disappeared. The better opening for most startups is one layer closer to the customer: adaptation to a specific robot, evaluation, task-specific post-training, safety constraints or a model tuned to a valuable vertical.
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 the revenue mix across customer segments in the humanoid robotics market
Is teleoperation still worth building for humanoid robots?
Teleoperation is still worth building for humanoid robots because every human intervention can help recover a failed task and create training data for the next attempt.
1X makes this visible in the consumer product itself. When NEO cannot complete a complex chore, the owner can schedule Expert Mode and have a human remotely supervise the task. Apptronik has used teleoperation and augmented reality with experienced Mercedes-Benz workers to transfer factory knowledge into Apollo. Agility's learning pipeline also uses teleoperated demonstrations alongside simulation and reinforcement learning.
Those examples suggest teleoperation will survive longer than many people expect. Better autonomy should reduce how often a human has to intervene, while simultaneously making each operator capable of supervising more robots.
The interesting businesses are therefore around intervention efficiency: deciding when autonomy should hand control to a person, connecting the right operator quickly, capturing a clean demonstration, handling several robots from one console and feeding difficult episodes back into training.
A company charging purely for remote labor will eventually face pressure as autonomy improves. Software that turns human help into faster autonomy gets more useful as the robot fleet grows.
Is humanoid robot fleet software still open?
Humanoid robot fleet software is still open because customers are beginning to put robots inside real factories and warehouses while each manufacturer currently brings its own management system.
Agility already sells Arc for facility mapping, workflow definition, fleet management and troubleshooting. UBTECH has built ROSA alongside its Robot Management Cloud Platform. These products are useful evidence that operating humanoids requires a software layer above the individual robot.
The messy part comes when a facility has Digit in one workflow, another humanoid somewhere else, AMRs moving pallets, industrial arms behind fences and an existing warehouse or manufacturing management system coordinating everything. Customers will eventually want task assignment, intervention queues, software updates, permissions, uptime tracking, cost-per-task reporting and audit logs in one place.
We have not found a dominant cross-vendor platform doing that today. OEMs naturally control the first deployments because their engineers are already on site, but independent software has a plausible opening once mixed fleets become common.
The commercial model also fits here. GXO and Toyota have used Robots-as-a-Service agreements with Agility, which moves some hardware and reliability risk away from the customer. Fleet software can eventually sit underneath those contracts and measure exactly what matters financially: robot availability, completed tasks, interventions, maintenance and actual productive hours.
This chart, featured in our humanoid robotics market deck, shows how factory humanoid robot technology has evolved over time
Is humanoid robot safety becoming its own market?
Humanoid robot safety is becoming a real market now, and the software, testing and certification stack is still far from settled.
NVIDIA made that unusually visible when it launched Halos for Robotics. The platform spans safety compute, operating software, sensors, applications and an inspection program designed to help robot companies reach third-party certification. Agility is the first humanoid manufacturer integrating parts of the stack into production robots.
Standards are moving at the same time. Industrial robot safety standards ISO 10218-1 and ISO 10218-2 received major revisions in 2025. ISO 13482, which covers personal and professional service robots and becomes especially relevant outside conventional industrial settings, is currently in final-draft approval for its second edition.
Learned behavior makes the problem harder. A traditional industrial robot can often be validated around a constrained trajectory. A humanoid running a changing AI policy has a much larger behavior space, and software updates can alter what the machine does after deployment.
That leaves room for independent testing, safety monitors, human-detection systems, simulated edge-case testing, collision-risk analysis, policy guardrails and tools that document what changed between model versions. Customers and insurers are likely to care about this well before robots become fully general.
Is humanoid robot maintenance still an open market?
Humanoid robot maintenance and uptime are still wide open because the industry has only recently accumulated enough operating hours to learn what repeatedly fails.
Figure's BMW experience offers a useful example. After running Figure 02 in production, the company identified the forearm as its largest hardware failure point and redesigned the wrist and forearm architecture in Figure 03, reducing cabling and electronics complexity. That kind of lesson rarely appears in polished demo videos; it emerges after thousands of physical cycles.
Power management belongs in the same uptime problem. Unitree lists about two hours of battery life for the G1. UBTECH took another route with Walker S2, which can autonomously swap its own batteries in roughly three minutes and is designed around continuous industrial operation. The commercial question is how many productive hours a customer gets from the machine, regardless of whether the answer comes from bigger batteries, faster charging or automatic swapping.
As robot fleets grow, maintenance should become a measurable software and service business. Joint wear, hand damage, cable fatigue, calibration drift, battery health and thermal problems all create data that can be monitored before failure.
We see room for predictive diagnostics, spare-parts logistics, independent repair networks and maintenance contracts tied directly to uptime. Those businesses look small while deployments are counted in tens or hundreds of robots. They become much more interesting if fleets move into the thousands.
In our humanoid robotics market deck, we identify pain points entrepreneurs should prioritize
What’s actually still open in humanoid robotics?
Humanoid robotics still has a lot of room for startups today, with the clearest openings around manipulation, vertical deployment, data, safety, fleet software and uptime.
Our biggest change in view comes from separating robot companies from the market around them. The number of well-funded humanoid manufacturers has exploded. At the same time, real deployments are exposing problems those manufacturers all have to solve: hands wear out, data is scarce, factories need integration, operators must handle exceptions, safety has to be certified and customers care about uptime more than impressive locomotion.
We would put vertical applications near the top of the list. A startup that becomes the best company in the world at automating one expensive physical workflow can use improving third-party hardware and foundation models instead of financing every layer itself. Dexterous manipulation, data infrastructure and reliability look similarly attractive because stronger robot adoption expands their addressable market.
Home humanoids remain the biggest unsolved end market, but they also carry the largest gap between demonstrations and dependable everyday operation. Actuators remain attractive where there is genuine differentiation or a non-China sourcing advantage. Horizontal robotics foundation models still have enormous potential, although the amount of capital entering that layer makes it a much tougher startup entry point now.
The least convincing idea today is another undifferentiated humanoid OEM. The open market is increasingly everything required to turn these increasingly capable machines into dependable workers.
| Opportunity | How open does it look now? | Our judgment |
|---|---|---|
| Vertical factory applications | Very open | Probably the strongest near-term startup wedge |
| Dexterous manipulation and hands | Very open | Major bottleneck with clear willingness to pay |
| Robot data, post-training and evaluation | Very open | Still fragmented and strategically important |
| Maintenance, diagnostics and uptime | Very open | Early market that should grow directly with deployments |
| Safety, testing and certification | Very open | Becoming necessary infrastructure |
| Fleet and deployment software | Very open | No clear cross-vendor winner yet |
| Warehouse applications beyond basic material movement | Open | Commercial proof exists, but many hard workflows remain |
| Teleoperation infrastructure | Open | Valuable when intervention improves autonomy |
| Non-China actuators and robotics components | Open | Strong need, difficult manufacturing competition |
| Home humanoid robotics | Wide open | Huge prize with very high technical risk |
| Wheeled humanoid-style mobile manipulation | Open | Increasingly credible alternative to full bipeds |
| Horizontal robotics foundation models | Crowded | Frontier opportunity with a much higher capital bar |
| Generic full-stack humanoid OEM | Very crowded | Needs a structural advantage to make sense |
If you want more recent data on this point, please see our latest humanoid robotics market report.
OUR METHODOLOGY
This analysis asks what is genuinely still open for a new startup in humanoid robotics today. We broke the market into competitive density, capital concentration, real-world deployment, technical bottlenecks, hardware economics, supply-chain maturity, enabling infrastructure and application-level gaps rather than treating humanoid robotics as a single market.
We used different evidence for different questions. Operating hours, production work and commercial agreements were given the most weight when judging whether robots are genuinely useful. Funding rounds, valuations and the concentration of leading research teams helped us judge competitive intensity. Product pricing and supply-chain estimates helped show where hardware is commoditizing and where structural cost advantages remain.
We also separated activity from maturity. A segment can attract billions of dollars and still contain major unsolved technical or commercial problems. At the same time, an immature category can already be difficult for a startup if capital, talent or platform advantages have concentrated around a small group of companies.
Repeatable evidence from operating environments was weighted more heavily than isolated demonstrations or announced roadmaps. Figure's production work at BMW, Agility Robotics' operating hours and commercial deployments, Apptronik's industrial programs, BMW's AEON project and 1X's consumer rollout therefore tell us more about current market readiness than a single impressive robot demo.
We treated humanoid robotics as an ecosystem. Full-stack robot manufacturers, hands, actuators, foundation models, training data, teleoperation, fleet software, safety, maintenance and vertical applications can be at very different stages of competitive maturity even though they all benefit from the same underlying growth in robot deployment.
The “open,” “very open,” “wide open” and “crowded” labels are editorial judgments rather than the output of a mechanical scoring model. They reflect how much unresolved customer need remains, how concentrated the existing competition has become and whether a new company could still build a meaningful and defensible advantage.
Key sources used for deployment and commercial evidence include Figure's account of Figure 02 production work at BMW, BMW Group's independent account of the deployment, Figure's Figure 03 sequencing deployment, Figure's Catalyst Brands agreement, Agility Robotics' SEC materials covering operating hours and committed Digit v5 orders, Agility Robotics' Toyota Motor Manufacturing Canada agreement, and BMW Group on the AEON deployment at Plant Leipzig.
For hardware, economics and product architecture, important sources include Apptronik on the bipedal and wheeled versions of Apollo 2, Unitree's current humanoid product range and pricing, and McKinsey's analysis of humanoid supply-chain economics, actuator costs and component manufacturing.
For robot intelligence and training infrastructure, we used Figure's Project Go-Big work on human-video-to-robot transfer, Apptronik's Robot Park data-collection program, Skild AI's Series C announcement, Generalist's GEN-1.5 work, Axios on Generalist's latest financing, and Google DeepMind on Gemini Robotics On-Device 2.
For safety and deployment infrastructure, key references include Agility Robotics' Arc fleet software, NVIDIA's Halos for Robotics safety platform, ISO 10218-1:2025, ISO 10218-2:2025, and the second edition of ISO 13482 currently in final-draft approval. For the emerging home market, we also used 1X's published NEO pricing, delivery plan and Expert Mode details.
This chart, featured in our humanoid robotics market deck, shows the regional revenue mix across Europe, Asia, North America, Africa, and South America in the humanoid robotics market
Related blog posts
- Who are the top investors in humanoid robotics?
- Is the Humanoid Robotics Market growing now?
- What are the top startups in the humanoid robotics market?
- What does the humanoid robotics startup landscape look like today?
