Who has the most advanced humanoid robot?

Why can't we simply rank the best humanoid robots?

We cannot produce a clean humanoid robot league table today because the leading companies still choose their own tests.

NIST has pointed out that the DARPA Robotics Challenge, held more than a decade ago, was the last serious event where several humanoids were measured side by side. Its proposed new benchmark would test walking, manipulation, coordinated movement and basic reasoning, but that common test is still being developed. With no common benchmark yet, we weigh who is ahead, and who is falling behind, in our humanoid robotics market report.

For now, every company shows the part of the problem it handles best. Boston Dynamics publishes heavy lifting, extreme reach and unusual body positions. Figure shows long household sequences, ladder climbing and learned factory work. Agility counts totes moved in a live warehouse. Unitree highlights running, sparring, recovery and price. Tesla talks mostly about future production.

The videos also hide different amounts of help. A task can be fully autonomous, teleoperated, replayed from recorded motion, guided by a human at key moments or assembled from several learned skills. Those distinctions often disappear once a short clip spreads online.

Figure's September 2026 home test is a step toward better evidence because it stated its rules: one fixed model, 30 unseen homes, no partial credit and any safety intervention counted as a failure. It is still a company-run test. We therefore give the most weight to repeated work, stated autonomy conditions, useful task variety and hardware that can survive outside a lab.

What does "most advanced humanoid robot" actually mean?

For this question, the most advanced humanoid is the robot that best combines autonomy, body control, useful hands, reliability and repeatable manufacturing.

A humanoid needs enough strength and balance to move through spaces built for people. It also needs to understand changing scenes, handle objects without constant retraining and recover when a grasp or step goes slightly wrong. Hands deserve special attention because most valuable human work ends with touching, holding, turning, pulling or fitting something.

Reliability separates a research success from useful equipment. Ten polished minutes tell us less than hundreds of ordinary hours with jams, shifted objects, warm motors and worn components. Production counts too. A company learns much faster when it can build hundreds of comparable robots instead of nursing a few custom machines.

Some things count more than others here. The biggest test is whether the robot can think and move its whole body through a changing task. Hands and reliability come next. Factory scale only helps once the machine is useful.

What we measure	Strong evidence	Why it changes the ranking
Autonomous intelligence	Long tasks, unseen places, changing objects, few human interventions	Shows whether the robot can adapt rather than replay motions
Whole-body control	Walking, balancing and manipulating at the same time	Reveals whether the humanoid shape adds real capability
Hands and manipulation	Two-handed work, touch, force control and object variety	Determines how much useful work the robot can actually perform
Reliability	Reported hours, cycles, failures and customer-site operation	Exposes weaknesses that short demonstrations hide
Manufacturing	Hundreds or thousands of comparable units, measured yields and fleet support	Shows whether progress can spread beyond a small lab fleet

Which humanoid robots are genuinely in the race today?

Figure, Boston Dynamics and Agility form the top tier, with Tesla, Unitree, UBTECH, 1X and Apptronik chasing from different directions.

Figure looks strongest overall. Helix links vision, reasoning, walking, balance and manipulation, and its Helix 2.5 version worked in 30 unseen homes. Figure says BotQ had delivered more than 350 Figure 03 units by April 2026, and the model is back at BMW's Spartanburg plant on a harder sequencing job.

Boston Dynamics has the most capable body. The production Atlas combines a 50-kilogram peak lifting capacity, a 2.3-meter reach, continuous rotational joints and industrial environmental protection. It is now training on parts logistics inside Hyundai's Georgia plant, but its public work record is still younger than Figure's or Digit's.

Agility's Digit has performed the most clearly measured commercial work. It passed 100,000 tote movements at a GXO facility, and Agility reports more than 65,000 hours of operation across its fleet. In September 2026 it unveiled Digit 5, designed to work near people without fences, with early access planned for 2027.

The chasing group contains serious threats. Unitree has built far more humanoids than anyone in the West and is now a listed company. UBTECH is commissioning new humanoid plants. 1X is preparing its first consumer deliveries. Apptronik remains a core Google DeepMind hardware partner. Tesla has the clearest route to enormous manufacturing scale if Optimus reaches the required capability.

Robot	Best evidence	Main gap	Our position
Figure 03	Broad learned autonomy, 30-home test, BMW experience and more than 350 units delivered by April (company-stated)	Most new work remains company-reported	Overall leader
Boston Dynamics Atlas	Best hardware specifications and a training site inside Hyundai's Metaplant	No published production throughput	Closest challenger
Agility Digit	Sustained commercial throughput at GXO; Digit 5 announced	Narrower manipulation and very small revenue	Worksite leader
Tesla Optimus	AI talent and a Fremont line in early production since the third quarter of 2026	V3 not shown publicly; no official unit count; outside customers planned for late 2027	High-upside follower
Unitree H2/G1	About 18,000 bipedal humanoids built, open sales and strong value	Weak public evidence of sustained useful work	Top-two volume maker and hardware ecosystem leader
UBTECH Walker S2	Industrial access, new factory capacity and factory software	Few comparable runtime totals	Leading Chinese industrial contender
1X NEO	Home design, new hands and production underway	Basic autonomy still needs remote help	Home-product leader
Apptronik Apollo 2	Large data operation and Google DeepMind partnership	Commercial results remain lightly quantified	Strongest sleeper

Who has built the best humanoid robot body?

Boston Dynamics has built the best humanoid body we can assess: Atlas is stronger, tougher and more mechanically free than its peers.

The published specifications are unusually convincing. Atlas stands 1.9 meters tall, weighs 90 kilograms, sustains a 30-kilogram load and can briefly handle 50 kilograms. Its 56 degrees of freedom and 2.3-meter reach let it work near the floor, above shoulder height and around awkward containers. Continuous rotational joints allow the robot to turn limbs through positions that would force a more human-like machine to reset its stance.

Atlas is also designed for rougher workplaces. Boston Dynamics lists an IP67 rating and an operating range from minus 20 to 40 degrees Celsius. Its four-hour battery sounds ordinary until the robot walks to a station and replaces the pack by itself. UBTECH's Walker S2 offers a similar three-minute battery swap, but Atlas combines continuous operation with greater strength, reach and joint freedom.

Others are closing specific gaps. Agility's Digit 5, announced in September 2026, lifts up to 50 pounds repeatedly and recharges a 90-minute battery in nine minutes, which Agility says allows more than 20 hours of work a day. Unitree remains impressive on dynamic movement. Its H2 has 31 degrees of freedom and joint torque reaching 360 newton-meters, and in September 2026 it launched a stronger G1+ at $15,000.

Those are good machines for specific jobs or for research. Atlas still looks more ready for heavy, varied industrial work. For a physically demanding job where hardware limits matter more than broad reasoning, Atlas is the obvious first choice.

Which humanoid robot has the smartest AI today?

Figure 03 has the best public record for the smartest full humanoid, and in September 2026 it published the strongest generalization test in the field.

Figure says Helix sends onboard vision, touch and body-position data through one learned system that controls the robot's full set of actuators. That lets the robot adjust its feet, torso, arms and hands as one action unfolds. Earlier demonstrations showed a four-minute dishwasher task and two robots making a bed together without a central coordinator.

The September 2026 Helix 2.5 release goes further. Figure tested one fixed model in 30 Bay Area homes where it had collected no data, on three jobs: tidying a living room, folding towels and making a bed. Figure reports 56% full-task success, against 9% for a model trained without its new Index dataset. Any safety intervention counted as a failure. That gap came from training data; we look at which technologies are gaining traction in our report on humanoid robots.

That number cuts both ways. Succeeding in a stranger's home more than half the time with no local training is a real step. Failing almost as often shows how far household autonomy still is from a product. Figure says plainly that general humanoid robotics is not solved.

Figure is also spending to stay ahead. In August 2026 it launched Index, a paid platform that it says has gathered more than 16 million human videos from 108 countries, and in September it signed a deal with Nscale for up to 100,000 NVIDIA GPUs, a ceiling rather than delivered capacity.

Google DeepMind is the main rival in robot intelligence. Its Gemini Robotics models now cover whole-body control and run on Apptronik's Apollo, and Atlas is set to use them too. NVIDIA's GR00T platform gives many manufacturers better foundation models. On the evidence available, Figure has shown the most convincing link between a general learning system and a capable full body.

Who has the best humanoid robot hands?

Figure leads demonstrated humanoid dexterity, although 1X has the most interesting hand hardware on paper.

Figure's advantage comes from what the whole robot has handled. Its systems have worked with boxes, thin sheet metal, dishes, towels, clothing, bedding, toys, carts and ladders. At BMW, Figure 02 placed sheet-metal parts within a five-millimeter tolerance and targeted more than 99% success for a full three-part cycle during a shift.

The household work adds two-handed and flexible-object control. Figure 03 has folded towels and handled comforters in homes it had never seen, operated a trash-can pedal while balancing on one leg and coordinated with another robot on a bed. Those jobs combine hand control with vision, foot placement and body posture.

1X has the most interesting hand design. NEO's tendon-driven hands, shown in July 2026, have 25 degrees of freedom, backdrivable joints and tactile skin intended to sense pressure and slip. Figure's CEO dismissed tendon-driven hands in August as a dead end, so the two companies are now betting on opposite designs. Unitree also unveiled a 22-degree-of-freedom Dex5-S hand at $6,500 in September, with little task data so far. Tesla, for its part, said on its October 2026 earnings call that a hand able to survive wear while staying soft enough for delicate objects is one of its hardest problems.

The proof is still catching up with the specifications. 1X has shown the hardware and described its sensing; Figure has published a broader record of autonomous manipulation. Figure keeps the lead for now, while 1X has the clearest path to taking it.

Is Agility's Digit still the most proven humanoid worker?

Yes. Digit remains the most proven paid humanoid worker, while Figure 02 completed the more delicate factory job. But Digit's record also shows how small the paid market still is.

Agility reported that Digit moved more than 100,000 totes at GXO's Flowery Branch facility, with about 98% accuracy while on task, and more than 65,000 hours of operation across its fleet. The robot works inside live deployments at GXO, Schaeffler and Toyota's Canadian plant, moving containers between mobile robots, conveyors and floor positions.

The money tells a harder story. Agility agreed in June 2026 to go public through a merger with Churchill Capital Corp XI at a $2.5 billion valuation. Its merger filings, declared effective by the SEC in October, showed $1.78 million in 2025 sales, a $138.1 million net loss and a going-concern warning. First-half 2026 sales were about $174,000. Shareholders vote on the deal in November 2026. Its $300 million in multi-year Digit 5 orders comes from one contract for 1,000 robots, tied to milestones. That is a backlog, not revenue.

Figure's BMW record covered fewer handled items but demanded more precision. Figure 02 accumulated over 1,250 hours, loaded more than 90,000 sheet-metal parts and contributed to the production of more than 30,000 BMW X3 vehicles, meeting an 84-second cycle. Figure retired that fleet in September 2026.

The others remain far behind on published totals. UBTECH works in several vehicle factories, Atlas is training at Hyundai's Metaplant and Apptronik runs fleets at partner sites. None has released a comparable mix of hours, cycles and completed units. For a warehouse buyer choosing a robot now, Digit is the safest bet on what we can verify.

Has Figure really built 1,000 Figure 03 robots?

Not on the record. Figure's last official count is more than 350 Figure 03 units delivered by late April 2026. That still makes Figure 03 a real production robot, even though much of the fleet feeds training rather than customer revenue.

In the same April update, Figure said BotQ had gone from one robot a day to about one an hour, a demonstrated rate rather than a cumulative total. In July 2026, industry outlets reported that Figure had built its 1,000th unit, but Figure has not published that milestone in its newsroom, and no production audit exists.

The earlier production detail gives the official count weight. Figure reported more than 9,000 actuators built, over 500 battery packs, a 99.3% first-pass yield on its battery line and more than 80 functional tests per robot. A fleet that size lets Figure split robots between AI training, reliability testing, household work and customer projects.

Figure also runs stress tests in public. In October 2026, its CEO said one Figure 03 had covered 348 miles on a treadmill test meant to expose failures. BMW supplied the earlier lesson: Figure 02's forearm was its most common failure point, and Figure redesigned the wrist electronics for Figure 03.

Figure 03 is now at BMW Spartanburg on a sequencing job that combines thin-part handling with pulling a heavy cart, and Figure has a commercial agreement with Catalyst Brands. The missing piece is customer throughput from the new generation. Figure's stated goal of 100,000 robots in four years is a target, not a forecast we can check.

Can Boston Dynamics Atlas take the lead soon?

Atlas can take the overall lead, but not yet. Its work record is still too thin to award it first place, and the main deployment milestones now sit in 2027 and 2028.

Boston Dynamics began manufacturing the production Atlas in Boston in early 2026, with 2026 units committed to Hyundai and Google DeepMind. In September 2026 it opened the Robotics Metaplant Application Center inside Hyundai's plant near Savannah, Georgia. Atlas is learning to unpack parts, sequence them for assembly and insert them into dense storage cubbies, which requires two-handed work.

Hyundai's plans are large. It says it intends to deploy 25,000 Atlas robots across Hyundai and Kia plants over the next few years and to build a U.S. factory able to produce up to 30,000 robots a year by 2028. Hyundai also plans to expand the training center about tenfold in 2027 and to start component assembly by 2030. Capacity targets for Atlas are plans, not output, so we judge who the likely future winners are in our humanoid robots market report.

The product already includes features many rivals still describe as future goals: autonomous or teleoperated modes, Orbit fleet software and self-swapping batteries. What we still lack is a public workload. Boston Dynamics has not released any equivalent of Digit's tote count or Figure's BMW runtime, and the reports on the Georgia center give no unit count or hours.

Atlas can do more physically and has a credible route to stronger intelligence through Google DeepMind. One quantified customer deployment could change the answer quickly. For now, Figure has simply shown more.

Has Tesla Optimus caught up now that Fremont production started?

No. Tesla started early Optimus production at Fremont in the third quarter of 2026, but it still has not published a unit count, a task result or an outside customer.

Tesla has several advantages that would be hard for a startup to copy. It already designs motors, power electronics, batteries, AI chips, vision systems and high-volume factories. Its own plants offer an enormous internal testing ground, and the company can fund a long development cycle without depending on one outside customer.

The schedule has slipped repeatedly. The Optimus V3 reveal moved from early 2026 to "middle of the year," and it still had not happened when Tesla reported third-quarter results on 21 October; images of the design leaked through Tesla's own app in September. Tesla ended Model S and X production at Fremont in May to make room. Its Q3 shareholder letter lists Fremont at "early ramp" and the Texas Optimus factory under construction, and says the first production-quality bots are now coming off the line. Musk warned the ramp would be "somewhat bumpy" and said reaching 20,000 robots a week would take at least another year. The one-million-unit annual line remains a capacity goal.

Tesla itself has published no unit count, no task-success rates, no external customers and no operating totals. On the October call, Musk said the design still needs tweaks, that the robot "can't wear out after a month," and that there is "still quite a lot of work" on the AI before Optimus handles open-ended tasks. He said Tesla may need 10,000 to 20,000 robots doing self-play to train it, a proposed fleet rather than one that exists. He also floated an unscripted Optimus video around November or December, AI5 chips around April 2027 and deliveries to outside companies likely toward the end of 2027. All of these are plans. The Information reported in September 2026 that Fremont was building several hundred robots a week in August, mostly for internal testing; Tesla has not confirmed that figure.

Optimus may become a fast-moving challenger once that output turns into robots doing measured work. Today, Tesla's strongest evidence concerns the factory it is building around the robot, and its own management says the robot is not finished. It still falls short of first place.

Are Chinese companies like Unitree already ahead in humanoids?

On volume, price and capital, yes. On broad autonomy and sustained work, Figure and Agility still hold the stronger public evidence.

Unitree is one of the two volume leaders, alongside AGIBOT. It shipped more than 5,500 humanoids in 2025 and says it had built about 18,000 bipedal humanoids by July 2026. In August 2026 it raised about 6.1 billion yuan, roughly $900 million, in a Shanghai STAR Market listing, and it remains profitable; by October its market value had cooled to about $31 billion. Its G1 sells from about $13,500, and NVIDIA chose its H2 Plus as the body for an open Isaac GR00T reference humanoid.

The market-share data is messy. Counterpoint put AGIBOT first in the first half of 2026 with about 9,700 units and Unitree second with over 7,000, while Smart Analytics Global put AGIBOT at about 8,400 and Unitree at about 5,900; both rank Unitree second. Either way, Chinese firms are shipping humanoids in numbers no U.S. company has matched.

UBTECH is further along in factory rollout. It delivered 1,079 full-size humanoids in 2025 and sold 921 in the first half of 2026 alone, reported about 1.27 billion yuan in first-half 2026 revenue, and in September commissioned a Liuzhou plant designed to produce one humanoid every ten minutes. In October, FAW-Volkswagen extended its UBTECH work into factory logistics.

The tension sits in Unitree's own words. In August 2026, CEO Wang Xingxing said the company was holding back from large-scale factory deployment because its robots still work slower than people and need retraining for each task. In October, the company warned investors that commercialization is at an early stage. China leads the industrial ecosystem. The best all-around robot still comes from the United States.

Does 1X have the best humanoid robot for the home?

NEO has the clearest home humanoid product, but Figure has shown more autonomous household ability.

1X has designed around the risks and inconveniences of a normal house. NEO uses a soft exterior, covered joints, tendon-driven movement and self-charging. Customers can reserve a $20,000 ownership version or choose a $499 monthly subscription. 1X says it plans first U.S. deliveries in 2026 and other markets from 2027. We could not confirm how many units have reached customers.

The new hands make the product more serious. They combine tactile sensing with backdrivable tendons and human-like movement. That should help with fragile objects and accidental contact, which matter far more around people, pets and furniture than on a fenced factory line.

Autonomy remains the weak point. 1X says early NEO units arrive with basic autonomy, and a company expert can remotely supervise the robot on unfamiliar tasks. That may collect useful data quickly, but the customer is partly buying a human-supported learning service. Funding is also unsettled: SoftBank was reported in August 2026 to be negotiating a majority stake at about $6 billion, below the $10 billion 1X had sought. No deal has been confirmed.

Figure has demonstrated richer independent household behavior, now including towel folding, bed making and room tidying in 30 homes it had never seen. Figure has yet to package that ability into a consumer offer with NEO's safety design, price and support. NEO leads as a home product; Figure leads the household autonomy race.

Is Apptronik the sleeper in the humanoid robot race?

Apptronik is the most credible sleeper, with Apollo 2 serving mainly as a learning platform before the commercial Apollo 3 arrives.

Apollo 2 comes in both walking and wheeled versions. That choice is practical. Wheels suit flat warehouses, use less energy and fit existing industrial safety rules more easily. Legs remain available for sites where steps, tight passages or human-designed layouts justify the extra complexity.

Apptronik's Austin Robot Park covers about 90,000 square feet, and Apollo 2 fleets operate at Google DeepMind, Mercedes-Benz, GXO and other partner sites. Apollo serves as a hardware target for Google's Gemini Robotics models, which added whole-body control in July 2026. In October 2026, Google showed a system teaching Apollo new skills from video, simulation and physical feedback, and Apptronik published a technical explainer on Apollo's whole-body control, without new customer or unit figures. Partnerships and demos are easy to announce; we check who has real customers and who only promises in our humanoid robots report.

The company has raised more than $935 million in its expanded Series A, with backing from Google, Mercedes-Benz and John Deere. In September 2026, its CEO warned that the U.S. lacks enough gear and actuator supply to build humanoids at scale, a problem that touches every Western rival.

Apptronik's own description explains why it remains below the top tier. Apollo 2 is the workhorse for data collection, and Apollo 3, expected in 2027, is meant to be the first true product. Public results still lack the hours, success rates and throughput needed to rank Apollo beside Digit or Figure. The setup is impressive. The proof is still missing.

Are any humanoid robots genuinely general-purpose today?

None of today's humanoids is general-purpose. The best systems can stretch one architecture across a growing set of trained tasks.

Figure's 30-home test is the closest anyone has come, and it still covered only three behaviors with 56% success. Each new capability needs relevant data, training and a controlled rollout. Digit can switch between several warehouse workflows, yet it cannot walk into an unfamiliar plant and decide how to operate every machine. Atlas is still being trained, task by task, at Hyundai's center.

NEO makes the boundary especially clear. The product handles basic learned chores and calls on remote experts for harder ones. Apptronik gathers teleoperated data to prepare future autonomous fleets. Unitree's CEO says his robots need retraining for each factory task.

Humans bring years of touch, movement and common sense to a new room. These robots start with much less physical experience and must avoid damaging expensive hardware while learning. Their apparent intelligence can also collapse when lighting, object shape, friction or timing changes beyond the training range.

The useful term is multi-purpose. A strong humanoid can learn several unrelated jobs on one body and carry parts of that learning into the next job. That is already valuable. General-purpose autonomy, where a robot safely handles an unfamiliar request without preparation, remains unsolved.

Who leads each part of the humanoid robot race?

The category leaders split cleanly: Atlas owns hardware, Digit owns workplace proof, Unitree shares the volume lead with AGIBOT, and Figure owns the best overall blend.

Figure wins our combined ranking because it stays near the front across almost every important test. Its body is capable enough for factory and household work. Helix has the broadest learned task record and the only published unseen-home test, Figure 02 supplied real deployment data, and BotQ had delivered more than 350 Figure 03 units by April.

Atlas loses points mainly through missing customer throughput. Its hardware lead is large enough that one strong industrial deployment could move it into first place. Digit faces the opposite problem. Its work record is excellent, while its manipulation range, task evidence and revenue remain narrow.

The rest of the field leads specific bets. Tesla has the best chance to reach huge manufacturing scale. Unitree builds humanoids at a scale no Western rival matches and offers the best accessible hardware. UBTECH has the best Chinese industrial package. NEO is closest to a product built for ordinary homes. Apptronik has assembled perhaps the best ingredients for a future jump.

Meaning of "most advanced"	Current leader	Why we chose it
Best overall humanoid	Figure 03	Strongest mix of AI, manipulation, deployment learning and production
Best physical hardware	Boston Dynamics Atlas	Highest strength, reach, joint freedom and environmental protection
Most proven worker	Agility Robotics Digit	Clearest sustained commercial throughput
Best demonstrated autonomy	Figure 03 with Helix 2.5	Long full-body tasks, now tested in 30 unseen homes
Best demonstrated manipulation	Figure 03	Widest useful mix of rigid, flexible and two-handed object work
Best affordable platform	Unitree	Capable humanoids sold openly at low prices, in very large volumes
Biggest scale potential	Tesla Optimus	Manufacturing assets and a Fremont line in early production
Strongest Chinese industrial robot	UBTECH Walker S2	Factory access and new production capacity
Strongest dark horse	Apptronik Apollo	DeepMind partnership, Robot Park and strong industrial backing

Who has the most advanced humanoid robot today?

Figure has the most advanced humanoid robot overall: Figure 03 powered by Helix. Its lead widened through the summer of 2026, and Tesla's October results did not narrow it.

The answer rests on the combined package. Helix has shown full-body autonomy across kitchens, living rooms, bedrooms, logistics and automotive work, and Helix 2.5 completed three household tasks in 30 unseen homes with 56% full-task success. Figure's previous generation logged more than 1,250 hours at BMW and loaded more than 90,000 parts. The company says it had delivered more than 350 Figure 03 units by April and has returned the new model to BMW.

Boston Dynamics comes closest. Atlas has the best body by a comfortable margin and is now training inside Hyundai's Georgia plant. Hyundai's plan for 25,000 Atlas robots could make it the largest deployment in the field, but that is a plan. We need months of quantified customer work before moving Atlas to the top.

Agility deserves the clearest commercial credit. Digit's 100,000 tote movements show that a humanoid can already become part of an ordinary warehouse process. Yet its $1.78 million in 2025 sales, and about $174,000 in the first half of 2026, show how early that business is, and its narrower manipulation keeps it behind Figure in an overall technology ranking.

Tesla, Unitree, UBTECH, 1X and Apptronik could all change the answer. Unitree already builds far more units than any Western rival. Tesla has started early production at Fremont but still shows no V3, no official unit count and no outside customer, which Musk expects only toward the end of 2027. Their strongest claims concern scale, affordability, factories, home design or future AI rather than a fully rounded, proven machine.

Figure 03 leads because it combines the best public autonomy record with useful hands, whole-body control, real factory learning and the largest disclosed Western fleet. Atlas has the body to overtake it, and Digit has the work record Figure still needs to match. Today, Figure has the best all-around humanoid robot.