Which physical AI startups are closest to real-world scale?

What should “real-world scale” mean for a physical AI startup?

Real-world scale in physical AI means robots doing useful work repeatedly for customers, across enough sites and hours that the deployment has moved well beyond a pilot.

That definition separates several very different kinds of progress.

AgiBot says it has produced 15,000 embodied robots. Figure has manufactured more than 350 Figure 03 humanoids and demonstrated a production cycle of one robot per hour. Agility Robotics says Digit has accumulated more than 65,000 operating hours. Skild AI reports more than $100 million in annual recurring revenue. Zipline has completed millions of autonomous deliveries. Locus Robotics has processed more than seven billion warehouse picks.

All of those numbers sound like scale, but they measure different things.

For this comparison, we give the most weight to robots actually deployed, how often they work, whether customers expand after the first installation, whether the same system works across several sites, and whether manufacturing can keep up with demand.

A robot that has left the factory tells us something about manufacturing. A robot completing thousands of customer tasks every week tells us much more about commercial scale.

Company	Strongest evidence today	What it tells us
Locus Robotics	17,000+ robots, 360+ sites, 7B+ picks	Large repeatable commercial fleet
Zipline	Millions of commercial deliveries	Large autonomous network working outside controlled sites
Exotec	10,000+ robots and more than $1B in systems sold	Mature warehouse deployment
AgiBot	15,000 robots produced	Exceptional manufacturing scale
Unitree	5,500+ humanoids delivered in 2025 alone	Large humanoid hardware market
Skild AI	$100M+ ARR and 60+ paying customers, according to Skild	Rapid software commercialization
Agility Robotics	65,000+ operating hours and 100,000+ totes moved at GXO	Deep humanoid operating experience
Figure	350+ Figure 03 units produced and BMW deployment	Fast Western humanoid production ramp
Serve Robotics	2,000+ robots deployed	Large delivery fleet, though utilization is lower
RobCo	1,000+ industrial robots sold	Broad commercial industrial adoption

Are humanoid robots actually leading physical AI right now?

No. Humanoids currently lead the attention around physical AI, while specialized robots still lead by a wide margin in real customer activity.

Locus Robotics says more than 17,000 of its robots now work across over 360 sites. The fleet has assisted more than seven billion warehouse picks.

Zipline has already completed millions of commercial autonomous deliveries. Its drones handle healthcare, food and retail orders while flying outside tightly controlled factory environments.

Exotec has also crossed 10,000 deployed warehouse robots and more than $1 billion in systems sold. Customers include major retailers such as Decathlon, Carrefour, Gap and Uniqlo.

No humanoid startup can show comparable task volume today.

Humanoids remain interesting because one robot body could eventually handle several jobs that currently need separate machines. That flexibility could become extremely valuable. For now, specialized robots are much further along because their jobs are narrower and easier to repeat reliably.

If you want more recent data on this point, please see our latest physical AI market report.

Is China already winning the humanoid robot scale race?

Yes. China currently has the clearest lead in humanoid production volume, and the gap with most Western startups is already large.

AgiBot reached its 15,000th produced robot during the middle of the year. The company had announced its 5,000th unit around the end of 2025 and its 10,000th only a few months later, so production has been accelerating quickly.

Unitree provides another strong data point. In an official clarification published earlier this year, the company said it delivered more than 5,500 humanoid robots to customers during 2025 and produced more than 6,500. Those figures exclude its quadrupeds and wheeled dual-arm systems.

That is already several times the publicly disclosed output of leading U.S. humanoid startups.

China also has an advantage deeper in the supply chain. Motors, actuators, batteries, gears and electronics can often be sourced locally, while many Western robotics startups still depend heavily on Asian component suppliers.

Commercial utilization is harder to compare. AgiBot has shown robots working at Longcheer Technology on tablet-production tasks including loading, unloading, transfer and inspection, but Chinese companies generally publish less standardized operating data than companies such as Agility or Locus.

So China's lead is clearest in manufacturing. At this point, the gap is too large to dismiss.

Are AgiBot and Unitree already at mass humanoid deployment?

AgiBot and Unitree are the closest humanoid startups to mass hardware deployment, although neither has yet proved that thousands of its robots are doing productive commercial work every day.

AgiBot has produced 15,000 robots, putting it in a category of its own among young embodied-AI companies. Its progression from 5,000 produced robots to 10,000 and then 15,000 also shows that manufacturing is no longer the main constraint.

The harder part is working out what all those robots are doing.

AgiBot describes deployments across factories, logistics, service applications and data-collection environments. The Longcheer deployment is useful evidence because robots there perform repetitive production-line jobs rather than staged demonstrations.

But the company has yet to publish a simple fleet-wide utilization number. We do not know how many of those 15,000 robots work every day, how many hours they operate, or how much customer labor they replace.

Unitree provides a different type of scale.

The company disclosed more than 5,500 humanoid deliveries during 2025, with production above 6,500 units. Unitree explicitly clarified that the number referred to robots sold and delivered to end customers rather than orders or production targets.

Its installed base spans universities, laboratories, developers, integrators and companies experimenting with embodied AI.

That breadth creates a large ecosystem, but it also makes the commercial picture harder to read. A university buying a G1 for research does not prove the same economics as a factory ordering 200 robots after a successful year-long deployment.

AgiBot therefore looks stronger on manufacturing volume, while Unitree has already built one of the broadest humanoid hardware ecosystems.

Both have moved far beyond prototype-scale production. What neither has yet shown publicly is thousands of humanoids earning their keep every day across commercial operations.

If you want more recent data on this point, please see our latest physical AI market report.

Is Figure AI close to running humanoids at real industrial scale?

Figure is currently one of the closest Western humanoid startups to large industrial deployment, although its active customer fleet is still small compared with mature warehouse robotics.

BMW gives us the most useful evidence.

Figure 02 robots worked inside BMW's Spartanburg plant on a production-line workflow. Figure says the robots accumulated more than 1,250 operating hours, handled over 90,000 parts and contributed to production involving more than 30,000 vehicles.

BMW has since started working with Figure 03 on a more complex logistics-sequencing task where the robot manipulates components while repositioning itself and moving a cart.

That progression is useful because BMW has already seen the previous robot work and is continuing with another generation and another task.

Figure is also starting to remove the manufacturing constraint.

Its BotQ facility had produced more than 350 Figure 03 robots by late April. Figure says the factory improved from roughly one robot per day to demonstrating a one-robot-per-hour production cycle in less than four months.

More recently, Figure has shifted much of its public attention toward training data and compute. Its Index program is collecting large amounts of humanoid data, while a partnership with Nscale could eventually give Figure access to up to 100,000 NVIDIA Vera Rubin GPUs.

The missing piece remains fleet size in customer production. We have BMW, Catalyst Brands and a fast-growing internal fleet, but not yet several hundred Figure robots simultaneously earning money across dozens of independent sites.

Is Agility Robotics actually ahead of Figure in humanoid deployment?

Yes. If we care most about how long humanoid robots have actually worked inside customer facilities, Agility Robotics currently has the stronger record.

Digit has logged more than 65,000 operating hours across customer environments. At GXO's Flowery Branch site, the robot has moved more than 100,000 totes while Agility reports roughly 98% accuracy when Digit is on task.

Agility says its robots have also worked with companies including Schaeffler, Amazon and Toyota Motor Manufacturing Canada.

The newest development makes this more commercially interesting. Agility now reports more than $300 million in multi-year orders for Digit 5, subject to contractual milestones.

Digit 5 was built partly around problems discovered during nearly three years of real deployment. One of the biggest changes is cooperative safety: Agility wants the robot to work around people without the safety barriers that limit conventional industrial automation.

Figure still appears to be moving faster on manufacturing capacity and general-purpose AI. Agility has the better operating history today.

If we had to choose which Western humanoid company has already proved the most in factories and warehouses, we would currently pick Agility.

If you want more recent data on this point, please see our latest physical AI market report.

Has Skild AI already turned robot foundation models into a real business?

Yes. Skild AI has moved from robotics research into commercial deployment unusually fast, and its latest numbers put it well ahead of most physical-AI model startups.

Skild says it passed $100 million in annual recurring revenue only ten months after its first commercial deployment. It reports more than 60 paying customers and says it has already recognized around $50 million in revenue.

The customer mix is what makes the number more interesting.

According to Skild, only around 10% of its revenue comes from mobility, including the autonomous mobile robot business associated with Fetch Robotics. Roughly 90% comes mainly from manipulation.

Skild is also spreading the same intelligence across different hardware. Its partners include ABB Robotics, Universal Robots, NVIDIA and Foxconn. At Foxconn, Skild's software is being deployed on dual-arm manipulators involved in assembly work for NVIDIA Blackwell systems.

A software company can potentially scale faster than a robot maker because every new customer does not require Skild to manufacture an entire machine.

The caveat is financial transparency. Skild is private, and the $100 million ARR figure comes from the company itself. ARR can also include recurring contracted business that differs from revenue already recognized.

Still, the $50 million recognized-revenue figure and 60-plus paying customers make the commercialization hard to dismiss.

As of now, Skild has the strongest public case for being the first robot-foundation-model startup to build a sizeable business.

Is FieldAI now a serious physical AI scale contender?

Yes. FieldAI has become one of the most commercially important physical-AI startups surprisingly quickly, even though its operating data remains thinner than Skild's.

Recent reporting says FieldAI has secured more than $135 million in revenue and customer contracts from more than 30 customers.

Those customers span industries including construction, data centers and defense, which fits FieldAI's core pitch: one general robot intelligence that can operate different machines in messy environments.

FieldAI was founded only in 2023, yet recent financing discussions have reportedly valued the company around $10 billion. The valuation itself tells us little about deployment, but the reported customer contract volume suggests investors are reacting to more than laboratory progress.

We would still rank Skild ahead commercially because Skild gives us clearer numbers for recognized revenue, paying customers and the split between manipulation and mobility.

FieldAI publishes less about active robots, hours of autonomous operation or tasks completed per month.

For now, FieldAI has clearly crossed into meaningful commercial deployment, although the depth of that deployment remains harder to measure.

Is Physical Intelligence commercially behind Skild AI?

Yes. Physical Intelligence looks like one of the strongest robotics-model research companies, while Skild is currently much further ahead in visible commercialization.

Physical Intelligence is trying to build general vision-language-action models that work across different robot bodies and tasks.

The company has already shown that those models can leave the lab.

Ultra, for example, has used Physical Intelligence models on robots packaging real e-commerce orders. Ultra reported that one version of the model reached 96.4% autonomous operation across a complete shift at a customer site.

Yet the scale is still much smaller than what Skild reports.

Physical Intelligence has deliberately focused heavily on the underlying model rather than rushing into a large sales organization. Its public communications still revolve mainly around better generalization, new model generations and expanding the range of tasks robots can perform.

Today, Physical Intelligence is closer to proving a powerful general robotics model than proving a large physical-AI business.

Are industrial startups like RobCo closer to scale than the flashy humanoid companies?

Yes. RobCo and several specialized industrial robotics startups are already selling large numbers of machines because they began with jobs factories know they need automated.

RobCo has now sold more than 1,000 robots to customers including BMW.

Its systems work on jobs such as machine tending, palletizing and material handling. These tasks are repetitive, easy for a factory manager to understand and usually come with a clear labor cost attached.

RobCo is now adding more AI to that existing commercial base.

Its new dual-arm Alfie system is designed to learn changing tasks without traditional reprogramming. That means RobCo can move toward more general physical AI without first having to convince customers that factory robots are worth buying.

The company has also just crossed a $1 billion valuation after an employee share transaction and new investment, doubling its valuation in roughly nine months.

The valuation matters less than the installed base, but the combination is telling: investors are now paying frontier-AI prices for a company that already has more than 1,000 industrial robots in customers' hands.

Dexterity, Ambi Robotics and Pickle Robot fit the same broader pattern.

Dexterity is expanding robotic trailer-loading work with FedEx. Pickle focuses on unloading trailers and has received follow-on orders from customers such as Ryobi. Ambi Robotics automates parcel sorting and palletizing for large logistics operators.

Each company has chosen a painful job where the customer immediately understands why a robot could be valuable.

That gives them a simpler path to deployment than a humanoid company trying to make one machine useful across many jobs.

If you want more recent data on this point, please see our latest physical AI market report.

Are Serve Robotics' 2,000 robots actually being used?

Only partly. Serve Robotics has built a large robot fleet, but its own numbers show that deploying 2,000 robots and keeping 2,000 robots busy are very different things.

Serve had more than 2,000 robots in its fleet by the end of 2025.

During the second quarter of this year, however, the company averaged 792 daily active robots. Serve defines that metric as robots actually performing deliveries during the period.

The gap is substantial.

Some robots are naturally being held for maintenance, market launches, fleet balancing or spare capacity, so we would never expect 100% utilization. Still, fewer than half of the headline fleet being active on an average day shows why installed robot counts need context.

Serve generated about $3.2 million in second-quarter revenue. That was more than four times the level of a year earlier, yet sequential growth was only around 9%. The company also reduced its full-year revenue outlook after delivery volumes through Uber came in below previous expectations.

Its acquisition of Diligent Robotics adds another useful piece. Diligent's Moxi robots have already completed more than one million hospital deliveries across healthcare facilities.

Serve has therefore reached meaningful fleet scale, but the utilization and economics are still catching up.

Why is Zipline already much closer to scale than most physical AI startups?

Zipline has already crossed the line from robotics startup into autonomous infrastructure.

The company passed two million commercial deliveries early this year. By the middle of the year, it reported more than 2.5 million, with roughly one million of those deliveries completed during the previous twelve months.

That acceleration is more useful than the absolute milestone.

Zipline spent years building its first million commercial deliveries. It added the next million much faster, and its U.S. operation is now becoming a much larger part of the business.

During the first half of the year, Zipline said the number of businesses available through its U.S. delivery marketplace increased thirteenfold. Roughly 70% of its flights were taking place in the United States by July.

Those flights are fully autonomous.

Zipline therefore has large cumulative task volume, fast current growth, repeat customer use and international deployment at the same time.

Its newer Platform 2 system could accelerate the shift from healthcare logistics toward everyday home delivery.

The company has also partnered with Uber around a much larger future network. Uber and Zipline have discussed reaching one million drone deliveries per day by the end of the decade.

That target remains far away, but Zipline starts from a radically stronger base than most physical-AI forecasts.

Among startups operating robots in uncontrolled real-world environments, Zipline currently has one of the strongest scale stories anywhere in physical AI.

Are Locus Robotics and Exotec still relevant to the physical AI race?

Absolutely. Locus Robotics and Exotec show what physical AI looks like once customers stop treating the robots as experiments.

Locus now has more than 17,000 robots across over 360 facilities, and its fleet has processed more than seven billion warehouse picks.

That installed base also keeps getting more capable.

Locus is moving deeper into robotic manipulation through systems such as Locus Array, pushing its existing fleet model toward increasingly autonomous warehouse workflows.

Exotec has followed a similar path.

The company has passed 10,000 robots and more than $1 billion in systems sold. Customers have gone on to expand deployments across several warehouses and countries.

Decathlon is a good example. Rather than using Exotec in one showcase facility, the retailer has expanded the system across multiple European sites.

That repeat purchase is one of the clearest forms of evidence in robotics because the customer has already dealt with installation, workflow changes, maintenance, software integration and daily operation.

Commercially, Locus and Exotec are already where most humanoid startups are trying to get.

Could robot foundation models scale faster than humanoid robot companies?

Yes. A company like Skild AI could potentially spread physical intelligence much faster than a vertically integrated humanoid startup can manufacture complete robots.

ABB and Universal Robots already have huge hardware footprints inside factories.

If Skild can make existing industrial robots handle more variable jobs, every software improvement can potentially reach machines that have already been installed.

FieldAI is pursuing a similar idea across a broader range of robots and environments. Physical Intelligence is trying to make general models portable between embodiments as well.

This creates a very different scaling model from Figure or Agility.

Figure has to design the hardware, secure components, manufacture robots, service them and improve Helix at the same time. A foundation-model company can let someone else handle much of the hardware layer.

The limitation is that robots vary far more than computers.

Different robot bodies have different sensors, joints, payloads, control systems and safety limits. A model that works perfectly on one arm cannot automatically be dropped onto another machine without adaptation.

That makes Skild's reported customer count particularly useful.

More than 60 paying customers across factories, warehouses, kitchens, construction sites and data centers would suggest the model is already surviving meaningful changes in hardware and environment.

If that pattern continues, some of the largest physical-AI companies may eventually sell intelligence rather than robots.

Company	Main approach	Commercial evidence today	Where we put it
Skild AI	General robot foundation model	$100M+ ARR claimed, 60+ customers	Closest model startup to scale
FieldAI	General autonomy across robot types	$135M+ in reported revenue/contracts, 30+ customers	Strong commercial contender
Physical Intelligence	General VLA models	Production work through partners	Strong technology, earlier business
Figure	Helix models tied closely to Figure hardware	BMW and Catalyst deployments	Strong vertical approach
AgiBot	AI plus several proprietary robot forms	Very large manufactured fleet	Hardware scale already substantial
Apptronik / DeepMind	Apollo hardware with external AI partnership	Industrial customer programs	Still early in deployment scale

Which physical AI startups are actually closest to real-world scale today?

Zipline, Locus Robotics and Exotec are currently the clearest large-scale physical AI operators; AgiBot and Unitree lead humanoid hardware volume; Skild AI leads among general robot-model startups; and Agility Robotics and Figure have the strongest Western humanoid cases.

The companies are solving different parts of the problem, so forcing all of them into one simple ranking would hide more than it reveals.

Zipline belongs at the top when we care about autonomous machines repeatedly working in the open world. Millions of commercial deliveries give it a level of real-world experience few robotics startups can match.

Locus Robotics and Exotec have reached comparable maturity inside warehouses. Their fleets are large, their task volumes are enormous and customers repeatedly expand deployments.

AgiBot has already reached a level of robot manufacturing that most humanoid startups still talk about as a future goal. Unitree has also built a genuinely large humanoid installed base, with more than 5,500 customer deliveries in a single year.

Skild AI is the standout software company. Its reported $100 million-plus ARR, $50 million in recognized revenue and more than 60 paying customers make it much more commercially advanced than the usual robotics-model startup.

Agility Robotics currently has the strongest Western evidence of humanoids doing sustained industrial work. More than 65,000 operating hours, 100,000-plus totes at GXO and over $300 million in multi-year orders give the company a much deeper commercial record than a typical pilot-stage robotics company.

Figure is moving faster on manufacturing and AI ambition. More than 350 Figure 03 robots, the BMW relationship and the BotQ production ramp make it one of the most likely companies to close the fleet-size gap.

RobCo deserves to sit surprisingly high. It has sold more than 1,000 robots, works with customers such as BMW and is now layering more general AI onto an established industrial business.

FieldAI also deserves more attention than it received even a few months ago. More than $135 million in reported revenue and contracts from over 30 customers would put it firmly inside the commercial group, although we still want better operating metrics before placing it beside Skild.

Dexterity, Ambi Robotics and Pickle Robot have narrower ambitions, but they are solving factory and logistics tasks that customers already pay to automate.

Serve Robotics has reached meaningful fleet size, while its own daily-active-robot numbers show how much work remains on utilization.

Apptronik remains one of the most credible companies immediately behind this group. Mercedes-Benz, Jabil, GXO and Google DeepMind give it an exceptional partner base, but we still need larger public operating numbers before calling Apollo a scaled deployment.

1X is earlier again. Putting NEO robots into homes could eventually generate one of the richest physical-AI datasets in the industry, but household commercialization is only beginning.

Physical Intelligence may eventually become more important than several companies ranked ahead of it here. Today, its commercial footprint is simply smaller.

Position today	Company	Why
Already at large scale	Zipline	Millions of autonomous commercial deliveries
Already at large scale	Locus Robotics	17,000+ robots and 7B+ warehouse picks
Already at large scale	Exotec	10,000+ robots and repeat multi-site deployments
Humanoid manufacturing leader	AgiBot	15,000 robots produced
Humanoid hardware leader	Unitree	5,500+ humanoids delivered in 2025
Physical AI software leader	Skild AI	$100M+ ARR claimed across 60+ customers
Most proven Western humanoid	Agility Robotics	65,000+ operating hours and $300M+ in orders
Fastest-scaling Western humanoid	Figure	350+ Figure 03 robots and rapid BotQ ramp
Established industrial contender	RobCo	1,000+ robots sold
Fast-rising model contender	FieldAI	$135M+ reported revenue/contracts
Real fleet, weaker utilization	Serve Robotics	2,000+ fleet but fewer than 800 daily active robots in Q2
Scaling specific workflows	Dexterity	Expanding industrial work with FedEx
Scaling specific workflows	Ambi Robotics	Production parcel-handling deployments
Scaling specific workflows	Pickle Robot	Repeat orders in trailer unloading
Approaching scale	Apptronik	Strong industrial partners, limited public fleet data
Earlier commercial stage	Physical Intelligence	Strong real-task results through partners
Earlier commercial stage	1X	Consumer humanoid deployments just beginning

Physical AI already works at real-world scale when the task is narrow enough. Warehouse movement, order picking, drone delivery and several industrial manipulation jobs have crossed that line.

General-purpose physical AI is much earlier.

No startup today has tens of thousands of general-purpose robots autonomously working across a wide range of unrelated customer tasks with consistently proven economics.

AgiBot is closest to that volume on the hardware side. Skild is closest to broad scale on the intelligence layer. Agility has the strongest Western humanoid operating record. Figure currently looks best positioned to combine increasingly general AI with a rapidly expanding humanoid fleet.

The company that eventually combines generality, high utilization, repeat customer orders and mass manufacturing will have achieved something the physical AI industry still has not.

If you want more recent data on this point, please see our latest physical AI market report.