Signals Inbox·July 15, 2026·Humanoid Robotics

Humanoid robotics: who is ahead?

Figure and Agility Robotics are ahead in humanoid robotics today, but for opposite reasons: Figure has built the most complete general-purpose system, while Agility has the strongest proof that humanoids can already work reliably for real customers.

Q1Who is building the smartest general-purpose humanoid robot today?

Figure currently has the strongest public case for general-purpose humanoid intelligence, with 1X in second place. The rest of the field is still showing narrower autonomy or keeping too much of its performance private.

The difference comes from task breadth. Figure has shown Helix 02 controlling the robot’s hands, arms, torso, balance and walking through one end-to-end system. Since January 2026, it has published a four-minute dishwasher workflow, a living-room cleanup and a two-robot bedroom reset involving clothes, furniture, rubbish and bed-making.

Most competitors can match individual pieces of that. Several robots can pick objects from bins. Others can walk while carrying something. A few can follow spoken instructions. Figure has connected more of these abilities inside longer tasks without visibly switching between separate control systems.

1X is the nearest rival because NEO is also being built around open-ended tasks rather than one industrial routine. Its Redwood system can learn and repeat household work, and 1X has shown autonomous attempts on unfamiliar tasks. However, the company openly says that an expert may remotely supervise difficult jobs at scheduled times. That makes NEO more practical today, but it means we cannot credit every completed household task as autonomous.

Boston Dynamics sits behind those two in demonstrated general-purpose intelligence. Atlas has the stronger body and now ships with software aimed at part sequencing, machine tending and order fulfilment. Yet its partnership with Google DeepMind only began in 2026, and the most interesting generalist AI work is still ahead of the first customer fleets.

Tesla cannot currently be placed in the leading group based on public evidence. Optimus may be learning useful internal factory tasks, but Tesla has released very little that allows us to compare task length, success rate, human intervention or generalisation with Figure and 1X. Keeping performance private may make competitive sense. It does not help Optimus in a public ranking.

Figure’s lead is clear in public demonstrations, though it remains one level below proof in daily customer operations. Several hours of continuous work, failure statistics or results verified by BMW would make the case much harder to dispute.

Q2Who is ahead in humanoid robot hands and delicate object handling?

Figure and 1X are currently the top two in humanoid manipulation. Figure is ahead in complete autonomous tasks, while 1X has revealed the more impressive hand hardware.

Figure’s strongest advantage is that its hands already work as part of a broader autonomous system. Helix has handled dishes, packages, clothing, containers and unfamiliar household objects while coordinating both arms, the torso and individual fingers. That puts Figure ahead of companies whose dexterous hands have mainly appeared in isolated laboratory exercises.

1X closed much of the gap in July 2026 when it revealed the production hand that will ship with NEO. It has 25 degrees of freedom, tendon-driven actuation and motors positioned farther up the arm to keep the hand light. The company showed it inserting a USB-C connector, manipulating grapes, pouring liquid and producing detailed finger movements.

On hardware alone, 1X may now have the strongest claim. Most of its latest evidence, however, shows what the hand can physically do rather than what NEO can perform autonomously and repeatedly. Figure has disclosed less mechanical detail, but its manipulation is already embedded in longer learned workflows.

Boston Dynamics remains close on heavy and awkward industrial manipulation. Atlas can reposition its whole body around large objects and reach from angles that are difficult for more human-shaped robots. It has yet to show the same range of finger-intensive tasks as Figure or 1X.

Unitree deserves a different kind of credit. Researchers can buy its platforms, add dexterous hands and build their own manipulation systems. Recent projects have used the G1 for autonomous grasping, tactile-data collection and even teleoperated surgical assistance. Much of that intelligence comes from outside laboratories rather than Unitree’s own autonomy stack, but the platform is clearly useful.

For now, Figure turns manipulation into useful autonomous behaviour more convincingly. 1X could move ahead once NEO performs its most delicate actions independently across long, unedited tasks.

Q3Who is best at making a humanoid walk, balance and work at the same time?

Boston Dynamics currently leads the humanoid market in physical capability. Figure is the closest challenger when the question becomes whether that movement can be connected to useful autonomous work.

Atlas can rotate joints beyond normal human ranges, move through cramped positions and adjust its body around heavy industrial objects. That gives it a wider mechanical operating envelope than the more human-shaped designs from Figure, Tesla, Apptronik or UBTech.

The lead becomes obvious when a task requires the whole robot rather than just its arms. Atlas can turn its torso, reverse joint directions and reposition without taking the small corrective steps many competitors need. No newer company has publicly matched that full range yet.

Figure is now the more interesting challenger because Helix 02 controls locomotion and manipulation together. Its June 2026 BMW demonstration showed Figure 03 stepping around parts while pulling a loaded cart and adjusting its posture through the task. It looked less spectacular than an Atlas demonstration, but much closer to something a factory would pay for.

Unitree belongs in the top group for locomotion itself. The G1 can run, recover from falls and support unusually dynamic research policies. A May 2026 paper demonstrated walking, running and recovery under one controller on real G1 hardware. Its broad availability also means independent teams can test it properly.

Atlas is ahead today. The open question is whether its exceptional movement will translate into reliable industrial output once production fleets reach Hyundai and Google DeepMind.

Q4Who is actually using humanoid robots in real factories and warehouses?

Agility Robotics is currently ahead across the humanoid robotics market in repeated external workplace use. Figure is very close behind, while UBTech is emerging as the largest Chinese challenger.

Digit had moved more than 100,000 totes in a live commercial deployment by November 2025. Agility later reported commercial relationships spanning manufacturing, logistics and distribution, including GXO, Mercado Libre and Toyota Motor Manufacturing Canada.

That puts Agility ahead of companies with long customer lists but few operating results. Digit’s work is narrow, yes, but it has crossed the line from pilot theatre into repeated customer activity. A tote moved for the hundred-thousandth time tells us more about commercial readiness than a robot completing ten unrelated tasks once.

Figure is close because its BMW deployment involved a more demanding part of the production process. Figure 02 worked ten-hour weekday shifts, loaded more than 90,000 sheet-metal parts and contributed to over 30,000 BMW X3 vehicles. A missed placement there could interrupt real manufacturing.

The distinction between the two leaders is fairly clean. Agility has stronger proof of repetition and a broader group of external customers. Figure has shown deeper integration into a high-value manufacturing line and a wider route toward general-purpose work.

UBTech may already have more industrial robots physically reaching customers than either company. It began mass production and delivery of several hundred Walker S2 units in late 2025, supported by announced orders above RMB800 million. What happens after delivery is less clear. UBTech publishes much less about completed work, operating hours and intervention rates.

Apptronik has excellent access to Mercedes-Benz, GXO and Jabil, but most public evidence still describes partnerships, trials and planned deployments. Apollo does not yet have an operating benchmark comparable with Digit’s tote count or Figure’s BMW numbers.

Tesla also ranks lower than its factory footprint might suggest. Testing Optimus inside Tesla plants is strategically useful, but the company has not released enough task, fleet or uptime data for a proper comparison.

Agility is slightly ahead today. Figure could overtake it by showing that Figure 03 performs several BMW workflows repeatedly with low human intervention.

Q5Who has built the most reliable and safest humanoid robot for work?

Agility Robotics currently has the strongest reliability and safety evidence among humanoid robotics companies, and the gap is meaningful.

Digit’s lead begins with repetition. More than 100,000 tote movements show that the robot can complete the same useful action again and again under commercial conditions. Figure’s 90,000 parts provide the closest comparable benchmark, but Agility has since expanded across more facilities and customers.

Agility also completed an OSHA-recognised NRTL field safety inspection at an ecommerce fulfilment site in November 2025. That does not certify Digit for every possible workplace, but it gives Agility a piece of outside validation that most humanoid companies still lack.

Figure ranks second. Its BMW results show that Figure 02 survived long shifts and repeated a production task over more than 1,250 runtime hours. We still do not know its uptime, maintenance frequency, reset rate or how often a person intervened. Without those numbers, Figure cannot quite match Agility’s case.

UBTech’s disclosures create a mixed picture. The company is moving quickly into deliveries, but it has acknowledged that Walker S2 currently performs at roughly 30% to 50% of human efficiency on suitable tasks. That is a useful reality check: UBTech may be scaling deployments faster than it is closing the productivity gap.

1X has probably designed the safest body for close contact with people. NEO is light, compliant and covered in soft materials, reducing the consequences of a collision. Household safety goes much further than soft hardware, though. A home robot must deal with children, pets, stairs, liquids and unpredictable objects, and 1X has not published enough unsupervised home runtime yet.

Boston Dynamics brings substantial safety experience from Spot and Stretch. Its new Atlas fleet has only just entered production, so that experience cannot replace Atlas-specific operating data.

Agility leads because it has combined repetition, external customers and third-party field inspection. The next useful numbers will be boring ones: uptime, interventions per shift, maintenance hours and failure-related stoppages.

Q6Who is actually building and delivering the most humanoid robots?

UBTech and Unitree currently look strongest in physical humanoid volume, while Figure has disclosed the clearest production ramp among Western general-purpose humanoid startups.

UBTech said it began mass production and delivery of several hundred Walker S2 robots in November 2025, with a goal of delivering 500 units by year-end. That was one of the largest clearly announced batches of full-size industrial humanoids at the time.

Figure reported more than 350 completed Figure 03 robots by April 2026. It also said BotQ had moved from one robot per day to one per hour in under four months. Those figures describe robots already built and a cadence already reached, rather than a distant capacity target.

The comparison between UBTech and Figure is messy. UBTech disclosed a larger delivered batch but has said less about how production continued after late 2025. Figure’s disclosed count is slightly smaller, yet its update is newer and easier to connect to one operating factory.

Unitree may be ahead of both in cumulative humanoid shipments because it has sold G1, H1 and other platforms through direct and distributor channels for longer. The company does not publish a clean cumulative delivery figure, so a confident ranking is impossible. Its widespread availability strongly suggests real volume, but that is not the same as an audited shipment total.

1X has started full-scale production at its 58,000-square-foot Hayward factory, which employs more than 200 people and has stated capacity for 10,000 NEO units a year. Until 1X releases completed-unit or delivery figures, the factory represents a strong position for the next phase, not leadership today.

Boston Dynamics began production of Atlas in early 2026, with the entire first-year deployment allocated to Hyundai and Google DeepMind. Genuine production, small fleet.

Tesla sits earlier in its ramp. As of July 2026, the Fremont Optimus line is being prepared for production, and Tesla has not published a verified current unit count. Reports that fewer than 500 robots shipped during 2025 underline the gap between its earlier targets and actual execution.

UBTech and Unitree are the likely volume leaders, although confidence is limited by incomplete totals. Figure gives us the best view into a current Western production ramp.

Q7Who sells the cheapest humanoid robot that developers can actually use?

Unitree is far ahead of every major Western humanoid company on price and open market availability.

A basic G1 has recently been listed around $13,500, while developer-oriented G1 EDU configurations can cost several times more depending on hands, compute and degrees of freedom. Even after making that distinction, Unitree remains dramatically cheaper than industrial humanoids sold through enterprise agreements.

The cheap headline price needs context. The base robot does not necessarily include the programmability, dexterous hands or onboard computing required for advanced research. A serious development configuration can move into the $40,000 to $70,000 range. Even there, laboratories get access to capable bipedal hardware without negotiating a major customer partnership.

This is where Unitree’s lead over Figure, Agility, Apptronik and Boston Dynamics becomes especially wide. Those companies currently sell an operating solution, a pilot or a strategic relationship. Unitree sells a robot.

That difference has already shaped the research market. Independent teams use the G1 for locomotion, teleoperation, manipulation, surgery experiments and foundation-model research. Every outside project creates more software, more training data and more engineers familiar with Unitree hardware.

1X is the closest Western company in consumer accessibility. Customers can reserve NEO with a $200 deposit, but that deposit tells us little about the full cost or what early owners will actually receive. NEO will also rely on scheduled expert supervision for tasks it cannot yet perform alone.

Tesla talks about a future robot priced low enough for mass adoption, but Optimus is not for sale. A projected price at future scale cannot be compared with a product developers can order now.

Unitree’s lead is one of the clearest in this tracker. The harder question is whether these inexpensive platforms will become reliable autonomous workers or remain mainly research and demonstration machines.

Q8Who has turned humanoid robot partnerships into the strongest real customer base?

Agility Robotics currently has the strongest proven customer base in humanoid robotics. Figure has the deepest publicly measured relationship with one major manufacturer.

Agility’s advantage comes from conversion. GXO moved from testing Digit to operating it commercially, and Agility has since added agreements with Mercado Libre and Toyota Motor Manufacturing Canada. Its robots are now positioned across distribution, logistics and manufacturing rather than one showcase facility.

Figure has fewer publicly documented operating customers, but its BMW relationship has produced unusually concrete results. The first generation completed 1,250 hours of work and handled 90,000 parts. Figure 03 then returned to the same plant in June 2026 for a more dynamic logistics workflow. BMW did not stop after the original pilot and publicity cycle.

Figure also signed a May 2026 agreement with Catalyst Brands. This could expand its reach into retail and logistics, although the announcement has not yet produced deployment metrics comparable with BMW.

Apptronik probably has the strongest list of strategic names outside the top two. Mercedes-Benz, GXO and Jabil cover manufacturing, logistics and contract production, while Google DeepMind brings a major AI partner. Its funding round expanded beyond $935 million in early 2026. Still, Apollo has not accumulated a public workload comparable with Digit or Figure 02.

UBTech has announced orders worth more than RMB800 million and relationships with several industrial groups. Those orders may give it the largest commercial backlog, but backlog, deliveries and productive usage are three different stages. Without site-level results, it cannot rank ahead of Agility.

Boston Dynamics sold out its initial 2026 Atlas production to Hyundai and Google DeepMind before deliveries began. That is excellent launch demand, although one owner and one research partner do not yet make a broad customer base.

Agility leads on customer adoption today. Figure trails narrowly because its proof is deeper but concentrated. Apptronik and UBTech have enough demand to challenge both; now they need to show what happens after the contracts are announced.

Q9Who could manufacture humanoid robots at the largest scale later?

Tesla has by far the strongest long-term manufacturing potential in humanoid robotics. Figure is currently executing a humanoid-specific ramp more convincingly.

Tesla can reuse factories, purchasing power, battery systems, motors, electronics, AI computers and global supplier relationships developed through its vehicle business. No humanoid startup has operated anything close to that industrial system.

The company is converting former Model S and Model X production space in Fremont for Optimus. Elon Musk said production should begin around late July or August 2026, with the Fremont line eventually designed for one million robots annually and a later Texas line targeting much more. Those numbers describe intended capacity. Tesla still has to prove the first stage.

Figure starts from a much smaller base, but its lead in current execution is straightforward. BotQ has already produced hundreds of Figure 03 robots and reached an announced pace of one unit per hour. The initial line was designed for annual capacity of 12,000 robots.

1X has also moved beyond planning. Its Hayward factory is vertically integrated across motors, batteries, transmissions, structures, electronics and soft components, with stated annual capacity of 10,000 NEO robots. Delivery figures will show how much of that capacity it is actually using.

Boston Dynamics has perhaps the strongest manufacturing partner after Tesla. Hyundai plans to supply Atlas actuators through Hyundai Mobis, deploy robots in its own manufacturing network and build a broader robotics ecosystem in the United States. Hyundai has discussed purchasing tens of thousands of Boston Dynamics robots, although that figure covers the wider portfolio rather than Atlas alone.

Unitree and UBTech benefit from China’s dense supply chain for actuators, batteries, electronics and precision components. Their early production volume proves that this advantage is already real. Neither has demonstrated the global service, consistent quality and million-unit output Tesla could eventually bring.

Tesla is far ahead in industrial ceiling. Figure is ahead in what we can verify today. The ranking will only converge once Tesla reports completed Optimus units, sustained weekly output and robots that survive deployment after leaving the factory.

Q10Who is ahead overall in humanoid robotics today?

Figure is currently the most complete humanoid robotics company, while Agility Robotics remains ahead in proven commercial work. They are the clear top two, but neither has built an unassailable lead.

Figure comes closest to covering the whole race. It leads public general-purpose autonomy, sits in the top two for manipulation, has completed meaningful work at BMW and is already producing Figure 03 in the hundreds. Its weakness is evidence depth: we know much more about what Figure can demonstrate than about how Figure 03 performs over thousands of customer hours.

Agility has almost the opposite profile. Digit performs a narrower range of work, but the work is real, repeated and increasingly spread across outside customers. It leads the categories that reveal whether a humanoid can already function as industrial equipment: repetition, operating maturity and safety validation.

Figure is ahead by breadth. Agility is ahead by proof.

Boston Dynamics belongs just behind them. Atlas remains the strongest physical platform and has now entered production, but its first customer fleets are only arriving in 2026. Strong Hyundai operating results and progress from the Google DeepMind partnership could move it into the top group quickly.

Unitree leads a different race. It has made dynamic humanoids affordable and available enough to become a default research platform. Its lead in price and developer adoption is large, although it has not shown the autonomous commercial work required for overall leadership.

UBTech may have the strongest near-term Chinese industrial volume. Hundreds of Walker S2 units and more than RMB800 million in orders are substantial. The company currently gives us much better evidence of delivery than productivity, and its own estimate that the robots perform at 30% to 50% of human efficiency is a useful reality check.

1X is the leading home-focused challenger. Its new hands, soft hardware and vertically integrated NEO factory make it one of the most differentiated companies in the field. Remote expert supervision and limited public household-runtime data currently keep it outside the overall top group.

Apptronik has funding and partners that match almost anyone, but Apollo still needs a public operating record. Tesla has an even larger gap between potential and proof. Its manufacturing ceiling is unmatched, yet Optimus currently trails the leaders in disclosed autonomy, customer use, reliability and output.

As of July 2026, humanoid robotics has a top pair rather than one winner. Figure has the strongest claim to technical completeness. Agility has the strongest claim to commercial reality. Boston Dynamics, Unitree and UBTech lead important parts of the race but remain less complete overall.

The next ranking-changing evidence is easy to identify. Figure needs several thousand externally verified autonomous operating hours. Agility needs to show that Digit can handle a meaningfully wider range of tasks. Boston Dynamics needs results from the first Atlas fleets. UBTech needs productivity and intervention data from delivered robots. Tesla needs an actual production ramp and quantified factory work.

Humanoid robotics leaders by category, July 2026