China's humanoid robots: models, data, and the cost curve continue to advance, while scaled deployment is still expected in 2027-2029

Overview

This report is based on Goldman Sachs' exchanges with 14 robotics companies during the GS Asia Communacopia + Technology Conference in Hong Kong and China AI robotics field research in Shenzhen and Beijing, covering both private and listed companies. The core conclusion is that China's humanoid robotics and embodied intelligence industry is moving from proof of concept toward a clearer technological and commercial path: on the model side, VLA/VTLA and world models are being integrated more rapidly; on the data side, greater emphasis is being placed on scalable, high-quality real-world data collection; on the application side, industrial handling, logistics sorting, picking, inspection, and other structured scenarios are the priority entry points; and on the cost side, costs continue to decline through scale and full-stack R&D control. At the same time, however, the report stresses that the industry is still some distance from broad scaled deployment, with model quality, data accumulation, POC conversion, cost thresholds, and safety and stability remaining key constraints.

Core views

First, the model path is evolving from standalone VLA to a combination of VLA/VTLA plus world models to improve planning, action validation, and robustness in real-world scenarios. Second, high-quality real-world data is the industry's main bottleneck, and companies are placing visibly greater emphasis on human-centric, first-person, tactile, wearable, VR, customer-site, and data-factory approaches. Third, commercialization is no longer limited to demonstrations; industrial, logistics, and service scenarios are advancing through POCs and small-batch validation, but large orders and cross-scenario generalization still require more time. Fourth, near-term hardware form factors are more pragmatic: wheeled robots with two- to three-finger grippers are considered capable of covering about 70%-90% of industrial applications, while bipedal robots with five-finger dexterous hands are more of a medium- to long-term direction. Fifth, from an investment perspective, the industry's long-term outlook remains attractive, but investors need to remain patient about large-scale deployment before 2027-2029.

Analysis framework

The report uses industry-chain research and one-by-one company interviews, conducting cross-company comparisons around models/embodied intelligence stacks, data strategies, commercialization progress, cost-reduction paths, and the competitive landscape, while observing private and listed companies within the same technological and commercialization framework.

Methodology notes

• Technology roadmapVLA/VTLA + World Model

  Combining vision-language-action and vision-tactile-language-action models with world models

  The report views this combination as a key path to improving robot planning, next-state prediction, pre-action validation, and robustness in real-world scenarios.

• Data moatHigh-quality real-world data collection

  Accumulating multidimensional data through data factories, customer deployments, wearable devices, VR, first-person, and human-centric collection

  The report believes the data issue has shifted from generic data recipes toward scalable, reusable, and verifiable data collection architectures.

• Commercialization funnelPOC-small batch-validation-pilot-scaled deployment

  Industrial customers typically first go through a 3-6 month POC, 2-3 rounds of validation, small-batch testing, and roughly 12 months of verification before moving to larger orders

  This framework is used to judge that the industry is still at an early commercialization stage, and order scale-up requires further maturity in model quality and scenario stability.

• Cost curveScale and full-stack R&D control

  Reducing costs through in-house development of core components, reuse of robot platforms, shipment expansion, and form-factor optimization

  The report emphasizes that cost reduction is not driven solely by lower component prices, but by the combined effects of product form factor, architectural choices, domestic supply-chain substitution, and scale effects.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• China's humanoid robotics and embodied intelligence industry chain

  Industry thematic asset

  Strengths

  The technology roadmap is becoming clearer, with VLA/VTLA, world models, real-world data, and cost reduction forming a common direction of progress.

  Weaknesses

  Most scenarios are still at the POC, pilot, and small-batch stage, and stable quality and cross-scenario generalization have not yet been fully validated.

  Comparison

  Compared with the pure concept stage, the industry has entered a more concrete phase of scenario, data, and cost validation.

  Risks

  Commercialization later than expected, insufficient model capability, insufficient data quality, slow customer order conversion, and weaker-than-expected cost reduction.

• Estun Automation (002747.SZ/2715.HK)

  Listed industrial automation company; the report lists ratings as Sell/Neutral

  Strengths

  Benefiting from domestic substitution, overseas expansion, and improved product mix in medium- to high-payload robots, with management emphasizing margins and growth quality.

  Weaknesses

  After losses in 2024, the company's strategy shifted to place greater emphasis on profitability and project quality, implying high execution dependence.

  Comparison

  Compared with early-stage humanoid robot companies, Estun is more focused on traditional industrial robots and automation platforms.

  Risks

  Intensifying domestic competition, pricing pressure, low-margin project selection, and execution risks in overseas expansion and cost control.

• Dobot (2432.HK)

  Listed collaborative robotics and embodied intelligence company; the report lists it as Not Covered

  Strengths

  Has a collaborative robot platform and major-customer base, with fast growth in embodied AI revenue, and products covering robotic arms, robot dogs, wheeled robots, and bipeds.

  Weaknesses

  Embodied AI still accounts for a low share of revenue, and the industry's model and data roadmap has not yet fully converged.

  Comparison

  Compared with pure humanoid robot companies, Dobot places more emphasis on one brain across multiple form factors and reuse of existing collaborative robot capabilities.

  Risks

  New product ramp falling short of guidance, insufficient industrial POC conversion, and changes in hardware form-factor choices and customer demand.

• Geek+ (2590.HK)

  Logistics robot company; the report lists it as Not Covered

  Strengths

  Has a clear scenario-first strategy, leveraging real warehouse task flows and a data closed loop, and focusing on deployable scenarios with clearly defined boundaries.

  Weaknesses

  Management believes industry models, data, and technology roadmaps have not yet converged, so it takes a relatively cautious approach toward cutting-edge embodied AI.

  Comparison

  Compared with the general-purpose humanoid robot path, Geek+ is more focused on deepening logistics scenarios and prioritizing reliability.

  Risks

  Constraints in resource delivery capability, customer education cycles, the pace of logistics automation demand, and new-customer conversion.

• UBTech (9880.HK)

  Listed humanoid robot company; the report lists it as Not Covered

  Strengths

  The report notes that the company believes demand for data factories in 2026 will remain strong or become even stronger, reflecting support from data-related demand.

  Weaknesses

  The provided excerpt lacks more complete company-level details on models, revenue, and costs.

  Comparison

  As a listed humanoid robot asset, its data and application progress can serve as a temperature check on industry demand.

  Risks

  Fluctuations in demand for data factories, product deployment pace, order quality, and profitability validation.

• Galaxea

  Private embodied intelligence company

  Strengths

  Covers the full stack from VLA plus world models to data and robot hardware, emphasizes a real-data moat, and has already open-sourced 500 hours of internal data.

  Weaknesses

  Cross-platform transfer still requires post-training iteration and is not yet fully suitable for broad deployment.

  Comparison

  Occupies a key focus position in the report in terms of real-world data, UMI gloves, and wheeled robot form factors.

  Risks

  Data scalability and quality, effectiveness of model post-training, customer deployment speed, and the application boundaries of wheeled form factors.

• Linkerbot

  Private dexterous hand and skill data company

  Strengths

  Has outstanding market share and mass-production capacity in high-DOF dexterous hands, along with the LinkerSkillNet skill database and multiple dexterous-hand product lines.

  Weaknesses

  Software/skills are currently mostly monetized together with hardware, and the standalone Skill Store faces difficulty monetizing domestically.

  Comparison

  Compared with full-machine humanoid robot companies, Linkerbot is more like an infrastructure supplier for dexterous hands and precision skills.

  Risks

  Standards for the dexterous-hand path have not yet been unified, the skill-payment model is immature, and there is pressure from customer adoption cycles and hardware cost declines.

• PaXini

  Private tactile, data, and sensing infrastructure company

  Strengths

  Has a human-centric non-teleoperation data collection system and 5 data factories, and tactile sensing costs have declined significantly versus historical levels.

  Weaknesses

  Current commercialization is more concentrated in data and tactile sensing, while scaling of full-machine applications still depends on downstream models and robot customers.

  Comparison

  Compared with vision-led VLA approaches, PaXini emphasizes the importance of tactile sensing and VTLA for physical interaction tasks.

  Risks

  Sustainability of data sales, competition in tactile approaches, downstream customer procurement pace, and progress in standardization.

• Spirit AI

  Private embodied model company

  Strengths

  Emphasizes the integration of VLA and world models, real interaction data, and open-model performance, and has passed POCs and entered small-batch procurement in industrial scenarios such as CATL.

  Weaknesses

  Management believes foundation models are still insufficient to support broad commercialization, with post-training time and cross-scenario generalization remaining constraints.

  Comparison

  More focused on the robot brain and data-factory route, with a commercialization pace that may lag hardware-oriented companies.

  Risks

  Model success rates staying below the threshold for scaled deployment, slow expansion in industrial scenarios, and high safety and stability thresholds in household services.

• One Robotics (6600.HK)

  Service and general-purpose humanoid robot company; the report lists it as Not Covered

  Strengths

  Targets an end-to-end embodied AI architecture, advancing VLA plus world models and combining household and service-scenario data factories.

  Weaknesses

  Consumer-grade deployment is still on an approximately three-year timeline, and the shortage of household-scenario data and the BOM threshold remain to be solved.

  Comparison

  Compared with industrial-scenario companies, One Robotics places greater emphasis on semi-structured services, government deployments, and eventual consumer scenarios.

  Risks

  The speed of reducing BOM to the US$3k threshold, safety and stability in household scenarios, consumer acceptance, and the sustainability of service-scenario orders.

Key data

• Research coverage14 companiesIncluding Daimon Robotics, Dobot, Estun Automation, Galaxea, Galbot, Geek+, LimX Dynamics, Linkerbot, Mech-Mind, One Robotics, PaXini, Spirit AI, UBTech, and X Square Robot.
• Large-scale deployment window2027-2029Many industry participants expect large-scale deployment to become more likely after accumulating tens of millions of hours of high-quality data and forming deployable models.
• Industrial POC cycleTypically 3-6 months, averaging 2-3 roundsThis is usually followed by small-batch testing and about 12 months of validation, before orders gradually scale up.
• Small-batch and pilot order sizeSmall batches are typically below 50 units; follow-up pilot orders are about 50-100 units per customerThis reflects that industrial deployment is still in a phased conversion process.
• Model scale discussionAbout 40B-80B parametersThe report mentions that discussion of the model stack is shifting from smaller billion-parameter pretraining systems toward larger-scale models.
• Coverage rate of pragmatic hardware form factorsAbout 70%-90% of industrial applicationsMany companies believe a wheeled base plus a two- to three-finger gripper is more realistic in the short term, while bipeds and five-finger dexterous hands remain more medium- to long-term.
• Dobot embodied AI revenueAccounted for 4% of revenue in 2025, up 4x year over yearManagement expects continued rapid growth in 2026; 2026 shipment guidance is about 300-500 units.
• Dobot humanoid robot pricing and gross marginASP about Rmb200k-500k, averaging about Rmb300k; GPM about 45%This indicates that some industrial scenarios already have a commercial pricing and gross-margin framework.
• Galaxea data targetExpand data volume to 1 million hours in 2026The company emphasizes high-quality data such as real-robot teleoperation, UMI gloves, exoskeletons, and first-person data.
• Linkerbot market and capacityMore than 80% global share in high-DOF dexterous hands; mass production of 1,000+ units per monthThe company says peak output can reach 4,000 units, with customers spanning leading humanoid robot companies, industrial, 3C, automotive, research, and data customers.
• PaXini data factory5 data collection factoriesLocated in Tianjin, Suqian, Wuhan, Zigong, and Ganzhou, focused on human-centric, non-teleoperation data collection.
• Spirit AI data deployment800+ robots; year-end target of 1 million hours of real-world dataThe company believes real interaction data is the core variable for embodied model capability.
• One Robotics long-term price conceptOnero gripper version about US$10k ASP; key BOM threshold for household adoption about US$3kThe company plans to enter through semi-structured service scenarios and government deployments, with the consumer side on an approximately three-year timeline.

Impact & implications

For investors, the report conveys a signal of clear technological progress but a commercialization pace that still requires patience. Model integration, real-world data infrastructure, and cost reduction enhance the sector's long-term investability; however, most revenue and orders are still at the POC, small-batch, or early validation stage, so near-term valuations need to be cross-checked against actual shipments, data revenue, customer repeat purchases, and gross-margin improvement. More direct beneficiaries may include companies with an industrial customer base, a real-world data closed loop, in-house core component capabilities, or infrastructure attributes in data, sensing, or dexterous hands.

Risks

• Model quality and cross-scenario generalization may fall short of expectations, making it difficult for POCs to convert into stable orders.
• Accumulation of high-quality real-world data may be insufficient, or data collection costs may be higher than expected.
• Industrial customer validation cycles are long, making it difficult for small-batch orders to scale up promptly beyond 50-100 units.
• Hardware cost declines may be slower than expected, affecting gross margins and end-customer ROI.
• Intensifying domestic competition and price competition may compress profitability.
• Technology paths such as tactile sensing, dexterous hands, wheeled robots, and bipeds have not yet been fully standardized, which may lead to repeated shifts in R&D and product direction.
• Household and consumer service robots face higher safety, stability, and regulatory thresholds.
• If demand from data factories or government-related demand fluctuates, it may affect some companies' expectations for data revenue.

What to watch

• Changes in each company's real-world data hours, data quality, and share of data revenue in 2026.
• Whether the success rate of VLA/VTLA plus world models on new tasks can improve from about 40%-50% toward the 60%-70% threshold closer to scalable deployment.
• Whether industrial POCs can steadily convert into small-batch procurement and further expand to 50-100 unit orders per customer.
• Changes in the actual order mix among wheeled robots, two- to three-finger grippers, bipeds, and five-finger dexterous hands.
• Whether domestic substitution of core components, BOM declines, and whole-machine ASPs lead to gross-margin improvement.
• The pace of shipments, orders, gross margins, and R&D investment at listed companies such as Dobot, Estun, Geek+, and UBTech, as well as leading private companies.
• Whether expectations for large-scale deployment in 2027-2029 are brought forward or pushed back.