Which humanoid robot actually works today?
In our humanoid robotics market deck, you will find everything you need to understand the market
SUMMARY
Agility Robotics’ Digit is the humanoid robot that most clearly works today. It performs narrow, paid industrial work and has the broadest public record of repeated customer use.
The useful dividing line is not whether a robot can complete a chore once. It is whether a customer can leave it inside a live workflow for repeated shifts without engineers rescuing every awkward movement.
Digit leads because its evidence stacks up: more than 100,000 totes moved, over 65,000 operating hours, four named commercial customers, a multi-year service agreement, and commitments spanning nine facilities.
Figure has the strongest single factory case. Figure 02 handled more than 90,000 parts during 1,250 operating hours at BMW, but Figure has less multi-customer production history than Digit.
UBTECH and AgiBot complicate any US-only ranking. UBTECH has disclosed the largest humanoid sales and revenue figures among the main contenders, while AgiBot recently ran several robots through a real tablet-production line for more than 64 hours across six days.
Unitree has already proved that humanoid hardware can be manufactured and sold at surprising scale. Its 5,500-plus deliveries do not yet prove that those machines are producing repeatable economic output for customers.
Current autonomy is real but narrow. The best robots can run a defined sequence on their own, while people still map the site, prepare objects, integrate software, supervise fleets, and recover unusual failures.
The economics probably work in a small set of awkward, repetitive jobs that are difficult to staff and poorly suited to fixed automation. Broad cost superiority remains unproven because companies rarely disclose integration, support, intervention, maintenance, and downtime costs together.
Atlas and Apollo 2 look close, while Optimus remains mostly a high-potential internal program with little published production data. All three could move quickly once customers report hours, throughput, intervention rates, and renewals.
No humanoid can switch reliably between unrelated jobs or operate as a dependable autonomous household helper. A few robots work today, but mainly by moving familiar objects through carefully prepared industrial workflows.
This market map, featured in our humanoid robotics market deck, highlights top companies and startups in the humanoid robotics market
What does it mean for a humanoid robot to actually work today?
A humanoid robot works today when a customer can keep it on a useful job for repeated shifts without engineers rescuing every difficult movement.
One successful pick tells us almost nothing. We want hundreds or thousands of completed cycles, a live production environment, known customers, and enough operating time for hardware faults and awkward edge cases to appear.
We do not expect human-level flexibility because no humanoid currently has it. One narrow job can still count, just as an industrial arm counts even though it cannot change departments or understand a manager’s instructions.
We also separate autonomous work from a service built around remote operators. A robot may create value with occasional human support, but the support level changes what the machine itself has achieved.
| Level | What we see | What it really proves |
|---|---|---|
| Demo | A prepared task completed once | The movement is possible |
| Pilot | Temporary work at a customer site | The robot can leave the lab |
| Production deployment | Repeated output inside a live workflow | The robot can do useful work |
| Repeatable commercial use | Several customers keep or expand deployments | The product may scale |
Why do humanoid robot demos fool people so easily?
Most humanoid demos show a movement once and tell us very little about what happens over a full shift.
A short video can hide retries, remote control, carefully placed objects, low speed, off-camera resets, and hours of preparation. Even an unedited clip usually answers only one question: could the robot complete this version of the task under these conditions?
Factories expose a different set of problems. Parts arrive slightly rotated. Lighting changes. A gripper heats up. A box is damaged. People cross the workspace. The robot must recover without turning every small surprise into an engineering call.
Teleoperation adds another grey area. It is useful for teaching robots, collecting data, and resolving rare failures. Yet a machine guided through difficult steps by a person has shown less autonomy than the finished video suggests. 1X is unusually open about this with NEO: early units have basic autonomy, while a remote expert can supervise unfamiliar chores at scheduled times.
We therefore give the most credit to long runs, customer renewals, disclosed operating hours, and actual production counts. Athletic movement and tidy household videos are interesting, but they sit much lower on the ladder.
As this chart shows, and as featured in our humanoid robotics market deck, search interest in where to buy robots has been rising steadily
Which humanoid robots are doing useful work now?
Digit, Figure, UBTECH’s Walker S2, and AgiBot’s G2 currently have the clearest public record of useful work.
The gap between them comes from the proof available. Digit has the widest customer record. Figure has the strongest detailed automotive case. UBTECH leads in disclosed sales and humanoid revenue. AgiBot recently showed several robots running through a real production line for six days.
The rest of the field is less settled. Unitree has shipped thousands of humanoids, mostly as platforms rather than documented workers. Apollo 2 is collecting data at customer and partner sites. Atlas has committed deployments but still lacks published factory output. Tesla describes Optimus as progressing toward mass production. Those companies may move up quickly, but today their public record is thinner.
| Robot | Best proof available now | What still holds it back |
|---|---|---|
| Agility Digit | More than 100,000 totes, over 65,000 operating hours, four named commercial customers | Work remains concentrated in narrow material-handling tasks |
| Figure 02 and 03 | 90,000 parts and 1,250 operating hours at BMW, plus new deployments beginning | Less multi-customer production history than Digit |
| UBTECH Walker S2 | The largest disclosed humanoid sales and revenue among the main contenders | Little public detail on uptime and output at each customer |
| AgiBot G2 | A six-day factory run spanning more than 64 hours and four workflows | One public factory case and a wheeled base |
If you want more recent data on this point, please see our latest humanoid robotics market report.
Is Digit the most proven humanoid worker today?
Digit leads today because no rival matches its combination of paid use, named customers, operating time, and repeated output.
GXO moved Digit from a pilot into a multi-year Robots-as-a-Service agreement. At the Spanx warehouse, the robot takes totes from mobile robots and places them onto conveyors. The task sounds simple because it is simple. That is why the case is useful: GXO put the robot into a regular logistics flow and kept it there.
Agility later reported more than 100,000 totes moved in commercial deployment. Its latest public information also lists commercial deployments with GXO, Schaeffler, Toyota Motor Manufacturing Canada, and Mercado Libre, alongside more than 65,000 hours of operation across commitments covering nine facilities.
The work remains narrow. Digit moves known containers between known points in mapped sites. It does not inspect damaged goods, improvise a repair, or switch departments after a verbal request. A humanoid earns the label “working” by doing a useful job reliably, and Digit has crossed that line more convincingly than any other biped.
Agility’s advantage also comes from the unglamorous parts of deployment: fleet software, facility mapping, workflow integration, safety evaluation, support, and troubleshooting. Competitors often show better hands or broader AI. Digit has more proof that the whole system survives contact with a customer.
This chart, featured in our humanoid robotics market deck, illustrates yearly venture capital funding for humanoid robotics startups
Did Figure prove more than Digit at BMW?
Figure proved the hardest single humanoid factory case so far, but Digit still has the broader commercial record.
At BMW’s Spartanburg plant, Figure 02 loaded sheet-metal parts into production fixtures. Figure reported more than 90,000 parts handled, over 1,250 operating hours, roughly 200 miles walked, and work linked to 30,000 BMW X3 vehicles.
Those totals imply about 72 handled parts per reported operating hour, or roughly one every 50 seconds. We should not read that as an exact cycle time because Figure has not explained every waiting period or how the hours were counted. Even so, the scale is far beyond a staged demonstration.
The deployment also revealed a recurring forearm problem, which Figure called its top hardware failure point at BMW. That admission is useful. Robots only develop boring, repeated failure patterns after they have spent real time doing the job.
Figure 03 is now taking over. Figure says it has built more than 350 units and demonstrated a production rate of one robot per hour at BotQ. The new model has returned to BMW for a logistics workflow involving parts and a wheeled cart, while Catalyst Brands has signed a commercial agreement for a Nevada distribution center. We have not yet seen Figure 03 match its predecessor’s published output, so the company remains a very close second rather than the current leader.
If you want more recent data on this point, please see our latest humanoid robotics market report.
Is UBTECH quietly ahead of the American humanoid companies?
UBTECH is currently ahead in disclosed humanoid sales, although its field performance is much harder to audit.
In its latest annual results, UBTECH reported 1,079 full-size humanoid robots sold and RMB820.6 million in related products and services. One year earlier, that revenue was only RMB35.6 million. Humanoids grew from 2.7% to 41.1% of company revenue and became UBTECH’s largest business.
Dividing revenue by unit sales gives roughly RMB760,000 per robot. That is only an order-of-magnitude estimate because the revenue line includes services and different products. It still shows that UBTECH has built a substantial commercial operation rather than a collection of free trials.
Walker S2 is aimed at material handling, sorting, inspection, and factory logistics. UBTECH says the model entered mass production and large-scale scenario deployment during 2025. The filing gives far less detail about uptime, interventions, cycle times, and output at individual customers.
The comparison is awkward. Digit tells us far more about work completed at customer sites, while UBTECH tells us far more about sales. We rank Walker below Digit today, although UBTECH gives the strongest reason to distrust any US-only view of the market.
| UBTECH measure | Earlier year | Latest year | Change |
|---|---|---|---|
| Full-size humanoid revenue | RMB35.6 million | RMB820.6 million | About 23 times higher |
| Share of company revenue | 2.7% | 41.1% | Became the largest category |
| Full-size humanoid sales | Tiny implied base | 1,079 units | Reached meaningful delivery scale |
| Company loss | RMB1.16 billion | RMB789.8 million | Improved, but remained large |
This chart, featured in our humanoid robotics market deck, shows how Agility Robotics is capturing share in humanoid robotics
Did AgiBot prove a humanoid can handle a real factory shift?
AgiBot has produced one of the clearest public factory tests, with several robots working for more than 64 hours across a six-day livestream.
At Longcheer Technology’s tablet factory, AgiBot robots carried out more than 64,828 production-line tasks across at least four workflows. The company reported a 99.99% task-success rate and 17,625 tablets of cumulative line output.
The strongest part of the test was its duration. A six-day broadcast gives failures more chances to appear and makes selective editing harder. The robots worked around moving materials, factory equipment, human operators, and the normal rhythm of an active line.
The figures still need careful reading. The 17,625 tablets were total line output, so we cannot treat them as output produced by one robot. AgiBot also did not publish the amount of human intervention, downtime by robot, or cost per completed task.
Even with those gaps, G2 belongs in the working group. Its wheeled base makes the comparison with Digit or Figure less pure, yet wheels may be the smarter choice on a flat factory floor. AgiBot also said its 15,000th robot across its wider product portfolio had left production shortly before the test, giving it manufacturing depth that most startups still lack.
Do Unitree’s 5,500 humanoid deliveries prove its robots work?
Unitree has proved that humanoids can be manufactured, delivered, and sold cheaply at scale. Productive labor is a separate question.
Unitree officially clarified that it delivered more than 5,500 humanoid robots to end customers during 2025 and produced more than 6,500. The total covers humanoids only, excluding its wheeled dual-arm machines and quadrupeds.
That is an extraordinary manufacturing result. The G1 starts around $13,500, while newer small models go lower, bringing humanoid hardware within reach of universities, laboratories, developers, and smaller businesses. A larger installed base also creates more experiments, software, training data, and third-party modifications.
What we rarely see is sustained customer output. Unitree publishes impressive locomotion and manipulation, but it has not released a Digit-style tote total, a Figure-style automotive case, or an AgiBot-style production run. Many buyers are purchasing a programmable platform whose purpose is to be developed.
Unitree may be the most successful humanoid hardware seller today. The G1 still needs customer proof showing that a meaningful share of those thousands of machines produce repeatable economic output.
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 funding for humanoid robotics startups
Is Tesla Optimus doing useful factory work today?
Optimus is still an internal development program rather than a publicly proven humanoid worker.
Tesla has shown Optimus walking, sorting, handling objects, and performing household-style tasks. The company also has obvious advantages in manufacturing, batteries, motors, computer vision, and its ability to test robots inside its own factories.
Its latest investor updates remain future-facing. Tesla describes Optimus as making progress ahead of mass production, while earlier guidance said the first production line was being prepared for a start before the end of 2026.
The missing facts are basic: how many robots perform regular production work, how many hours they operate, what parts they handle, how often people intervene, and whether any Tesla line now depends on their output. Tesla has published none of these at a level comparable with Agility or Figure.
A few ordinary factory metrics could change our view quickly. Until Tesla publishes them, Optimus stays outside the proven group.
Which humanoid robots are close to working but still lack enough proof?
Atlas and Apollo 2 are serious working systems in preparation, although neither has published enough customer output to join the leaders.
Boston Dynamics has begun manufacturing the product version of Atlas, and all planned 2026 deployments are committed. Fleets are scheduled for Hyundai’s Robotics Metaplant Application Center and Google DeepMind. Atlas can swap its own batteries and comes from a company with a long record of shipping commercial robots, which gives the program unusual credibility.
The missing piece is what happens after installation. A committed deployment shows that a customer wants the robot; productive hours and throughput show whether it earns a place in the factory. Atlas will probably move up once Hyundai publishes those results.
Apptronik’s Apollo 2 is further along as a data-collection fleet. The company says bipedal and wheeled robots are active across Robot Park locations and customer sites, including work connected with Mercedes-Benz, GXO, and Google DeepMind. Apptronik also says those runs combine teleoperation with autonomous execution.
That mix keeps Apollo below the leaders for now. We still need customer throughput, autonomous completion rates, operating hours, and proof that the robots remain in production after the data campaign ends.
This chart, featured in our humanoid robotics market deck, compares the main business model options for humanoid robot manufacturers
How autonomous are the humanoid robots working today?
Today’s best humanoids can run a defined workflow on their own, but none can manage a whole job the way a person can.
Digit can navigate a mapped facility, coordinate with other machines, pick up containers, and deliver them without someone steering every step. Figure says Helix handles perception and full-body movement on board. AgiBot’s factory run also involved several robots coordinating across linked workflows.
Humans still build the world around that autonomy. They map the site, define safe zones, integrate software, choose compatible objects, train the task, watch the fleet, and recover unusual failures. The robot owns the repeated sequence; people own most of the surrounding judgment.
The gap becomes obvious when the task changes. A warehouse employee can notice a torn package, ask a supervisor what to do, find a different tool, and resume work. Current humanoids usually need a known recovery routine or outside help.
That gap between running a sequence and managing a job explains why real deployments remain narrow.
Can humanoid robots safely work beside people now?
Humanoids can work safely in controlled shared facilities today. Free movement through crowded human workspaces is a tougher problem.
A tall biped can fall, pinch, strike, or lose balance after an emergency stop. Learned control adds uncertainty because every response is harder to describe in advance than a fixed industrial sequence.
Companies manage the risk through restricted zones, lower speeds near people, emergency stops, monitored access, safety controllers, and site-specific assessments. Digit has passed field evaluations for commercial locations, which helped Agility move beyond a laboratory setting.
The standards are still catching up. ISO 25785-1, which covers safety requirements for dynamically stable industrial mobile robots, remains under development. Existing machinery and industrial robot rules still apply, while the unfinished standard shows that legged robots raise questions the older framework did not fully address.
This chart, featured in our humanoid robotics market deck, shows the revenue mix across customer segments in the humanoid robotics market
Are humanoid robots economical today?
Humanoids probably save money in a few awkward, repetitive jobs today, but nobody has proved a broad cost advantage across factories and warehouses.
GXO’s decision to turn its Digit pilot into a multi-year service agreement is the clearest commercial clue. Toyota also signed a commercial agreement after testing Digit. Large operators do not extend every robotics experiment, so those decisions deserve weight.
Early customers may still be paying for more than labor. They gain experience, influence the product, prepare their facilities, and learn where humanoids fit before competitors do. A deployment can make strategic sense even while its direct labor economics remain mediocre.
The full cost includes the robot, software, integration, support, spare machines, charging, maintenance, remote intervention, site changes, and downtime. Companies rarely publish that complete figure. UBTECH’s fast-growing humanoid revenue has also arrived while the company still loses substantial money, showing that building and supporting these robots remains expensive.
The best economics currently appear in jobs that are repetitive, difficult to staff, physically unpleasant, spread across human-designed spaces, and awkward for a fixed robot arm. In a stable high-speed process, conventional automation will usually win.
If you want more recent data on this point, please see our latest humanoid robotics market report.
Can any humanoid switch jobs like a human worker?
No current humanoid can move between unrelated jobs with human reliability.
Digit moves containers. Figure loaded sheet-metal parts and is now testing logistics work. Walker S2 focuses on handling, sorting, and inspection. AgiBot’s robots covered several connected steps on one tablet line. Each case required preparation, integration, and a limited range of objects.
AI models are broadening what robots can learn. Figure has shown one system loading a dishwasher, folding laundry, tidying rooms, and manipulating packages. AgiBot has released a large dataset spanning many tasks. Those demonstrations point toward faster retraining.
Reliability across unusual situations remains the wall. A tangled item, reflective surface, shifted fixture, damaged box, hidden object, or vague instruction can expose how narrow the learned skill still is.
A truly flexible worker would learn a new task quickly, remember it, perform it thousands of times, know when it is confused, ask for the right help, and repeat the result at another site. We have not seen that full chain from any robot.
This chart, featured in our humanoid robotics market deck, shows how factory humanoid robot technology has evolved over time
Does any humanoid robot work in the home today?
No humanoid currently works as a dependable autonomous household helper, although 1X NEO is the closest product people can actually order.
1X offers NEO for $20,000 or through a $499 monthly subscription, with US deliveries scheduled to begin in 2026. The company has also started full-scale production at its California factory.
The autonomy remains limited. 1X says early NEO units arrive with basic autonomy, learn routines over time, and can receive remote expert supervision for difficult chores. A customer may get the task completed, but part of the service can still come from a person outside the home.
Domestic work is brutal for robots. Every house has different furniture, clutter, lighting, fragile objects, pets, children, fabrics, cables, and storage habits. Factories remove variation wherever possible; homes generate new variation every day.
NEO may become useful for early adopters who accept supervision and gradual learning. Today, it belongs closer to an assisted robotics service than a household appliance that can be trusted to finish chores alone.
Which humanoid robot actually works today?
Agility Robotics’ Digit is currently the clearest answer: it performs narrow, paid industrial work with the broadest public record of repeated customer use.
Digit wins through accumulation rather than one spectacular capability. It has a multi-year customer agreement, deployments with four named enterprises, work spread across nine committed facilities, tens of thousands of operating hours, and a six-figure tote count.
Figure comes closest. Its BMW deployment is the strongest detailed factory case, and Figure 03 is now entering new industrial sites. UBTECH may already lead the world in commercial volume, but we cannot inspect its robots’ day-to-day performance as closely. AgiBot has added a strong recent factory run and could climb quickly if it repeats that result with more customers.
Unitree’s thousands of deliveries show that humanoid hardware can scale without proving that most units earn money through labor. Atlas and Apollo 2 look credible but still need output from customer operations. Optimus remains a high-potential program with remarkably little published production data.
The answer is narrower than the videos suggest. A few humanoids work today, mainly by moving objects through carefully prepared industrial workflows. Digit has the strongest claim to first place. No robot can currently walk into an ordinary workplace or home and take over whatever needs doing.
If you want more recent data on this point, please see our latest humanoid robotics market report.
In our humanoid robotics market deck, we identify pain points entrepreneurs should prioritize
OUR METHODOLOGY
We treated a humanoid as “working” only when the public evidence showed repeated useful output in a live customer or production workflow. A prepared demonstration counted as proof that a movement was possible; it did not count as proof of dependable work.
We looked at deployment stage, named customers, operating hours, task or production counts, customer renewals, commercial agreements, autonomy, intervention levels, safety, and disclosed economics. Long runs and repeat deployments carried more weight than short videos, athletic movement, manufacturing targets, or future commitments.
We did not force every company into one metric. Operating hours and throughput are the clearest evidence for Digit and Figure, disclosed sales and revenue are more informative for UBTECH, and the six-day factory run is the strongest public evidence for AgiBot. Unit deliveries were treated as evidence of manufacturing scale, not automatically as evidence of productive labor.
Autonomous work and teleoperated service were kept separate. Occasional human help does not disqualify a deployment, but it changes what the robot itself has proved. We therefore gave more credit when the machine could complete the repeated workflow without someone steering each difficult step.
We included wheeled humanoids when they performed the same kinds of industrial jobs as bipeds, while noting the difference where it affected the comparison. The ranking is about useful humanoid work, not a contest to reward legs for their own sake.
Simple calculations derived from company figures, including Figure’s approximate parts handled per reported operating hour and UBTECH’s rough revenue per unit sold, were used only to understand scale. They are analytical context, not independently reported operating metrics.
Key sources include Agility Robotics and its customer deployment pages, Figure’s BMW and production updates, BMW Group announcements, UBTECH investor materials and Hong Kong Stock Exchange filings, AgiBot, Longcheer Technology, Unitree Robotics, Tesla investor updates, Boston Dynamics, Apptronik, 1X Technologies, GXO Logistics, Toyota Motor Manufacturing Canada, Mercedes-Benz Group, Google DeepMind, and ISO for the developing safety standard.
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 has the most advanced humanoid robot?
- Which company has the lead in humanoid robots?
- Who is building the best humanoids for factories?
- Which startup is closest to making home humanoids a reality?
