Late January 2026, in a convention hall tucked into Shanghai's Zhangjiang Robotics Valley, DroidUp unveiled what it called the future of eldercare. The MOYA robot—standing roughly five-and-a-half feet tall, weighing about 75 pounds, equipped with synthetic skin warmed to human temperature—moved across the stage with what its creators described as "92% human-like walking accuracy." Twenty-five micro-servos animated its face. Company literature promised "emotional connection."
Journalists scribbled. Some outlets amplified the specifications verbatim. Yet in the months that followed, a quieter conversation began percolating through Chinese tech circles. By July, a Bilibili video analyst posed the question more directly: Had anyone actually verified those numbers?
That 92% figure—never explained, never independently verified—has become something of an emblem for a broader tension now rippling through China's accelerating humanoid robotics sector. The country is racing to build machines capable of caring for its aging population, and manufacturers are leaning hard into "emotional AI" as a differentiator. But the gap between what companies claim and what they can prove is widening at precisely the moment when regulatory scrutiny, vulnerable populations, and serious money are converging.
A Policy Push Meets Demographic Reality
China's Ministry of Industry and Information Technology set the terms in November 2023 with guidance targeting breakthrough humanoid capabilities by 2025 and the emergence of two or three globally influential firms. The document landed like a starter's pistol. By early 2026, product launches, factory pilots, and production announcements were cascading across Shanghai, Shenzhen, Dalian.
The forces propelling this aren't subtle. China counted 323 million people aged 60 and older at the end of 2025, according to China Daily reporting in March 2026—a figure that climbs by millions annually. National guidance issued January 7, 2025 explicitly promoted humanoid robots, brain-computer interfaces, and AI systems in eldercare. Qingdao launched a training hub and deployment pilot in April 2026. Other cities followed.
Against that backdrop, Chinese firms escalated their rhetoric. UBTECH Robotics, which raised approximately HK$1 billion (US$130 million) in a December 29, 2023 Hong Kong IPO, announced in January 2025 that its Walker S2 humanoid had entered mass production with cumulative orders topping RMB 800 million—about $110 million at mid-2025 exchange rates. The company reported factory pilots with NIO, Zeekr, BYD, and other automakers through 2024 and 2025. Unitree priced its G1 humanoid at roughly $16,000 in May 2024, opening university and lab access at a fraction of what similar platforms once cost. AgiBot claimed in a January 8, 2025 press release that it had produced its "5,000th mass-produced humanoid robot," though the claim has not been independently audited or verified, and shipment data remains conspicuously absent from public records.
MOYA entered this crowded field with a narrower pitch. Li Qingdu, DroidUp's founder and CEO—a professor and executive director at the Institute of Machine Intelligence at the University of Shanghai for Science and Technology with over two decades in biped research—emphasized what he called "scenario first" thinking in interviews through late 2025 and early 2026. Speaking to Chinese outlets including The Paper in April 2026, Li argued that humanoids must clear two gates to justify deployment: safety and return on investment. DroidUp's stated target scenarios included caregiving, education, and hospitality concierge roles where "emotional presence" might set the robot apart from purely task-oriented machines—though such capabilities have not been independently verified.
Whether that presence can be engineered is another question entirely.
Market Momentum, Scientific Uncertainty
The push toward emotionally expressive humanoids reflects both technological maturation and market calculation. Industrial robot installations globally reached 542,000 units in 2024, according to the International Federation of Robotics' World Robotics 2025 report published last September. Medical robot sales surged 91% year-over-year to approximately 16,700 units. Service robots—a category spanning entertainment, education, and personal assistance—grew 30% in 2023. Goldman Sachs Research projected a $38 billion global humanoid market by 2035 in analyses published February 27, 2024. Morgan Stanley sketched scenarios ranging from $357 billion in 2040 to somewhere between $4.7 trillion and $5 trillion by 2050, depending on adoption curves.
Emotion AI itself commands varied forecasts, though the wide range suggests methodological inconsistency more than clarity. Research and Markets estimated the sector at $4.71 billion in 2025, growing to $5.99 billion in 2026. Fortune Business Insights placed it at $3.4 billion with projections extending to 2034. SNS Insider, in a July 6, 2026 press release, offered a far more aggressive outlook: $91.76 billion by 2035. Directional momentum is clear, even if the precision isn't.
Technical enablers have advanced. Actuators, harmonic reducers, distributed sensor arrays, and large language model inference at the edge have all seen cost reductions and performance gains from 2023 through 2026. Chinese firms including AgiBot publicized proprietary embodied datasets and vision-language-action model pipelines in 2024 and 2025, signaling investment in the data infrastructure needed for adaptive robot behavior. Unitree's sub-$20,000 price point for the G1 illustrated how quickly commoditization can reshape a market.
Yet the claim that a robot can convincingly simulate human emotion—or accurately infer emotional states from facial expressions, vocal tone, or physiological signals—remains contentious in the scientific literature. A May 2025 paper in Perspectives on Psychological Science by Lisa Feldman Barrett and colleagues stressed the high variability in how individuals express emotions, challenging simple inference models that map facial configurations to discrete emotional labels. The European Union's AI Act, which entered into force June 13, 2024, cited "serious concerns" in Recital 44 regarding the scientific basis of emotion detection systems. It banned their use in workplaces and educational institutions outright, classifying other emotion recognition applications as high-risk AI subject to stringent compliance obligations.
That regulatory skepticism is worth sitting with for a moment. It reflects a growing gap between what vendors market and what researchers believe they can reliably measure.
What 92% Actually Means (If Anything)
DroidUp's positioning of MOYA centers on warm skin—maintained at 32 to 36 degrees Celsius—and a micro-expression system leveraging 25 servos in the head and face to render what the company describes as "six basic emotions" plus nuanced micro-expressions. The robot stands between 1.65 and 1.70 meters tall, weighs 32 to 34 kilograms, and integrates 3D navigation, distributed pressure sensors, and LLM-based dialogue.
The "92% human-like posture accuracy" claim first appeared in reports by outlets including IT之家 on January 31, 2026 and was echoed by NewsBytes, UOL, and others in February. The figure specifically refers to walking gait similarity, not emotional realism or overall human likeness, and has not been independently validated.
No methodology or third-party test protocol for the 92% figure has surfaced in public technical papers or patent filings accessible as of mid-2026. The July 2026 Bilibili video explainer questioned whether the metric had undergone independent validation. DroidUp's company site, updated through July 2026, highlighted a "1,000-unit strategic cooperation" signing at an expo forum on July 19, 2026, but order quantities and delivery schedules remain unverified by external auditors.
In other words: we have a number, but not much else.
EX Robots, based in Dalian, pursued a different strategy. Factory tours covered by the South China Morning Post in June 2024 showcased highly lifelike android faces with detailed micro-expressions. Pricing ranged from 1.5 to 2.0 million RMB per unit, with build times of two to four weeks. Deployments as of 2024 and 2025 concentrated on museums, experience centers, and exhibitions rather than functional eldercare or industrial roles. The emphasis on expressive realism over bipedal locomotion or task execution reflects a market segmentation: robots designed primarily for visual engagement versus those targeting operational utility.
Engineered Arts' Ameca, developed in the United Kingdom, represents a Western benchmark in expressive humanoid design. Ameca's facial actuation system supports nuanced emotion rendering and has been deployed in research studies examining human-robot interaction dynamics. A March 24, 2026 paper in Frontiers in Robotics & AI documented methods for transferring human facial expressions to Ameca's mechanical face; a December 2025 ACM conference paper explored service-interaction scenarios. Ameca's "Gen 3" iteration appeared at the International Conference on Robotics and Automation in 2025. These robots serve exhibitions, research labs, and human-robot interaction studies, but mobility remains limited compared to fully bipedal platforms.
UBTECH's Walker S and S2 models illustrated the industrial trajectory. Video documentation from NIO factories in 2024 and CnEVPost coverage of a Zeekr pilot on August 5, 2024 showed Walker units performing repetitive tasks in controlled manufacturing environments. A July 2026 partnership announcement with Hitachi and UBTECH's claim of over RMB 2 billion in revenue for 2025—including RMB 820 million attributed to S-class humanoids—suggested traction in factory automation. Yet these deployments emphasized precision and endurance over emotional expression or warm synthetic skin.
Hume AI, a U.S.-based venture, offered a software-centric counterpoint. The company announced a $50 million Series B round on March 25, 2024 alongside the launch of its Empathic Voice Interface, or EVI, designed to modulate conversational agents' tone and pacing based on inferred user affect. Hume's approach bypassed embodiment entirely, focusing on voice-based empathy for contact centers and healthcare communication. Investor and media coverage through 2025 and 2026 framed EVI as "emotionally intelligent AI," though it operated in the auditory domain rather than physical robotics.
Regulations Playing Catch-Up
The regulatory environment surrounding emotionally expressive humanoids is fragmented, evolving, and—perhaps predictably—struggling to keep pace with the technology. The EU AI Act's prohibitions on emotion recognition in workplaces and schools, effective since June 13, 2024, directly constrain deployment scenarios in Europe. Applications outside those banned contexts fall under high-risk categorization, triggering requirements for risk management systems, data governance, logging, transparency documentation, and human oversight.
The U.K.'s Information Commissioner's Office published biometric recognition guidance in February 2024 and updated its AI and biometrics strategy in March 2026, emphasizing necessity, proportionality, and data protection impact assessments. In the United States, the Federal Trade Commission took enforcement action in December 2024 against IntelliVision Technologies for deceptive claims about facial recognition accuracy—a signal of heightened scrutiny in biometric and affective computing domains. The FTC's May 2023 policy statement warned about biometric harms. State-level frameworks including California's Consumer Privacy Rights Act treat biometric data as sensitive. Regulations adopted by the California Privacy Protection Agency in September 2025, effective January 1, 2026, require risk assessments for high-risk processing involving biometrics.
China's Personal Information Protection Law, effective November 1, 2021, classifies biometric information as sensitive personal data requiring separate consent and strict necessity justifications. On March 21, 2025, Chinese authorities issued "Administrative Measures for the Security Management of Facial Recognition Technology Applications," effective June 1, 2025, adding another compliance layer for any public-facing robot employing facial analysis or emotion inference.
Safety standards present a parallel challenge. UL 3300, covering service, communication, information, education, and entertainment robots, saw its 2024 edition published and recognized by OSHA's Nationally Recognized Testing Laboratory program in 2025. UL Solutions announced the first UL 3300 global certification for a public-facing robot—Simbe's Tally inventory platform—in April 2026. ISO 13482:2014 addresses personal care robot safety but entered revision in 2021. ISO/TS 15066 governs collaborative robots in industrial settings. McKinsey partners noted in an October 15, 2025 analysis that autonomous humanoids navigating unstructured public environments fall into a standards gap, lacking clear certification pathways for mobility, collision avoidance, and unpredictable human interaction.
Li Qingdu's emphasis on "intrinsic safety" in interviews through 2025 and 2026 acknowledged these uncertainties. Discussing DroidUp's design philosophy with Chinese outlets including Jingwei in September 2025 and in a June 2026 Autohome forum interview, Li stressed speed limits, elastic materials to reduce collision forces, and scenario selection criteria aimed at minimizing risk exposure before comprehensive standards solidify.
In other words: proceed cautiously while the rulebook is still being written.
What Happens Next
The next 12 to 24 months will likely clarify which humanoid applications achieve commercial traction and which remain demonstration projects. Growth in concierge, hospitality, and exhibition deployments appears probable, supported by existing UL 3300 certifications and manageable regulatory burdens for robots operating under human supervision in bounded environments. Factory pilots for task learning and repetitive assembly—already underway with UBTECH's Walker series and other platforms—should expand as automakers and contract manufacturers refine procurement specifications emphasizing safety margins and cost per task.
Eldercare deployments face a longer horizon. China Daily reported on March 16, 2026 that while pilot programs were proliferating, "truly humanoid" service robots in nursing homes remained roughly five years away from meaningful scale. The technical challenges are formidable: safe navigation in cluttered residential settings, adaptive manipulation of fragile objects, emotionally appropriate interaction with cognitively impaired individuals. All of this must be solved at production cost points. Regulatory pathways for continuous biometric and affect monitoring of elderly users under China's PIPL and facial recognition measures—or under the EU AI Act's high-risk provisions—will require data governance infrastructure that early-stage robotics vendors have yet to demonstrate.
Verification of vendor claims will become a competitive differentiator, perhaps the competitive differentiator. As McKinsey analysts noted in reports from June and October 2025, the humanoid sector's credibility hinges on transparent testing, third-party audits of performance metrics, and adherence to emerging safety standards. DroidUp's 92% gait accuracy figure, AgiBot's 5,000-unit production assertion, and similar claims await independent substantiation through peer-reviewed publications, certified test facilities, or disclosed shipment data. Investors and procurement officers conditioned by decades of industrial robotics' measurable ROI metrics are unlikely to accept marketing copy as a substitute for validated performance benchmarks.
The scientific debate over emotion recognition's validity adds another layer of uncertainty. Barrett's May 2025 analysis and the EU AI Act's cautious stance suggest that robots marketed on "emotional realism" may face both technical limitations and regulatory pushback, particularly in healthcare and education contexts where vulnerable populations are involved. The gap between a robot displaying scripted facial movements and a system that genuinely perceives and responds to human emotional states remains vast. Overselling the latter invites both consumer disappointment and enforcement risk.
Goldman Sachs' $38 billion forecast for 2035 and Morgan Stanley's multi-trillion-dollar scenarios for the 2040s and 2050s reflect optimism grounded in demographic aging, labor shortages, and continued AI advancement. But those projections embed assumptions about regulatory harmonization, safety certification pathways, cost curves on actuators and sensors, and breakthroughs in embodied learning—all of which remain contingent. The International Federation of Robotics' 2025 data showing 91% growth in medical robots and steady expansion in service categories provides a credible baseline. Yet medical robots today are largely surgical platforms or telepresence carts operating under clinical supervision, not autonomous humanoids navigating public spaces.
Where the Opportunity (Really) Lies
For healthtech founders, the near-term opportunity lies less in deploying fully autonomous emotional companions and more in augmenting human caregivers with robots that handle physically repetitive tasks—lifting, fetching, monitoring—while humans retain responsibility for emotional labor and clinical judgment. Robotics investors should prioritize teams that publish technical validation, engage with standards bodies, and design compliance architecture into their platforms from the outset rather than retrofitting after deployment.
AI researchers face the dual challenge of advancing multimodal affect recognition while acknowledging its scientific and ethical limitations, particularly as regulatory frameworks codify skepticism into law. Policy makers confront the harder task of balancing innovation incentives with precautionary principles in domains where the stakes are high and the technology is immature. The EU AI Act's tiered risk approach, China's sensitive data classifications, and emerging U.S. state-level biometric frameworks suggest convergence toward heightened scrutiny of affective AI, even as national industrial strategies push humanoid development. Navigating that tension will require ongoing dialogue among technologists, ethicists, regulators, and the populations most affected by deployment decisions.
MOYA's warm skin and expressive micro-servos represent one company's bet on what elder care and service work will demand. Whether that bet pays off depends not on a single percentage claim echoed across press releases, but on rigorous testing, transparent reporting, regulatory compliance, and a willingness to acknowledge the distance between what current technology can reliably deliver and what marketing materials promise.
In an industry where a 92% claim can circulate unchallenged for months, that gap between promise and proof has rarely mattered more. The robots are coming—that much seems certain. Whether they arrive with the capabilities their makers advertise is a different question entirely.
