Unitree filings put humanoid work claims in context

The 2026 World Robot Conference puts China's humanoid-work claims beside a more practical question: which machines can prove useful in constrained workflows? The Beijing E-Town event opened on August 19 with more than 300 companies and more than 2,000 exhibits, according to geekpark. Vbot used the conference to launch its Vbot ATOM humanoid embodied-intelligence terminal and open reservations. The scale signals intense commercial ambition, not proof that general-purpose humanoid coworkers are ready.

The useful comparison is already visible on the floor. geekpark reported that Future Far Robot demonstrated its F2 robot preparing pasta autonomously from start to finish, while Variable Robotics showed logistics work with two robotic arms and ordinary grippers rather than a full humanoid and five-finger hands. This article separates public spectacle from deployment evidence: task fit, service history, data feedback, and the economics of keeping a robot working.

The conference floor shows ambition, not deployment speed

At the 2026 World Robot Conference in Beijing E-Town, which opened on August 19, more than 300 companies showed more than 2,000 exhibits. The floor made embodied AI look broad and immediate: Vbot launched its Vbot ATOM humanoid embodied-intelligence terminal and opened reservations, while Future Far Robot showed its F2 robot preparing pasta autonomously from start to finish with its WAM world action model.

Those demonstrations describe ambition. They do not by themselves establish deployment speed.

Future Far Robot said it had entered more than 500 paying households and accumulated more than 50,000 service hours, a stronger signal than a stage demonstration because it points to repeated use in homes. Yet a claimed household footprint still leaves open the hard question: how reliably can a robot handle the variation of real environments over time?

The logistics exhibit offered a more grounded contrast. Variable Robotics used two robotic arms with ordinary grippers, rather than a full humanoid and five-finger hands. In a previously public livestream, its system ran continuously for one hour, sorting 1,816 items per hour with accuracy above 98%. That is a task-defined measurement, tied to throughput and accuracy.

According to a report, Huang Jinping argues that investors often overestimate short-term deployability because real-world robots need far more data accumulation and model capability. He also warns that excessive capital can push firms toward short-term results instead of better technology and products. The conference therefore shows many possible forms for robots; it does not erase the work needed to make them dependable.

Unitree's revenue still points to research buyers

Unitree's G1 can make humanoid capability highly visible: geekpark described combination punches, spinning kicks and flying kicks. Those movements demonstrate balance and control, but they do not establish that buyers are paying for general-purpose work. Unitree's own revenue mix points elsewhere. In the first three quarters of 2025, scientific research and education accounted for 73.6% of its humanoid-robot revenue, while industry applications supplied 9.01%, according to woshipm.

That is a research-led customer base, not evidence of broad workplace substitution.

The narrower operational categories are smaller still. Woshipm reports that intelligent manufacturing, intelligent inspection and logistics delivery generated 15.702 million yuan for Unitree humanoid robots in the first three quarters of 2025, equal to 2.64% of total humanoid-robot revenue. The distinction matters because a robot's human shape can attract attention while its paid use remains concentrated among buyers who need a research platform, demonstration system, or training asset.

Task-specific logistics systems offer a different standard of proof. Variable Robotics showed a system with two robotic arms and ordinary grippers, rather than a full humanoid body and five-finger hands. In a previously public livestream, geekpark reports, it ran continuously for one hour while sorting 1,816 items per hour at accuracy above 98%. Stardust Intelligence's Xingdong M7 has deployments in logistics centers including China Post and can process up to 1,200 items per hour, according to geekpark. These figures do not settle every deployment question, but they make the relevant comparison clear: buyers can ask what task is completed, at what rate, and with what accuracy.

A public-market signal can hide a crowded field

Unitree Robotics' listing turned embodied AI into a public-market confidence signal. According to geekpark, its intraday market capitalization once exceeded 400 billion yuan after it rang the bell on the opening day of the 2026 World Robot Conference. Yet another report cites an IPO valuation of 61 billion yuan, showing how quickly a listed symbol can become a broader wager on the category rather than a narrow reading of current business results.

The demand signal was intense. Unitree's STAR Market subscription had a winning rate of only 0.02% on August 10, and its offering implied a static price-to-earnings ratio of about 219 times, versus 38.56 times for general equipment manufacturing.

That premium can validate investor belief in embodied AI, but it does not establish that every company has a durable product or deployment path. The field is already crowded: DeepRobotics entered exchange acceptance or review, while Agibot said in July 2026 that it had begun a Hong Kong listing process. A listing pipeline can make the sector appear to be moving as one market, even when companies differ sharply in revenue, technical maturity, and customer fit.

Huang Jinping warns that excessive capital can produce disorderly competition, encouraging short-term results over better technology and products. He says many companies valued above 10 billion yuan may generate revenue only in the tens of millions of yuan. His market example is blunt: if 20 companies split a 500 billion yuan opportunity, only three may do especially well; firms ranked fourth through tenth may do reasonably well, while others may not survive. As viable models reach scaled revenue, he expects capital to concentrate further on leaders, making financing harder for the middle and lower tiers.

Deployment signals across embodied-AI approaches

Full-stack hardware is an advantage with tradeoffs

Unitree Robotics makes motors, reducers and encoders in-house, according to woshipm. That vertical reach matters because actuator choices shape how a robot moves, how it is repaired, and which components can be changed without waiting on an outside supplier. It does not make the machine generally useful by itself.

The company chose electric motors rather than hydraulics, a report says, and built a quasi-direct-drive system around low-speed motors and planetary reducers. That differs from the pairing of Swiss Maxon high-speed motors with Harmonic Drive reducers. Huang Jinping said that Unitree's electric-drive approach can reduce hardware costs while serving quadruped and humanoid robots. The potential advantage is therefore a shared hardware base, not a guarantee that either form has solved a buyer's task.

Unitree also supplied self-developed open underlying interfaces that let university researchers modify algorithms, according to the report. Open interfaces can make hardware a more useful research platform. They do not supply the task data, training process, or evaluation evidence needed for dependable deployment.

Its filing position and planned spending make that distinction sharper. As of January 31, 2026, Unitree held 262 patent rights, including 20 domestic invention patents, woshipm says. It plans to direct 20.22 billion yuan, or 48.13% of total fundraising, toward intelligent robot model research and development. Patents can protect engineering choices, while model spending signals that hardware self-reliance has not closed the software question. For practitioners, the relevant test remains whether the combined machine and model can learn from a bounded workflow, be maintained there, and produce measurable operating results.

Choose embodied-AI systems by workflow evidence, not humanoid spectacle

• You have a repeatable logistics task such as sorting, and the work does not inherently require legs or a humanoid body. Start with task-specific automation. Variable Robotics demonstrated a system with two robotic arms and ordinary grippers that ran continuously for one hour, sorted 1,816 items per hour, and achieved accuracy above 98%. Treat a full humanoid body and five-finger hands as requirements to prove, not default specifications.
• A vendor demo is visually impressive, but you need evidence that the system can sustain production work. Ask for task-level throughput, accuracy, continuous-run results, and named deployment evidence. Stardust Intelligence says its Xingdong M7 is deployed in logistics centers including China Post and can process up to 1,200 items per hour; Variable Robotics published a one-hour sorting run. Dancing, combat movements, and transforming mechs are not substitutes for those measures.
• You are considering a humanoid robot for industrial use today. Treat the purchase as an early deployment rather than evidence that general-purpose humanoid labor is ready. In the first three quarters of 2025, 73.6% of Unitree's humanoid-robot revenue came from scientific research and education, while 9.01% came from industry applications. Intelligent manufacturing, intelligent inspection, and logistics delivery together represented 2.64% of its humanoid-robot revenue.
• You need a robot platform that can improve after installation instead of remaining a fixed demonstration. Prioritize a disclosed data and update loop, and verify what it means for your workflow. Vbot released models covering interaction understanding, world-model decision-making, cross-embodiment evolution, and real-world feedback loops; it also claims that Big Head completed four OTA updates in the previous three months. The sources do not disclose the operational impact, reliability, or customer controls of those updates.
• Lifecycle cost, repairability, and supply continuity matter as much as model capability. Examine how much of the hardware stack the supplier controls and request service terms separately. Unitree develops and produces motors, reducers, and encoders in-house, which is relevant to supply-chain control, but service coverage, spare-parts availability, repair turnaround, and maintenance pricing are not disclosed in sources.

Buy a constrained task, then demand deployment evidence

Buy the task before buying the body. Variable Robotics demonstrated logistics work with two robotic arms and ordinary grippers, avoiding the cost and complexity implied by a full humanoid and five-finger hands. Its previously public livestream ran continuously for one hour, sorting 1,816 items per hour with accuracy above 98%, according to geekpark. That is the kind of baseline a buyer can inspect: throughput, error rate, and sustained operation under a defined workload.

A procurement review should demand the same evidence for the proposed site: uptime records, task-success definitions, integration APIs, teleoperation procedures, and ownership of the resulting data. Ask how maintenance is scheduled, what happens during downtime, which safety requirements apply, and who provides on-site support. Unitree's full-stack self-developed open underlying interfaces let university researchers modify algorithms, a report says. Openness can help integration, but a buyer still needs proof that interfaces, service commitments, and recovery procedures work in a paid deployment.

Claims of use should be separated from claims of readiness. Future Far Robot said it had entered more than 500 paying households and accumulated more than 50,000 service hours; Vbot said Big Head completed four OTA updates in the previous three months. Those are useful questions to pursue: how often does the system change, and who bears deployment risk after each update?

For industrial work, request site-specific evidence rather than a viral demonstration. Stardust Intelligence said its Xingdong M7 operates in logistics centers including China Post and can process up to 1,200 items per hour. Compare that claim with Unitree's revenue mix: woshipm reports that industry applications supplied 9.01% of humanoid-robot revenue in the first three quarters of 2025. The purchase case should rest on a constrained task, measurable economics, and a supportable operating model.

Treat a vendor's valuation as a signal to investigate, not proof that a robot is ready for your site. According to reporting, Huang Jinping warns that real environments demand far more accumulated data and model capability than short-term deployment narratives suggest.

Ask for evidence from paid work: task success rates, uptime, maintenance terms, integration APIs, and teleoperation or data policies. Then test the machine against a repeatable physical bottleneck in a controlled warehouse, lab, facility, retail setting, or service environment.

Watch whether a supplier can turn trials into scaled revenue and a viable business model. Huang expects the gap between leaders and weaker companies to become clearer within the next two to three years, while financing becomes harder outside the leaders. That makes serviceability and sustained support practical procurement questions, rather than details to defer until after a demo.

For readers outside China

• Availability: Availability outside China is not disclosed in sources. The material documents conference launches, reservations, deployments, and listings in China, but does not state which robots can be bought, supported, or operated internationally.
• Pricing: The sourced consumer-oriented price point is Unitree's Go2 Air quadruped, which had a starting price below 10,000 yuan in 2025. Historical Unitree product pricing cited in the sources ranged from $20,000 to $30,000, later rising to $45,000. No purchase price is disclosed for Vbot ATOM, Big Head, Galbot ET1, Variable Robotics' logistics system, or Stardust Intelligence's Xingdong M7.
• Closest Western equivalents: Boston Dynamics Spot is the closest named Western comparison for Unitree's quadruped robots; the sources state that Spot's cheapest version cost $75,000.; Boston Dynamics Atlas is the closest named Western comparison for electric-drive humanoid development; Boston Dynamics retired the hydraulic Atlas in April 2024 and adopted an electric-drive approach.
• Data residency: Data residency, cross-border data transfer, cloud hosting, model-operation location, and customer data controls are not disclosed in sources. This is particularly important where vendors describe real-world feedback loops, OTA updates, or large-model interactions; buyers should obtain contractual answers before deploying systems that capture workplace or household data.

Sources

• geekpark 逛完 WRC，我们发现机器人行业最该回答的 5 个问题 https://geekpark.net/news/369388
• 36kr 对话宇树早期投资人：未来具身赛道可能只有三家过得好，十家过得去，剩下的活不下去｜硬氪专访 https://36kr.com/p/3942012454419592
• 36kr 千亿IPO：山顶上的宇树 ｜ 深氪 https://36kr.com/p/3933488511270017
• woshipm 200亿身家，王兴兴告别王兴兴 https://woshipm.com/it/6443501.html