Humanoid Manufacturers · Embodied Intelligence in 2026

4.0 How to read this chapter

Every company here builds bodies. They differ on four axes, and the axes matter more than the specs.

Do they build the brain too? Figure, Tesla, 1X and XPeng say yes. Apptronik, Boston Dynamics and Unitree effectively say no, or not alone.
Where does data come from? Factory deployment logs (Figure, UBTech, Agility), customer homes with human teleoperators (1X), simulation (Galbot, Boston Dynamics), retargeted human motion (Figure's S0, Tesla), or nowhere in particular because they sell hardware (Unitree).
Is the humanoid form load-bearing? Figure, Tesla and XPeng argue yes on first principles. Agility, Galbot, Dexterity and Walden argue wheels are better engineering and legs are a tax.
What is actually deployed? This is the axis that separates the chapter into two halves, and almost nobody's marketing tells you which half they are in.

4.1 Figure AI

Founded 2022 by Brett Adcock (Vettery, Archer Aviation), seeded with $100M of his own money. Figure closed over $1B at a $39B post-money valuation on 16 September 2025, led by Parkway Venture Capital with Brookfield, NVIDIA, Macquarie, Intel Capital, LG Technology Ventures, Salesforce, T-Mobile Ventures and Qualcomm Ventures.

Thesis: vertical integration wins. One company owns the hardware (Figure 03), the manufacturing (BotQ), and a single end-to-end neural policy (Helix), because the data advantage only compounds if you control the body. Helix is explicitly hierarchical: S2 for semantic reasoning at 7–9 Hz, S1 for visuomotor control at 200 Hz, and S0 for whole-body balance at 1 kHz — the last trained on 1,000+ hours of retargeted human motion and replacing, per Figure, 109,504 lines of hand-engineered C++.

Data: the BMW Spartanburg deployment — eleven months, 1,250+ hours of runtime, 90,000+ parts, contributing to 30,000+ vehicles at 84-second cycle time and 5 mm tolerance. Plus Project Go-Big with Brookfield to harvest home video across a 100,000-unit residential portfolio, plus egocentric human video and teleoperation.

Embodiment: maximally pro-humanoid and pro-own-hardware. Figure's master plan argues the choice is between "millions of different types of robots serving unique tasks or one humanoid robot with a general interface."

Skeptic: $39B on no disclosed product revenue. The BMW deployment is one repetitive part-loading station — 1,250 hours is roughly seven months of one human's working time. Home tidying demos remain curated single clips.

4.2 1X Technologies

Norwegian-American, led by co-founder Bernt Øivind Børnich; backed since 2023 by the OpenAI Startup Fund, Tiger Global, EQT Ventures and Samsung NEXT. Reported in September 2025 to be raising up to $1B at $10B+ — not confirmed closed.

Thesis: the home is the only environment with enough long-tail messiness to force general intelligence, and the only way to get home data is to put robots in homes now. NEO is $20,000 outright or $499/month.

Data: explicitly teleoperation-first, and unusually honest about it. Customers summon remote "experts" to teach NEO tasks, feeding the Redwood model, which trains on teleoperated and autonomous episodes including failures. 1X also runs a world model programme framed primarily as an evaluation substrate — its stated "holy grail" question is "can you predict how well a robot performs before you test it in the real world?"

Embodiment: the distinctive bet is soft embodiment — tendon drives, a knit covering, 66 lb mass, 25-DOF tendon-driven hands. Safety by compliance rather than by control. Børnich also frames the humanoid as the data-compatible body: making NEO as close to human as possible is what lets it learn from internet-scale video.

Skeptic: shipping a teleoperated robot into homes ships a privacy problem at scale — 1X's own FAQ has to address whether an expert can "connect and enter my NEO at any time." The unit economics of paying operators against $499/month invert only if autonomy actually arrives.

4.3 Tesla Optimus

Internal Tesla programme, no external funding.

Thesis: manufacturing is the bottleneck, not intelligence. Tesla believes FSD's vision-only, video-pretrained stack transfers to a biped, and that only a company with automotive-scale supply chain can reach a sub-$20,000 bill of materials.

Data: end-to-end video imitation from human demonstrations, in-house teleoperation capture via motion-capture suits and VR rigs, and factory task logs.

Embodiment: maximally committed to both the humanoid form and full vertical integration, down to actuators and lead screws.

Reality check: on the Q4 2025 earnings call (28 January 2026) Musk acknowledged that no Optimus robots were doing useful work in Tesla factories — against a prior target of ~10,000 units with 1,000 in productive work. In April 2026 he said Fremont production would start "late July or August," initially "quite slow," describing Optimus as the most difficult product to scale in Tesla's history because "virtually none of the components have a mature supply chain." Reporting through 2026 indicates the Model S/X line at Fremont was dismantled to install Optimus production, with 20,000–30,000 units targeted for the year, all internal.

Skeptic: three consecutive years of missed timelines, early demos containing teleoperated segments that were not initially labelled as such, and zero external units. Optimus is the demo-ware pole of this chapter.

4.4 Agility Robotics

Spun out of Oregon State's Dynamic Robotics Lab in 2015 by Jonathan Hurst, Damion Shelton and Mikhail Jones — the deepest bipedal-locomotion lineage of any humanoid firm (ATRIAS, Cassie). CEO Peggy Johnson (ex-Microsoft, Magic Leap). In June 2026 it announced a SPAC merger with Churchill Capital Corp XI at a $2.5B pre-money equity value, ~$620M gross proceeds, ticker "AGLT."

Thesis: narrow, boring, measurable. Win logistics tote-handling first with a robot engineered for uptime, not for viral video.

Data: real deployment logs above all — filings cite 65,000+ hours of real-world operation and 100,000+ totes moved at GXO.

Embodiment: pragmatically anti-purist. Digit has ostrich-like reverse-knee legs and, for most of its life, no hands and no head. Form follows the warehouse task.

Deployed: the strongest real-deployment record in the West — GXO (the industry's first humanoid RaaS contract), Schaeffler, Toyota Motor Manufacturing Canada (three units scaling to ten), Mercado Libre.

Skeptic: filings show opex rising from $71M (2024) to ~$111M (2025) with roughly $100M burn and undisclosed revenue. The headline ">$300M committed orders" is milestone-contingent and leans on one unnamed 1,000-robot customer. A SPAC is not usually the exit of a company with pricing power. Former Chief Product Officer Melonee Wise gave the sector's sharpest internal critique: "I don't think anyone has found an application for humanoids that would require several thousand robots per facility", and "currently AI is not robust enough to meet the requirements of the market."

4.5 Apptronik

Austin-based, spun out of UT Austin's Human Centered Robotics Lab (NASA Valkyrie lineage), co-founded 2016 by Jeff Cardenas and Luis Sentis. A $415M Series A (February 2025) plus a $520M extension in February 2026 brings the Series A to over $935M and total capital raised to nearly $1B, with B Capital, Google, Mercedes-Benz, PEAK6, AT&T Ventures, John Deere and the Qatar Investment Authority. (A $5B valuation was reported by CNBC; Apptronik itself has confirmed only that the extension priced at 3× the Series A.)

Thesis: best-in-class actuators plus somebody else's best-in-class brain. Apptronik's heritage is force-controlled actuator design; it partners with Google DeepMind and runs Gemini Robotics rather than building a frontier VLA alone. Its SharpaWave 22-DOF hand is among the most dexterous shipping today.

Data: teleoperation in dedicated robot training facilities, plus deployment logs from Mercedes-Benz, GXO and Jabil (which also manufactures Apollo).

Embodiment: humanoid-committed and hardware-first, but deliberately brain-agnostic — the exact inverse of Figure's posture.

Skeptic: nearly $1B raised on pilots, and outsourcing the policy layer means the moat is actuators — a component Korean and Chinese suppliers are commoditising quickly.

4.6 Boston Dynamics

Founded 1992 as an MIT spin-off by Marc Raibert; owned by Hyundai Motor Group since 2021. Retired hydraulic Atlas in April 2024 and revealed the all-electric successor the next day. Partnered with Toyota Research Institute since October 2024 to pair its control stack with TRI's Large Behavior Models.

Thesis: athletic whole-body control is the hard-won asset, and learned behaviour models should ride on top of a model-predictive controller rather than replace it. Boston Dynamics argues real work requires "a broadening of what we mean by physical intelligence" — shoulders, forearms and hips, not just fingertips.

Data: custom VR teleoperation with stereoscopic feed, haptics and foot tracking, feeding a 450M-parameter diffusion-transformer flow-matching policy running at 30 Hz on 48-action chunks. Critically, BD and TRI train one policy jointly across Atlas (50 DoF) and a 29-DoF manipulation test stand — one of the strongest public demonstrations that intelligence pools across embodiments. Sim-to-real is unusually easy for them: only two actuator types and a symmetric design mean the gap is "very small."

Deployed: a production Atlas debuted at CES 2026; Hyundai plans deployment at its Georgia Metaplant from 2028. Everything else is demo.

Skeptic: two decades of the best robot videos in the world and still no humanoid revenue. A 2028 first deployment means Atlas ships after Chinese rivals have shipped tens of thousands.

4.7 Sanctuary AI — the company that changed its mind

Vancouver, founded 2018 by Geordie Rose (D-Wave, Kindred) with Suzanne Gildert, Olivia Norton and Ajay Agrawal. Roughly $140M+ raised. Gildert departed in 2024; Daniel Friedmann is now CEO.

Original thesis: the most maximalist in this book — human-like general intelligence via a hydraulic, tactile-rich hand and pilot teleoperation, on the argument that dexterity, not locomotion, is the bottleneck.

The pivot: on 17 June 2026 Sanctuary announced it would deploy its Physical AI on existing third-party commercial platforms rather than wait for humanoid hardware to commercialise — an explicitly hardware-agnostic stance, with humanoids reframed as a later target. It came with a hard number: 99.5%+ success on wire plugging at 2.54-second cycle times, benchmarked against a Tier 1 automotive supplier's production line.

Why this entry matters more than its size suggests: Sanctuary is the first well-funded humanoid company to publicly abandon the form factor as a near-term strategy. Read charitably, it is evidence that the intelligence layer is separable from the body. Read skeptically, it is a capitulation dressed as strategy — a company that raised on human-like general intelligence now selling contact-rich skills for industrial arms, in a crowded market, without hardware differentiation.

4.8 Unitree Robotics

Hangzhou, founded 2016 by Wang Xingxing. The outlier: profitable, and audited. Its STAR Market IPO process through 2026 produced the field's first real humanoid P&L — ~$248M 2025 revenue, 59.8% gross margin, 3,701 humanoids produced and 3,551 sold in the first nine months of 2025, raising roughly $610–620M. (Reporting differs on whether the listing itself landed in March or July 2026; the financials are consistent across sources.)

Thesis: price, not intelligence. Ship cheap, capable hardware to researchers, integrators and entertainment buyers, let the world write the software, and ride the component cost curve. Average selling price fell from $85,000 in 2023 to $25,000 in 2025 — a 71% decline in two years, driven by in-house actuators.

Data: essentially outsourced. Unitree sells the platform; its sim-to-real RL locomotion work is excellent but there is no proprietary data flywheel.

Embodiment: hardware-first and form-agnostic in practice — quadrupeds, wheeled bases and bipeds all ship.

The number that defines the sector: Unitree's demand mix is 73.6% research and education, 17.4% demonstrations and displays, 9.01% industrial. The only company with audited humanoid revenue is overwhelmingly selling to people who study robots and people who put on shows.

Skeptic: a hardware margin business dressed as an AI story. If value accrues to the policy layer, Unitree is the Foxconn of this cycle.

4.9 UBTech Robotics

Shenzhen, founded 2012 by Zhou Jian; listed on the HKEX (9880.HK) since December 2023 — the first pure-play humanoid public company, and loss-making its entire public life.

Thesis: industrial humanoids win by cost curve and supply-chain localisation, not frontier models. UBTech targets 90% domestic component sourcing and sub-$20,000 unit cost by 2027–2030, running factories in Shenzhen and Liuzhou.

Data: factory deployment logs from BYD, FAW-Volkswagen, Geely and Foxconn lines — and, distinctively, UBTech sells data infrastructure, including a ¥159M contract for a "Humanoid Robot Data Collection Center" in Zigong. Walker S2's autonomous hot-swap battery exists precisely to keep robots — and data collection — running continuously.

Deployed: the most credible order book in the sector. Walker S2 entered mass production and delivery in November 2025 with orders exceeding ¥800M, later cumulative Walker orders above ¥1.1B. Roughly 500 units delivered in 2025, with targets of 5,000 in 2026 and 10,000 in 2027.

Skeptic: a large share of orders come from state-linked or strategic buyers — a ¥264M border-security contract, a government data centre — which is subsidised demand, not product-market fit.

4.10 Galbot

Beijing, founded May 2023 out of Peking University's embodied-AI group. Raised over $300M at a $3B valuation in December 2025, bringing total funding to roughly $800M, with CATL among strategic backers.

Thesis: simulation first. Galbot's signature claim is pre-training on large-scale synthetic action datasets and post-training on real data — an explicit bet that compute buys you past the teleoperation bottleneck.

Embodiment: deliberately semi-humanoid. G1 is a wheeled base with a human-like torso and dual arms. Galbot rejects bipedalism as an unnecessary cost and reliability tax while claiming cross-embodiment transfer of navigation and manipulation.

Deployed: the strongest commercial reality of any Chinese startup here — Galbot Store autonomous retail in 30+ Chinese cities, warehouses running 24/7 for over a year, hospital pharmacy pilots, and partnerships with CATL, Bosch, Toyota and Hyundai.

Skeptic: an autonomous convenience store is a vending machine with an arm. The sim-to-real thesis is unproven outside pick-and-place, and dropping legs forfeits exactly the stairs-and-clutter environments that justify humanoids in the first place.

4.11 AgiBot / Zhiyuan Robotics

Shanghai, founded February 2023 by Deng Taihua (ex-Huawei) and Peng Zhihui ("Zhihui Jun"), a Huawei "Genius Youth" hire with a large engineering following. Backed by Tencent, JD.com, Sequoia China, Hillhouse, SAIC Capital, BYD and CATL. In July 2025 it agreed to take a controlling stake in Shanghai-listed Swancor for ~¥2.1B as a listing route. (A 2026 Hong Kong IPO process is widely reported in trade press but could not be confirmed against a primary source.)

Thesis: data is the product. AgiBot World is the largest purpose-built real-robot manipulation corpus in existence — 1,001,552 trajectories, 2,976.4 hours, collected by 100+ robots in a 4,000 m² facility, with the GO-1 policy reporting a ~30% average gain over Open-X pretraining. This is a deliberate attempt to become the ImageNet of embodied AI and set the standard others train on.

Embodiment: aggressively multi-embodiment — humanoid (Yuanzheng), wheeled (Jingling), and the dual-mode Lingxi X2-N — with one foundation model intended to span them.

Deployed: 5,100+ humanoids shipped in 2025; revenue reported growing from ¥300K in year one to over ¥1B in 2025.

Skeptic: revenue growth that steep in a market this immature usually means selling robots to integrators, other robot companies and state-backed demonstration projects. A reverse merger into a wind-turbine composites firm is a financing manoeuvre, not validation.

4.12 Fourier Intelligence

Shanghai, founded 2015 by Gu Jie — uniquely, out of rehabilitation robotics, with products in 2,000+ clinical institutions across 40+ countries. Raised a Series E of roughly ¥800M (~$110M) in early 2025; SoftBank is a prior backer.

Thesis: the humanoid's first real market is care, not the factory. Fourier calls GR-3 a "Care-bot" and targets hospitals, community centres and eldercare — B2B, explicitly not consumer.

Data: clinical and rehabilitation deployment history is the differentiator, supplemented by an open-source play — the Fourier N1 platform and a developer framework released to crowdsource skills.

Embodiment: builds its own full-size humanoids (GR-1 2023, GR-2 2024, GR-3 August 2025 — 165 cm, 55 DoF), arguing the humanoid form is necessary specifically because care is an interpersonal, physically intimate task.

Skeptic: "warm tech companionship" is the softest claim in the sector and the hardest to price. Eldercare is the most regulated, most liability-exposed, lowest-margin environment imaginable, and a three-hour battery does not survive a care shift.

4.13 Neura Robotics

Metzingen, Germany; founded 2019 by David Reger. In June 2026 it announced a Series C of up to $1.4B — the largest robotics round in European history — backed by Tether, Qualcomm, Amazon, NVIDIA, Bosch, Schaeffler, the European Investment Bank, imec.xpand and others.

Thesis: "cognitive robotics" — sensor-rich, inherently safe robots that work without cages are the precondition for humanoids, and Europe must own its physical-AI stack rather than rent it.

Data: the most distinctive strategy in this chapter. NEURA Gyms are large-scale physical training environments where fleets learn by doing, feeding the Neuraverse, an app-store-like ecosystem where skills learned by one robot propagate to all. Reger's framing is that intelligence is a network good, not a per-robot asset.

Embodiment: builds its own hardware (MAiRA cobot, MiPA service robot, 4NE-1 humanoid), but the Neuraverse is explicitly cross-embodiment — Neura wants to be the Android of physical AI.

Skeptic: 4NE-1 has been shown publicly since 2023 with no disclosed commercial deployment. "Orderbook and strategic pipeline exceeding $1B" is a deliberately soft composite metric, and Tether as anchor investor in a hardware company invites questions about capital quality.

4.14 Kepler, Astribot and the Chinese long tail

Kepler Robotics (Shanghai, 2023) builds the Forerunner K2 "Bumblebee": 178 cm, up to 52 DoF, 11–12 DoF hands with 96 fingertip tactile contact points, 15 kg single-hand load, 2.33 kWh battery. It began SAIC-GM plant testing in April 2025 and claims mass production with "thousands of preorders" at a reported $20K–$30K. Funding is undisclosed; the preorder count is company-stated and unaudited.

Astribot (Stardust Intelligence, Shenzhen), backed by Ant Group, builds the S1 — 1.7 m, 7-DoF arms — and markets on speed and manipulation dexterity rather than locomotion. Funding figures unverified.

The shared thesis: China's entrants bet on the component supply chain, not frontier models. Domestic harmonic reducers, actuators and rare-earth magnets let them undercut Western bills of materials several-fold, and cheap hardware in volume generates the data.

Skeptic: near-identical spec sheets across a dozen firms, heavy reliance on industrial policy and demonstration-project demand, and "mass production" announcements that consistently outrun independently verified deliveries.

4.15 Walden Robotics — the wheeled-humanoid turn

Cambridge, MA. Emerged from stealth 15 July 2026 with $300M at a $1.1B valuation, co-led by Toyota and Deviation Capital, with Nvidia and Boeing reported as participants. CEO: Russ Tedrake, simultaneously MIT professor and TRI's robotics head. Spun out of TRI, building wheeled-base humanoids trained by large behaviour models, and already running in a Toyota North American plant.

Why this entry is disproportionately important: Tedrake is the most credentialed legged-robotics academic in the United States, and he founded a company that does not build legs. His stated reasoning is the sharpest anti-humanoid argument in this book:

"I thought about legs for 20 years; that's the class I teach at MIT. There are many reasons to build a robot with legs. But the question is, what's the addressable market?... Factories already have autonomous mobile [wheeled] robots. They already have safety cases built around AMRs. You can piggyback on that."

The market is voting with him at the margin. 2026's largest robotics rounds include Ai² Robotics ($735M at $3B) and Holiday Robotics' FRIDAY ($105M) — both wheeled humanoids.

Skeptic: a $1.1B valuation six months from founding, pre-product, with a sitting MIT professor as CEO. Even if the engineering argument is right, the price is a bubble artifact.

4.16 The new entrants: Meta, XPeng, and the Korean component play

Meta established a robotics product group inside Reality Labs, later folded into Superintelligence Labs, after a February 2025 internal memo positioning robotics as its next AR-scale bet. It hired Marc Whitten (ex-Cruise CEO) as VP of robotics, and MIT's Sangbae Kim reportedly leads humanoid hardware. In May 2026 it acquired Assured Robot Intelligence, bringing in Lerrel Pinto (NYU) — who now leads robotics at Meta Superintelligence Labs — and Xiaolong Wang (UCSD). The stated plan is dual-track: build consumer humanoid hardware and license AI, sensors and software to other manufacturers, with reported talks involving Unitree and Figure. Its research substrate is genuinely strong: FAIR's Digit 360 and Sparsh tactile stack, the PARTNR benchmark, Habitat, V-JEPA 2 world models, and Project Aria — the largest egocentric-video collection apparatus in existence. Nothing has shipped; no launch date is disclosed. Meta has cycled through robotics strategies since 2019.

XPeng unveiled next-generation IRON in November 2025, reportedly with 82 DoF, 22-DoF 1:1-scale hands, solid-state battery, three Turing chips, and a VLA 2.0 model that skips the language token entirely — vision straight to action — shared across cars, robots and flying vehicles, with mass production targeted for end-2026. (Flag: XPeng's own robotics pages were unreachable during verification and these specifications could not be independently confirmed. Treat the numbers as company-stated.) If accurate, the cross-domain VLA is the strongest industrial claim that intelligence transfers across bodies.

Korea is selling shovels. LG launched its Axium actuator brand; Samsung Electro-Mechanics and Samsung Electronics are entering actuators and dexterous hands; Hyundai Mobis actuators ship inside production Atlas. With actuators at 40–70% of humanoid cost depending on whose estimate you use, this may be the better trade — and it quietly concedes that nobody knows which humanoid wins.