Humanoid Robotics: what is actually working now?

Are humanoid robots actually doing useful work now?

Humanoid robots are doing useful work today, but almost all of the convincing evidence still comes from a small set of repetitive factory and warehouse jobs.

The strongest cases have something important in common: we can count the work.

At BMW’s Spartanburg plant, Figure 02 spent roughly 1,250 hours on the production floor, moved more than 90,000 sheet-metal parts and contributed to production of more than 30,000 BMW X3 vehicles. BMW says the robot worked ten-hour shifts, five days a week during the deployment.

Agility Robotics has crossed a similar threshold in logistics. Digit has now moved more than 100,000 totes at GXO’s Flowery Branch facility. GXO started testing Digit in 2023 and later signed a commercial Robots-as-a-Service agreement, so this progressed well beyond an experiment that disappeared after a few weeks.

China now gives us a third type of evidence. AGIBOT and electronics manufacturer Longcheer ran eight G2 robots inside a real tablet-production workflow during a six-day factory validation. Longcheer reported more than 64 hours of robot operation, 64,828 production tasks and 17,625 units of line output across the test.

Those numbers finally give us something better than polished demonstration videos. Humanoids can already survive thousands of repetitions inside operating factories and warehouses.

The scale is still tiny beside conventional automation. The International Federation of Robotics counted 4.66 million industrial robots operating worldwide at the end of 2024, with another 542,000 installed during that year. Verified humanoid deployments remain a rounding error beside that installed base.

So humanoid robotics has genuinely started working, just in a much narrower way than the idea of a robot replacing a person across an entire shift.

What we can verify	Best evidence today
Repetitive factory handling	Figure at BMW
Warehouse tote movement	Digit at GXO
Multi-robot electronics work	AGIBOT at Longcheer
Large-scale general labor	Still unproven

What should count as a humanoid robot that actually works?

A humanoid robot should count as “working” when it repeatedly completes a useful customer task in normal operations without engineers constantly rescuing or staging the job.

That sounds obvious, but it changes how we read most humanoid announcements.

A robot picking up a strange object once proves that the manipulation is possible. Walking across a stage proves locomotion. Producing 10,000 units proves manufacturing capability. Signing a large order proves that customers are interested.

Operational humanoid robotics needs more than that.

The useful progression is fairly simple. First comes a demonstration. Then a customer pilot. After that comes a robot doing repeated production work. The real test of scale arrives when the same system can be deployed across several sites without months of custom engineering each time.

Agility offers a good example because its commercial disclosures make that progression unusually visible. The company reported nine committed customer-facility deployments and more than 65,000 hours of robot operations across its deployment base as of May. Customers include GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre.

Those 65,000 hours should be read carefully. They cover Agility’s robot operations rather than 65,000 hours of completely autonomous paid production. The customer-specific workload at GXO is more useful because we know what Digit was actually doing and how often it did it.

That is the standard we should keep throughout the humanoid market: real task, real site, repeated work, measurable output.

Which humanoid robot has the strongest real-world track record?

Agility Robotics’ Digit currently has one of the strongest commercial track records because we can follow the robot from pilot to paid deployment and then into repeated work across several customers.

GXO first tested Digit before signing a multi-year commercial deal. Digit subsequently passed 100,000 totes moved at the company’s Flowery Branch logistics operation.

Toyota Motor Manufacturing Canada followed a similar path. Toyota piloted Digit and then signed a commercial agreement in February 2026 to use the robot in manufacturing, supply-chain and logistics work.

Mercado Libre has also signed a commercial deployment agreement. Schaeffler has used Digit for industrial material handling. Agility says nine customer facilities are now committed to deployments.

The more interesting number may be what comes next. Agility disclosed more than $300 million of multi-year orders for its upcoming Digit v5. The figure covers roughly 1,000 robots under multi-year arrangements and remains subject to contractual milestones, so we should not treat it as current revenue.

Still, a customer pattern is forming. One logistics company did not simply test Digit and walk away; other large industrial companies are now moving through the same pilot-to-commercial sequence.

For humanoid robotics, that repeated buying behavior is more persuasive than almost any single robot demo.

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

Has Figure proved that humanoid robots belong in car factories?

Figure has proved that a humanoid robot can handle one real automotive-production job for a meaningful amount of time, and BMW’s results are among the cleanest evidence we have.

Figure 02’s job at BMW was simple to describe and hard to fake at scale: retrieve sheet-metal parts and position them for welding.

Across the deployment, BMW recorded more than 90,000 handled components, around 1,250 operating hours and roughly 1.2 million robot steps. Figure 02 contributed to production of more than 30,000 X3 vehicles.

That works out to roughly 72 handled components per operating hour across the reported runtime. More importantly, BMW describes the task as repeatable production work requiring millimetre-level positioning accuracy.

The deployment also showed the hidden work behind a successful humanoid installation. BMW changed parts of its safety setup, added barriers and partitions and improved 5G coverage around the work area.

That makes the result more credible, actually. Figure 02 could perform the job, while the factory still had to be adapted around the robot.

Figure has since brought Figure 03 back to Spartanburg for a harder logistics-sequencing workflow. The newer robot has to find parts that are not always sitting in exactly the same position, manipulate them while repositioning its body and pull carts through the workspace.

Figure’s Helix 02 system controls the hands, arms, torso and legs together during these movements. We have seen strong demonstrations of the new workflow, although we do not yet have the same long-duration production record that Figure 02 built.

Figure has therefore proved the first step. The current Figure 03 deployment is testing whether the company can move from one repetitive industrial motion toward work with more variation.

Figure at BMW	Reported result
Figure 02 operating time	1,250+ hours
Components moved	90,000+
X3 vehicles supported	30,000+
Estimated robot steps	1.2 million+
Figure 03 current challenge	Parts sequencing and logistics

Can humanoid robots actually do several different jobs?

Humanoid robots can already learn several useful skills, but nobody has publicly shown one robot handling anything close to the range of work a normal employee handles during a shift.

This is where the phrase “general-purpose robot” gets ahead of the evidence.

Digit can walk, carry containers, manipulate totes and perform more than one material-handling workflow. Figure 03 can combine walking, grasping and cart movement. AGIBOT’s G2 has worked on tablet inspection, material transfer and sorting. UBTECH’s industrial robots have been demonstrated across logistics, inspection and automotive manufacturing tasks.

That is genuine flexibility. The robots can deal with more variation than a fixed industrial arm following the same programmed trajectory all day.

A normal warehouse or factory worker still operates on another level. Someone might replenish one station, notice damaged packaging, clear an obstruction, find an unexpected part and help a colleague within the same shift. Humans handle new situations without a robotics team collecting demonstrations and retraining a policy first.

The industry is gradually shrinking that gap. Figure’s move from sheet-metal loading to parts sequencing is a good example because the second task requires much more perception and whole-body coordination.

For now, “general purpose” describes the direction of travel. The jobs being done today are still quite specialized.

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

Is humanoid robot AI autonomous enough to be useful?

Humanoid robot AI is autonomous enough to create value inside carefully chosen workflows, while the surrounding deployment still needs much more engineering than hiring another person.

Figure shows how far the software has moved. Its Helix models turn camera inputs into robot actions, and Helix 02 coordinates the robot’s full body rather than treating walking and manipulation as completely separate systems. Figure 03 can adjust its grasp and body position when parts appear differently inside a container.

Agility mixes learned behaviors with more traditional control, teleoperation data, simulation and reinforcement learning. Digit can navigate facilities, manipulate objects and return to charge while Agility’s Arc software manages the workflow and fleet.

AGIBOT’s Longcheer deployment gives us a more quantitative example. Across more than 64 hours of operation, the company reported 64,828 production tasks and a 99.99% task-success rate during the six-day factory program. Longcheer separately said the eight robots ran through the public validation without an interruption.

Those are company-reported figures, and we should treat them accordingly. Even so, a multi-day live production run gives us much more information than a short edited video.

The limitation appears when the environment stops being predictable. Setting up a production deployment can still involve months of task selection, onsite testing, safety work, data collection and software tuning.

Today’s autonomy works best when we first make the problem manageable.

Why are humanoid robots mostly moving boxes and parts?

Humanoid robots currently work best at moving totes, bins, components and other predictable objects because those jobs give the robots lots of repetition without asking for human-level judgment.

Once we put the serious deployments side by side, the pattern becomes almost comically obvious.

Digit’s best-documented workload is tote movement. Figure 02 loaded sheet metal. Figure 03 is working on parts sequencing. AGIBOT has robots loading, transferring and sorting electronics components. Mercedes-Benz has tested Apptronik’s Apollo around repetitive manufacturing and intralogistics jobs.

Factories like these tasks because performance is easy to measure. Either the object reaches the correct place at the required pace or it does not.

Robot developers like them for another reason: repetition produces data. A factory robot may attempt essentially the same manipulation thousands of times. Every failure, awkward grasp and unusual object position can feed the next software iteration.

Household chores offer nothing like that consistency. Folding one shirt, clearing one table and finding a dropped toy are three different problems in three different physical setups.

So the first big humanoid use case was probably never going to look spectacular. Moving boring objects for hours is exactly where the technology has found its first solid footing.

Do factories really need robots with two legs?

Factories often need robots that fit human workplaces, but two legs themselves are much less essential than the humanoid industry sometimes makes them sound.

The argument for human-shaped machines is straightforward. Factories and warehouses were designed around our height, reach, aisles, shelves, doors, carts and tools. A robot with arms and roughly human proportions can potentially automate a job without rebuilding the whole facility.

Figure’s BMW sequencing workflow shows that advantage. Figure 03 reaches into existing containers, walks around carts and moves parts through infrastructure designed for people.

Digit takes the same idea into logistics. Its shape lets the robot work between autonomous mobile robots, conveyors and warehouse equipment without installing a new fixed automation cell for every movement.

Several manufacturers are quietly showing us that legs are optional.

Apptronik has developed Apollo in both bipedal and wheeled forms. AGIBOT sells wheeled embodied robots for industrial work. BMW’s new Leipzig pilot uses Hexagon’s AEON, whose humanoid upper body sits on a wheeled mobility system.

That is an important clue about where the market may go. Arms, hands, useful reach and the ability to move through existing spaces create most of the practical advantage. Wheels are simpler whenever the floor allows them.

Two legs become worth the extra complexity when stairs, obstacles or a truly human-specific workspace demand them.

The industrial winner may look human enough to use our world without copying every part of the human body.

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

Is China actually winning the humanoid robot race?

China is currently winning the manufacturing race in humanoid robotics, while leadership in proven autonomous work is still shared between several Chinese and American companies.

The production gap is becoming difficult to ignore.

AGIBOT says its 15,000th robot rolled off the line in June 2026. The company had needed roughly a year to move from its first 1,000 robots to 5,000, then only three months to reach 10,000. Omdia counted 5,168 AGIBOT humanoid shipments in 2025, equal to 39% of the market it tracked.

Figure gives us a useful comparison. Its BotQ factory had produced more than 350 Figure 03 robots by late April, while Figure said production speed had improved from roughly one robot per day to one per hour.

UBTECH has now added financial evidence to China’s manufacturing story. According to the company’s first-half 2026 accounts released in August, revenue from full-size embodied humanoid products and services reached RMB590.3 million, up 1,445% from RMB38.2 million a year earlier. UBTECH sold 921 full-size units during the six months, compared with 45 a year earlier.

Full-size humanoids consequently went from 6.1% to 46.5% of UBTECH’s total revenue in twelve months. That is one of the clearest signs we have seen that Chinese humanoids are becoming an actual business rather than an R&D category.

We still have to separate robots sold from robots doing autonomous production work. AGIBOT’s 15,000-unit figure spans a broader robot portfolio and several use cases. UBTECH’s shipments likewise include customers at different stages of deployment.

China has another structural advantage underneath the humanoid race. IFR says the country installed 295,000 conventional industrial robots in 2024, representing 54% of worldwide installations, and already had more than two million industrial robots operating.

That gives Chinese humanoid companies customers, suppliers, integrators and manufacturing knowledge on an unusually large scale.

The U.S. still has some of the strongest individual production records through Digit and Figure. China currently looks stronger when we measure how quickly companies can manufacture machines, ship them and build a domestic ecosystem around them.

Current race	Strongest position
Humanoid manufacturing volume	China
Publicly documented long-running U.S. customer work	Agility / Figure
Full-size humanoid revenue disclosure	UBTECH
Existing industrial-robot ecosystem	China
Clear overall winner	Too early

Are customers really spending money on humanoid robots?

Customers are now spending serious money on humanoid robots, although much of the biggest contract value still depends on robots being deployed successfully over the next few years.

Agility gives us the clearest American example.

The company disclosed more than $300 million of multi-year Digit v5 orders as of May 2026, covering roughly 1,000 robots. Most of those contracts use a Robots-as-a-Service structure, where customers pay over time rather than buying the hardware upfront.

The important detail is that those amounts depend on contractual milestones. We cannot simply turn $300 million of orders into $300 million of present-day revenue.

UBTECH gives us the opposite view: actual recognized sales.

Its latest first-half accounts show RMB590.3 million of revenue from full-size humanoid robots and services, roughly $88 million at recent exchange rates. The category generated almost half of UBTECH’s group revenue during the period.

The comparison is useful. Agility shows that large Western industrial customers are willing to make multi-year commitments before the next hardware generation fully ramps. UBTECH shows that hundreds of full-size robots can already turn into meaningful booked revenue.

We have moved beyond a market funded purely by venture capital and research budgets.

Now comes the tougher test: do customers order the second hundred robots after living with the first ten?

Can humanoid robots actually be cheaper than workers?

Humanoid robots can probably beat human labor costs in a few high-utilization jobs, but we still lack enough customer data to say that the economics work routinely.

Agility has published one of the clearest cost models in the industry.

For Digit v5, the company’s investor materials model roughly $500,000 of cumulative five-year revenue per robot under Robots-as-a-Service and about $400,000 under an ownership model. The ownership example includes an estimated $200,000 robot price, a deployment fee and continuing software and maintenance costs.

Agility assumes that a mature Digit deployment could eventually work two ten-hour shifts per day. At that utilization level, spreading the machine’s cost across thousands of productive hours starts to look attractive beside fully burdened industrial labor.

We should be careful with that model because Agility labels the numbers illustrative. Robot downtime, maintenance, onsite engineering and human interventions can quickly change the result.

BMW’s deployment shows why. Figure 02 did productive work, but BMW also needed safety changes, stronger connectivity and integration work around the robot.

A humanoid becomes economically interesting once one technician can oversee many reliable robots and each machine keeps working for long shifts with few interventions. If one engineer has to babysit each robot, the labor-saving argument collapses.

As of now, we have a credible route to good economics in repetitive high-utilization work. We do not yet have enough independent customer numbers to call cheap humanoid labor a solved problem.

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

How reliable do humanoid robots need to become?

Humanoid robots need reliability far beyond “usually works,” because even a tiny failure rate becomes painful once a machine repeats the same action thousands of times per shift.

This is one area where normal AI benchmarks can be misleading.

Imagine a robot completing a task correctly 99% of the time. Across 10,000 attempts, that still implies 100 bad outcomes unless the robot catches and fixes its own mistakes. At factory scale, that can be disastrous.

That is why the recent long-duration tests deserve attention.

AGIBOT says its Longcheer robots completed 64,828 production tasks with a 99.99% task-success rate across the six-day program. Even at that reported level, the more important question is what happens during the failures: does the robot recover by itself, retry, call for help or stop the line?

BMW’s Figure deployment gives us a different kind of reliability proof. More than 90,000 component movements across 1,250 operating hours tell us the robot survived enough repetition to expose problems that a twenty-minute demo would never find.

Safety sits inside the same problem. BMW still used barriers and partitions during the Figure 02 program. Agility is designing Digit v5 around cooperative safety so that the robot can operate more freely around human coworkers.

Those restrictions affect productivity. A robot that has to work slowly, remain behind a barrier or stop every time somebody comes near may be technically safe while still making poor economic sense.

The numbers we really want from humanoid companies now are boring industrial ones: uptime, mean time between interventions, autonomous recovery rate and productive cycles between human assists.

When companies routinely publish those figures across months rather than hours, we will know humanoids are becoming normal industrial equipment.

Is Tesla Optimus actually doing factory work yet?

Tesla Optimus is still behind the best-documented commercial humanoid deployments today, even though Tesla could become extremely dangerous once its manufacturing ramp really starts.

Tesla has shown Optimus performing factory-style manipulation and intends to use its own factories as the first major proving ground.

The company’s latest formal updates still place much of the story around production preparation. Tesla has been installing dedicated Optimus manufacturing lines, while management has described the humanoid as an unusually difficult manufacturing ramp because so many components and production processes are new.

That distinction matters when we compare Tesla with Figure or Agility.

BMW can tell us how many components Figure moved and how many production hours it logged. GXO can point to more than 100,000 totes moved by Digit. We still do not have an equivalent long-duration production dataset for Optimus.

Tesla’s advantage sits elsewhere. The company already knows how to manufacture electromechanical products in enormous volumes, develops AI systems internally and controls much of its motors, electronics, battery and compute stack.

If Optimus reaches robust autonomy, Tesla has a plausible path to scaling hardware much faster than a robotics startup.

Today, though, the robot race should be judged on completed work rather than manufacturing potential. By that standard, Optimus remains a major contender with more to prove on the factory floor.

Are humanoid robots ready to work in our homes?

Home humanoid robots are still early, and the gap with factory robots is much larger than the marketing makes it look.

1X is currently making the most direct push into ordinary homes with NEO.

The company has opened a 58,000-square-foot factory in California with capacity for up to 10,000 NEO robots per year. 1X plans to expand manufacturing much further as production becomes more automated, and early customer machines are expected to enter homes during 2026.

The hardware is moving toward a real consumer product. The autonomy problem remains brutal.

A warehouse can standardize the lighting, floors, containers, wireless network and task flow. Homes contain loose clothing, glassware, food, pets, children, stairs, doors, clutter and objects that may appear in a different place every day.

A factory task may repeat 5,000 times. Someone at home might ask a robot to empty a dishwasher once, find a sock once and clean up a completely different mess an hour later.

1X has deliberately used human assistance and extensive in-home data collection to close that gap. That approach could make early NEO units genuinely useful because a human operator can help when the autonomous system gets stuck. It also tells us how far full autonomy still has to go.

Figure 03 is being designed with eventual home use in mind as well. The robot is lighter, uses softer exterior materials and includes safety and charging improvements that make much more sense in a house than earlier industrial prototypes.

But the evidence remains heavily tilted toward factories.

If someone wants a humanoid that can walk around an ordinary home and reliably take over a broad range of chores by itself, we are not there yet.

So what is actually working in humanoid robotics right now?

Humanoid robotics is genuinely working today in repetitive industrial jobs, while the dream of a flexible robot employee who can handle almost anything remains well ahead of the evidence.

We can now point to several things with confidence.

Humanoids can move factory components for long shifts. They can transfer totes through warehouse workflows. They can manipulate objects whose positions vary within a limited range. Multiple robots can work inside a live electronics-production line. Large industrial customers are signing commercial contracts. Chinese companies are manufacturing and selling robots in numbers that would have sounded unrealistic a few years ago.

The strongest jobs today are parts loading, tote movement, sorting, transfer, inspection and other forms of repetitive material handling.

That concentration tells us where the technology really stands. Humanoids are starting to occupy the space between traditional fixed automation and human labor: they are more flexible than a machine bolted to one workstation, yet nowhere near as adaptable as a person.

The scale gap remains enormous. As seen above, conventional industrial robots already number about 4.66 million worldwide. Humanoid robots doing independently verifiable productive work still represent a tiny fraction of that.

What has changed is the quality of the evidence. We have moved from asking whether a humanoid can perform a useful factory task once to measuring tens of thousands of repetitions, real operating hours, paid deployments and actual robot revenue.

China is pushing hardware volume fastest. Agility currently has one of the clearest commercial deployment records. Figure has produced one of the best-documented automotive case studies. UBTECH’s latest financial results show that full-size humanoids can already become a meaningful revenue line.

The next breakthrough will come from reliability and reuse. If one robot can move from task A to task B with days of setup instead of months, run two shifts with very few interventions and repeat that performance across dozens of customer sites, the economics change quickly.

We are not there yet.

But we are now far enough along that dismissing humanoid robotics as a collection of demos misses what is happening inside real factories. The first useful humanoid workers have arrived. They are doing narrow, repetitive, boring jobs — and that is exactly why we should take them seriously.

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

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.