Embodied AI enters the multi-scenario commercialization stage, with WAMs, real data and dexterous hands becoming competitive focal points

Overview

The report summarizes visits in July 2026 to six embodied AI companies: Galaxea AI, OneRobotics, Yunji, Linkerbot, Orbbec and PaXini. The research covers robot complete-machine manufacturers and core component suppliers, focusing on application scenario expansion, WAM evolution, real-world data collection, cost reduction in dexterous hands, and industrialization of visual and tactile sensing.

Core views

First, industrial and logistics scenarios are expected to become the main deployment scenarios for embodied AI over the next several years, while home and service markets offer greater long-term potential. Second, by integrating world models with VLA, WAM is expected to improve robots' understanding of physical environments, generalization capabilities and ability to complete complex tasks lasting several minutes. Third, model competition will increasingly depend on the scale, quality and scenario diversity of real-world data, and UMI and first-person-view solutions can significantly reduce collection costs. Fourth, dexterous hands and tactile sensors are benefiting from economies of scale, production process improvements and modular design, with cost declines supporting commercialization. Fifth, Greater China complete-machine and component companies play an important role in the global humanoid robot supply chain.

Analysis framework

The research uses company site visits and management discussions to conduct a horizontal comparison of three robot complete-machine companies and three key component companies, and assesses industry trends based on product portfolios, model architectures, data collection methods, shipments and market share, cost reduction paths and management operating targets.

Methodology notes

• Model ArchitectureWorld Action Model (WAM)

  Combining world models with vision-language-action models

  This architecture aims to enable robots to better learn the physical laws of the real world, improve cross-scenario generalization capabilities, and support long-horizon complex tasks such as laundry, folding clothes, meal preparation and dishwashing.

• Model ArchitectureVision-Language-Action Model (VLA)

  Unified modeling of visual inputs, language instructions and robot actions

  VLA is one of the foundational architectures of embodied AI; the report believes that combining it with world models to form WAM can address the shortcomings of pure VLA in learning physical laws and handling long-cycle tasks.

• Multimodal ModelVision-Tactile-Language-Action Model (VTLA)

  Adding tactile information on top of VLA

  Dexterous hands and multidimensional tactile sensors can provide data such as contact, pressure, torque and slip, supporting fine manipulation and high-quality multimodal model training.

• Data CollectionUMI and First-Person-View Data Collection

  Collecting real-world operation data in a low-cost manner

  Companies use a combination of UMI, first-person-view data and teleoperation to balance data quality, scale and cost; the report cites company information indicating that the cost of the first two methods can be about 90% lower than teleoperation.

Asset mapping & comparison

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

• OneRobotics (6600 HK)

  Complete-machine company for home service, tennis training and companion robots

  Strengths

  Has the “One Brain, Multiple Embodiments” platform, SwitchBot product base and OneModel 1.7, and plans large-scale collection of home-scenario data.

  Weaknesses

  New embodied AI products are still in the commercialization expansion stage, with accuracy for complex home tasks at about 70%, leaving room for improvement.

  Comparison

  Compared with industrial and hotel robot companies, it is more focused on home, sports and companionship scenarios, with a larger long-term potential market but also higher task complexity.

  Risks

  Failure to achieve high revenue growth targets, home robot demand below expectations, insufficient model reliability and low returns on data collection investment.

• Yunji (2670 HK)

  Hotel and commercial service robot company

  Strengths

  Has delivered over 60,000 robots and covers more than 40,000 hotels in China, with mature scenarios and customer base.

  Weaknesses

  The current business is highly dependent on hotel scenarios and needs to expand growth potential through new markets, new services and AI agent revenue.

  Comparison

  Compared with home robot companies, tasks such as hotel delivery are more standardized and commercial deployment is more mature.

  Risks

  Slower hotel penetration, cross-scenario expansion falling short of expectations, longer hardware replacement cycles and slow increase in the share of high-margin AI agent revenue.

• Orbbec (688322 CH)

  Supplier of robot 3D vision systems and data collection equipment

  Strengths

  Has full-stack capabilities in chips, structured light, iToF, binocular vision, LiDAR, algorithms and software, with a market share of over 70% in China's service robot 3D vision system market.

  Weaknesses

  Robot-related business currently accounts for about 30% of total revenue, and the business structure transformation still needs to continue materializing.

  Comparison

  Compared with complete-machine manufacturers, Orbbec captures supply-chain opportunities across multiple types of robot platforms through visual sensing and data collection tools.

  Risks

  Long-term camera price declines compressing value, customer concentration, changes in technology routes and the robot business share rising less than expected.

• Galaxea AI (unlisted)

  Embodied AI foundation model and robot hardware company

  Strengths

  Has the open-source VLA model G0.5, Fast-WAM, and a portfolio of wheeled robots, dual-arm mobile robots and robotic arms, with an emphasis on real-world data.

  Weaknesses

  Commercialization scale, profitability and model stability in large-scale deployment have not been sufficiently disclosed.

  Comparison

  It has a high degree of integration between models and hardware, with data collection covering various scenarios such as logistics, catering, hotels and retail.

  Risks

  Data targets falling short of expectations, insufficient model generalization, difficulty expanding hardware deliveries and pressure from sustained R&D investment.

• Linkerbot (unlisted)

  Supplier of high-DoF dexterous hands and manipulation data

  Strengths

  Covers three technology routes: tendon-driven, direct-drive and linkage mechanisms; has a relatively high market share in high-DoF dexterous hands and achieves significant cost reductions through scale and modularization.

  Weaknesses

  The business is concentrated in the still-early-stage dexterous hand market, and end demand is sensitive to the mass-production progress of humanoid robots.

  Comparison

  Compared with complete-machine manufacturers, it benefits more directly from upgrades in fine manipulation capabilities and demand for operation data collection.

  Risks

  Delayed industry mass production, price competition, higher product durability requirements and declining market share.

• PaXini (unlisted)

  Supplier of multidimensional tactile sensors, dexterous hands, wheeled robots and data platforms

  Strengths

  Has 6D Hall-effect tactile technology, significant sensor cost-reduction capabilities and plans for five data factories, which can provide multimodal data for VTLA models.

  Weaknesses

  Large-scale data factory construction, data commercialization and cross-product synergies still need to be validated.

  Comparison

  Compared with pure sensor companies, PaXini covers tactile hardware, robots, data collection and data sales platforms, with a more complete industry-chain layout.

  Risks

  Delays in data factory construction, data quality or demand falling short of expectations, continued sensor price declines and insufficient business model realization.

Key data

• Number of surveyed companies6 companiesIncluding Galaxea AI, OneRobotics, Yunji, Linkerbot, Orbbec and PaXini.
• Leading companies' real-world data targetsOver 1 million hoursGalaxea AI plans to collect over 1 million hours in 2026; OneRobotics targets 1 million hours of robot data.
• Cost of UMI and first-person-view dataAbout 90% lower than teleoperationData disclosed by Galaxea AI; actual costs depend on collection scenarios and quality requirements.
• OneRobotics 2026 revenue expectationRMB1.4-1.5bnThe company expects YoY growth of over 50% and expects a revenue CAGR of no less than 50% in 2026–2028.
• OneRobotics home-task accuracyAbout 70%Applicable to home scenarios disclosed by the company, such as laundry, folding clothes, meal preparation and dishwashing.
• Yunji robot delivery scaleOver 60,000 unitsIt has served more than 40,000 hotels in China.
• Linkerbot global share in high-DoF dexterous hands80%Corresponds to the market for high-DoF dexterous hands with more than 20 degrees of freedom, based on the company's disclosure.
• Linkerbot O6 Lite selling priceAs low as RMB3,999 (USD590)Cost reductions come from economies of scale, production process improvements and modular design.
• Orbbec share in China's service robot 3D vision marketOver 70%Based on the company's disclosure for China's service robot 3D vision system field.
• Value per humanoid robot vision systemAbout RMB2,000-3,000Estimated based on the use of two RGBD cameras per robot.
• PaXini tactile sensor costReduced from RMB100,000 to RMB200/unitThe company says this was achieved within 5 to 6 years through product design improvements and economies of scale.
• PaXini data factory plan5 facilities, annual output of over 10 billion data entriesThe plan covers Tianjin, Suqian, Wuhan, Zigong and Ganzhou, with scenarios including logistics, industrial manufacturing and retail.

Impact & implications

Industry value may extend further from pure hardware sales to models, data and AI agent services. Near-term beneficiaries include service and humanoid robot complete machines, dexterous hands, RGBD vision, tactile sensors and data collection infrastructure; medium- to long-term competitive barriers will depend more on real-data closed loops, complex-task generalization, scaled manufacturing capabilities and the pace of cost reduction. For listed companies, a higher revenue contribution from robotics businesses and commercialization of new products may provide valuation catalysts, but management targets still need to be validated through orders, deliveries, revenue and profit realization.

Risks

• WAM, VLA and VTLA are still in a rapid iteration stage, and generalization capabilities, stability and safety in complex environments may fall short of expectations.
• Real-world data collection costs remain high, and million-hour targets and data factory plans face execution risks.
• Demand and penetration speed for robots in industrial, logistics, service and home scenarios may be lower than management expectations.
• Rapid price declines in dexterous hands, vision and tactile sensors may expand demand but could also compress supplier margins.
• Market shares, growth targets and technical indicators disclosed by companies mainly come from management information and still need to be validated by subsequent operating data.
• Intensifying supply-chain competition, changes in technology routes and customer concentration may weaken the advantages of existing leading companies.
• BofA Securities may have business relationships with issuers covered by its research, and investors should use this report only as one factor in decision-making.

What to watch

• WAM success rates, generalization capabilities and real commercial deployment progress in complex tasks lasting several minutes.
• The completion progress of Galaxea AI and OneRobotics toward their million-hour real-world data targets in 2026.
• The actual cost, efficiency and data quality of UMI and first-person-view data relative to teleoperation.
• The realization of OneRobotics' 2026 revenue target of RMB1.4-1.5bn and its 2026–2028 CAGR target of no less than 50%.
• Yunji's expansion into scenarios beyond hotels and the increase in the share of high-margin AI agent revenue.
• Whether Orbbec's robot-related revenue share can rise from about 30% to over 50% within two years.
• Changes in Linkerbot's dexterous hand shipments, prices, durability and high-DoF market share.
• PaXini's five data factories' commissioning progress, annual data capacity and external sales of OmniSharingDB.