2026 was supposed to be the year humanoid robots proved themselves commercially. The volume is there. According to IDC, about 18,000 humanoid robots were shipped worldwide in 2025, up roughly 508% year on year [1]. In the first half of 2026 alone, shipments reached close to 25,000 units, and Chinese vendors built more than 95% of them [2].

Volume, however, is not the same as value. Many products look alike, real and recurring demand is still unproven, and a wave of companies is racing to the stock market. Between a convincing prototype and a profitable product at scale lies what engineers call the "valley of death": the stage where risk and cash burn are highest and most products fail. This article looks at what China's humanoid industry has to solve to get across.

Where the industry stands

China leads on volume and supply chain. Morgan Stanley attributes 63% of the global humanoid robot supply chain to China [4]. The cost effect is large: Morgan Stanley estimates the bill of materials for Tesla's Optimus Gen 2 at about USD 46,000 with Chinese suppliers, compared with about USD 131,000 without them [5]. China's lead is especially clear in the dexterous hand, one of the most expensive components at up to around a third of the bill of materials [6]. Prices are falling accordingly: Unitree's average selling price for humanoid robots dropped from about CNY 593,000 in 2023 to about CNY 168,000 in the first nine months of 2025 [12].

The US leads on the "brain". In Morgan Stanley's "Humanoid 100" value-chain map, AI and computing are dominated by US companies such as Nvidia, Alphabet and Microsoft [4]. US robot makers, however, have a thinner supply chain and limited production capacity.

Europe sets the rules and has the demanding applications. Europe installed about 85,000 industrial robots in 2024, with Germany alone accounting for almost 27,000 [7]. Labour shortages make European industry an important potential user of humanoid robots, and EU rules are a benchmark that manufacturers worldwide must meet. European developers are scaling up too: Germany's Neura Robotics raised up to USD 1.4 billion in June 2026 [16].

Real deployments exist, but they are still the exception. Some Chinese makers already have robots working on production lines; we map ten of them, and how far each has got, in a separate article. Yet IDC's data shows how early the market is. In the first half of 2026, research and education, performances and displays, and government data-collection centres together still took 69% of all humanoid shipments, down from 83.8% in 2025 [3]. Less than a third went into manufacturing, logistics and other commercial uses.

Four hurdles on the way across

1. Can demonstrate, can't yet work, doesn't yet pay

At trade shows, humanoid robots work as cashiers or baristas. It looks futuristic. But a machine that takes money and hands out goods already exists: the vending machine. A coffee machine makes coffee. Both are faster, more reliable and cheaper. In many applications, mature technology still beats the humanoid on cost and failure rate.

Warehousing is the most frequently cited first use case:

• AgiBot and JD Logistics presented the Genie G2 Max at WAIC in Shanghai in July 2026. It is a heavy-payload, force-controlled robot for the inbound handling and palletising that people still do by hand in otherwise automated warehouses. Neither company has published uptime figures yet [13].

• Agility Robotics in the US is further ahead. According to its SEC filings, its Digit robot has been deployed or committed at nine customer sites and has logged more than 65,000 hours of real-world operation [14]. Customers include the logistics group GXO and the German motion-technology company Schaeffler. Schaeffler has also taken a stake in Agility and plans to use humanoids across its global plant network [15].

The industry broadly agrees that walking and basic movement are largely solved. Grasping and manipulation, especially tactile sensing, are not, and they remain the main barrier to deployment at scale. Commercial users need measurable economic value, not showmanship.

2. Safety: the overlooked priority

Cyber security is far behind. In April 2026 the Chinese security research firm DARKNAVY and two partner organisations published a white paper on embodied-AI security. Once the researchers had a commercially available robot from a well-known brand in hand, it took them less than eight hours to go from finding vulnerabilities to complete remote control. By comparison, the authors note, a full remote compromise of a flagship smartphone usually takes a professional team at least several months, and a connected car even longer [9]. Attacks can also spread from robot to robot. At the GEEKCON security contest in Shanghai in 2025, researchers took control of one robot, then used its vulnerable Bluetooth service as a relay to pass malicious code to nearby robots, including one that was not connected to any network, and made it perform dangerous movements [9, 23].

AI creates new kinds of risk. When a large model controls a robot, a wrong output becomes a wrong action in the physical world:

• From prompt to takeover. DARKNAVY describes a case in which carefully crafted prompts got around the safety limits of a service robot's cloud-based language model. The model then issued a task containing malicious code, which exploited a flaw in the robot's on-board agent and gave the attackers full control of the device [9].

• Jailbreaks. The RoboJailBench benchmark defines 18 categories of safety violations for embodied AI, from excessive force to unauthorised image capture [8]. Carefully crafted commands can trick a robot into dangerous behaviour.

• Manipulated perception. Adversarial patches in a scene can make a vision model misread signs or miss defects.

• Poisoned training data. Hidden triggers in demonstration data can make a model behave normally until a specific signal appears.

• Wide attack surface. The white paper describes seven typical attack scenarios: unauthorised access through cloud services; hijacked accounts or leaked tokens; attacks from the local network; Bluetooth and other short-range wireless links; physical access through maintenance and debug ports; manipulation of the environment to fool the robot's perception; and poisoned firmware, model or software updates that spread to every device at once [9].

Functional safety needs new answers for humanoids. The established principles of ISO 13849 and IEC 61508 still apply, but humanoids add problems that conventional robots do not have:

• Falling. A bipedal robot relies on dynamic balance. The IEEE Humanoid Study Group points out that an emergency stop that cuts power may make it fall and create a bigger risk [10]. Safety concepts therefore need controlled stopping and fall-energy absorption.

• Contact. Two arms, two hands and two legs mean far more potential contact points than a robot arm, which makes force limitation and collision detection much harder.

• Trust. People readily treat human-shaped machines as human, overestimating what they can do and underestimating the risk.

• AI-made decisions. Safety-relevant decisions may come from a model that cannot be fully explained or verified. A consensus is forming that the deterministic safety loop must be strictly separated from AI control.

Some makers are starting to design for this. Shanghai-based DroidUp, which unveiled its biomimetic companion robot Moya in January 2026 [24], described a three-layer approach in an interview with Taas Humanoid Robotics:

• a compliant structure that, according to the company, absorbs energy on contact instead of passing it on as a hard impact;

• an independent intent check between the AI model and the motors, so that no model output goes straight to the drives;

• behaviour that can be interrupted, downgraded or explained, with a defined safe state for every fault.

"Our aim is not a model that never makes mistakes," says Dr Yang Yichao of DroidUp, "but a robot in which a model's mistake never turns into a physical action." [25]

Dedicated standards are still being written. The draft ISO 25785-1, for example, points towards controlled descent instead of collapse when power is lost [11].

3. A capital bubble and look-alike products: a shakeout is coming

When Unitree listed on Shanghai's STAR Market in August 2026, its shares opened more than seven times above the issue price [22]. Within a month they had fallen by more than half from their peak [18]. In September, Reuters reported that China's securities regulator was using informal "window guidance" to slow humanoid robot IPOs. Applicants are expected to show recurring revenue and a path to narrower losses, and regulators are questioning whether revenue from local-government-backed projects reflects real demand [18].

The underlying problem is sameness. China's National Development and Reform Commission counted more than 150 humanoid robot companies in late 2025. Many were start-ups or newcomers from other industries, and the commission warned against highly repetitive products squeezing out research and development [17]. Many of these companies assemble bought-in components with open-source algorithms. Not all of them will survive the coming consolidation.

4. From samples to products

A large share of the supply chain does not mean mass production has been mastered:

• Consistency at volume. Output cannot be raised simply by adding workers, and the industry still lacks common processes and test standards that guarantee consistent quality. Joint modules are still made in small batches across many variants, without the high-volume supply base the car industry has.

• Integration. A humanoid combines motion control, computing, AI, sensors and mechanics. Few companies master all of them, and unlike a robot arm repeating one programmed task, a humanoid must work reliably in changing environments.

China is pushing hard on all of this. Manufacturers are transferring know-how from electric vehicles and consumer electronics. Since June 2026, a government programme has been pushing companies to prove their robots in more than 100 real application scenarios [19]. In February 2026, China also published its first national standards framework for humanoid robots and embodied AI [20].

Crossing the valley

The valley of death is not about whether the technology works. It is about whether the business case closes. Crossing it does not require general artificial intelligence. It requires commercial common sense:

• find applications with real demand;

• make reliability industrial-grade;

• bring costs down to a level customers can justify;

• build service and support that last.

The companies that prove their unit economics in specific applications, with real orders instead of showcase deployments, will make it across. Those that rely on assembling hardware, chasing concepts and collecting subsidies will not. On the other side of the valley, robots stop performing and start working.

What this means for European buyers

A note from Humanoid Europe

1. Check the supplier, not just the robot. Consolidation is coming, so before you buy, ask whether the manufacturer will still exist in three years, and who provides spare parts, updates and support in Europe.

2. Clarify conformity early. From 20 January 2027, the EU Machinery Regulation applies. It explicitly covers machinery with self-evolving behaviour [21]. A CE mark on the box is not the same as a complete conformity file. Ask for the declaration of conformity, test reports and technical documentation.

3. Know where the data goes. A humanoid robot collects images, sound and location data on your shop floor. If that data is processed in a cloud outside the EU, GDPR rules on international transfers apply [21]. For connected products, the Cyber Resilience Act adds cyber-security obligations [21].

4. Start with one use case. Prove value in a single, well-defined application with clear acceptance criteria before scaling.

Sources

Where available, we link to primary sources. Sources marked (in Chinese) are original Chinese publications.

Market data and research

1. IDC, worldwide humanoid robot market analysis, 2025 data. Reported by CGTN, 24 January 2026. https://news.cgtn.com/news/2026-01-24/IDC-report-China-leads-the-global-humanoid-robot-rise-in-2025-1KccOGZyVGM/p.html

2. IDC, worldwide humanoid robot tracker, H1 2026. Reported by TechNode, 29 September 2026, https://technode.com/2026/09/29/china-accounts-for-77-9-of-global-humanoid-robot-shipments-in-h1-idc-says/ and Forbes, 30 September 2026, https://www.forbes.com/sites/johnkoetsier/2026/09/30/humanoid-robot-shipments-jump-432-china-has-95-market-share/

3. Humanoids Daily, "IDC Puts H1 Humanoid Shipments Near 25,000, With AGIBOT Leading" (IDC application breakdown). https://www.humanoidsdaily.com/news/idc-h1-2026-humanoid-shipments-agibot

4. Morgan Stanley, The Humanoid 100: Mapping the Humanoid Robot Value Chain, February 2025. Summarised by China Industry Watch. https://www.ciw.news/p/humanoid-robots-100

5. Morgan Stanley estimate of the Optimus Gen 2 bill of materials. Reported by Interesting Engineering. https://interestingengineering.com/ai-robotics/chinese-parts-cloud-tesla-humanoid-robot-growth

6. Yole Group, Humanoid Robots 2026: Market and Technology Trends. https://www.yolegroup.com/product/report/humanoid-robots-2026/

7. International Federation of Robotics, World Robotics 2025, press release, 25 September 2025. https://ifr.org/ifr-press-releases/news/global-robot-demand-in-factories-doubles-over-10-years

8. RoboJailBench: Benchmarking Adversarial Attacks and Defenses in Embodied Robotic Agents, arXiv:2605.19328, May 2026. https://arxiv.org/abs/2605.19328

9. DARKNAVY, with CIIPA and Cybersecurity Reviews (数说安全), 具身智能安全技术白皮书：机器人篇 [Embodied AI Security Technical White Paper: Robots], April 2026 (in Chinese). https://www.darknavy.org/whitepaper/Embodied-AI-Security-Humanoid-Robots-2604.pdf

10. IEEE Humanoid Study Group on emergency stops for dynamically stable robots. Reported by MIT Technology Review, "Why humanoid robots need their own safety rules", 11 June 2025. https://www.technologyreview.com/2025/06/11/1118519/humanoids-safety-rules/

11. i-SCOOP, "ISO 25785-1 explained and what it means for humanoid robot safety". https://www.i-scoop.eu/iso-25785-1-explained-and-what-it-means-for-humanoid-robot-safety/

Company information and filings

12. Unitree IPO prospectus. Reported by The Robot Report. https://www.therobotreport.com/unitree-ipo-shows-a-real-hardware-business-the-humanoid-case-is-still-early/

13. AgiBot, "AGIBOT Unveils Four New Products at WAIC 2026", July 2026, https://www.agibot.com/article/231/detail/85.html, and The Robot Report, https://www.therobotreport.com/agibot-unveils-four-embodied-ai-products-real-world-operations-waic/

14. GeekWire, "'Digit' maker Agility Robotics to go public in $2.5B deal", 2026, https://www.geekwire.com/2026/digit-maker-agility-robotics-to-go-public-in-2-5b-deal-heres-what-the-filings-say-about-its-finances/ and Agility Robotics, "Agility Unveils Digit 5", https://www.agilityrobotics.com/content/agility-unveils-digit-5-humanoid-robot-built-for-cooperatively-safe-work-at-scale

15. The Robot Report, "Schaeffler plans global use of Agility Robotics' Digit humanoid". https://www.therobotreport.com/schaeffler-plans-global-use-agility-robotics-digit-humanoid/

16. Germany Trade & Invest, "Robotics company sets new German VC financing record" (Neura Robotics), June 2026. https://www.gtai.de/en/invest/industries/robotics-company-sets-new-german-vc-financing-record-2004850

Policy and regulation

17. Reuters, "China state planner asks humanoid robot firms to avoid repetitive products", 27 November 2025, via Yahoo Finance, https://finance.yahoo.com/news/china-state-planner-asks-humanoid-074820513.html. See also Bloomberg, 27 November 2025, https://www.bloomberg.com/news/articles/2025-11-27/china-warns-of-bubble-risk-in-booming-humanoid-robotics-industry

18. Reuters, "China slows humanoid robot IPO rush as hype outruns reality", September 2026, via Investing.com. https://www.investing.com/news/stock-market-news/china-slows-humanoid-robot-ipo-rush-as-hype-outruns-reality-4908236

19. Pandaily, "MIIT and SASAC Launch 2026 Humanoid Robot Real-Scene Training Initiative", June 2026, https://pandaily.com/miit-sasac-humanoid-robot-real-scene-training-2026-jun2026, and Window to China, 16 June 2026, https://window-to-china.de/2026/06/16/chinese-government-launches-initiative-for-over-hundred-humanoid-robot-application-scenarios/

20. SESEC (Seconded European Standardization Expert in China), "China's First Standards System for Humanoid Robots and Embodied Intelligence", 1 April 2026, https://sesec.eu/2026/04/01/chinas-first-standards-system-for-humanoid-robots-and-embodied-intelligence/, and TechNode, 28 February 2026, https://technode.com/2026/02/28/china-releases-first-national-standard-framework-for-humanoid-robots-and-embodied-ai/

21. EU legislation (EUR-Lex):

– Machinery Regulation (EU) 2023/1230: https://eur-lex.europa.eu/eli/reg/2023/1230/oj

– GDPR (EU) 2016/679: https://eur-lex.europa.eu/eli/reg/2016/679/oj

– Cyber Resilience Act (EU) 2024/2847: https://eur-lex.europa.eu/eli/reg/2024/2847/oj

Media reports and interviews

22. NBC News, "Unitree, China robot maker, surges in stock market debut amid AI boom", August 2026. https://www.nbcnews.com/world/asia/unitree-china-robot-maker-stock-market-ai-humanoids-tech-trump-rcna593278

23. South China Morning Post, "Chinese researchers show how 1 word could allow spies to take control of a robot army". https://www.scmp.com/news/china/science/article/3337300/chinese-researchers-show-how-one-word-could-allow-spies-take-control-robot-army

24. iTechGuides, "New Realistic Humanoid Robot Moya Unveiled by DroidUp: Specs, Price and Availability", 2026. https://www.itechguides.com/new-realistic-humanoid-robot-moya-unveiled-by-droidup/

25. Interview with Dr Yang Yichao, DroidUp, conducted by Taas Humanoid Robotics, 2026 (unpublished).