WRC 2026 Shows That the Focus of Robotics Competition Has Shifted from Demonstration Capabilities to Productivity, Reliability and Deployment Data

Overview

The report summarizes observations from WRC 2026 and discussions with multiple robotics companies. Its core conclusion is that the customer base, application scope and data investment for humanoid robots and other forms of physical AI are expanding, but the industry has yet to cross the threshold for large-scale commercialization. The next stage of development will be determined not by basic motion demonstrations, but by whether robots can approach human productivity in real-world environments and operate continuously with sufficiently high reliability, data efficiency and economic viability.

Core views

First, robot deployment is broadening before deepening. The average number of customers per integrator has increased from approximately 10 at the beginning of 2026 to dozens, and some customers have moved from proof-of-concept trials to repeat purchases. For example, Foxconn made repeat purchases from UBTECH after its first deployment in late 2025. However, individual customers generally still deploy only dozens of robots, leaving the industry far from large-scale adoption. Robots have improved their generalization to unfamiliar objects, changes in position and certain environmental disturbances. Specific tasks such as sorting can match or even exceed human performance in exhibition demonstrations, but their speed and accuracy on complex or high-precision tasks still lag humans significantly. Semi-structured environments are currently a relatively suitable starting point for deployment: they vary enough that conventional fixed automation struggles to cover them, but their complexity still has clearly defined boundaries. Galaxea's solutions cover fulfillment, sorting, tote handling, picking and packaging and reportedly can handle approximately 10,000 SKUs and unfamiliar objects at 70–80% of human speed. UBTECH's industrial deliveries in 2026 are focused mainly on depalletizing and material handling, with depalletizing reaching 60–80% of the cycle time required for production. The report also emphasizes that robot intelligence does not depend solely on new humanoid platforms. Mech-Mind upgrades conventional robots with 3D vision, software/AI and end effectors. It shipped approximately 10,000 systems in 2025, recorded approximately 70% year-over-year revenue growth in the first quarter of 2026 and achieved similar order growth in the first half. Quadruped robots have also achieved clearer product-market fit in applications such as power-grid inspection, security, firefighting, steel, and oil and gas. Deep Robotics grew by more than 100% in the first half of 2026 and aims to increase shipments of industrial-grade quadruped robots from approximately 3,000 units in 2025 to approximately 10,000 units in 2026. Household use cases were significantly more prevalent than at WRC 2025, but commercial value, efficiency and safety issues will take time to resolve. The report does not expect household humanoid robots to scale rapidly; AI pets, companion robots, desktop robots and automated food and beverage workstations may be adopted sooner. Corporate targets remain aggressive. Geekplus plans cumulative shipments of 10,000 embodied-intelligence products over the next three years, approximately two-thirds of which will be Gino wheeled humanoid robots and one-third robotic arm workstations. UBTECH targets approximately 10,000 units in 2026 and 20,000–40,000 units in 2027. Linkerbot plans to sell approximately 20,000 dexterous hands in 2026, compared with approximately 10,000 in 2025. AI² Robotics targets approximately RMB500 million in revenue and deliveries of several thousand units in 2026 and plans annual capacity of approximately 30,000 units in 2027. DexForce targets revenue of approximately RMB250–300 million, compared with RMB60 million in 2025. OneRobotics targets a revenue CAGR of approximately 50% from 2026 to 2028. Deep Robotics targets deliveries of approximately 10,000 industrial-grade quadruped robots in 2026. These targets reflect the supply side's willingness to expand, but the report continues to regard actual productivity and commercial ramp-up as key areas requiring validation. For dexterous hands, WRC 2026 featured significantly more manipulation tasks, with the industry shifting from basic motion to multi-step operations. Dexterous hands are suitable for tasks requiring greater flexibility, such as using human tools, tightening screws and handling multiple SKUs, but conventional grippers still dominate structured applications such as standardized sorting. Different architectures correspond to different cost and performance ranges: Linkerbot's low-DoF linkage products cost approximately RMB3,000–6,000, high-DoF direct-drive products approximately RMB20,000–50,000, and tendon-driven products up to approximately RMB90,000–100,000. The company believes 15–20 degrees of freedom can cover most practical operational needs. Linkage systems have the lowest cost and relatively better reliability but less flexibility. Direct-drive systems provide more precise control but increase cost, thermal load and integration complexity. Tendon-driven systems most closely resemble human hands, but service life and cost impede scaled adoption. Even though laboratory grasping tests have exceeded 1 million cycles, the longest continuous powered test remains under six months. Long-term operation may also encounter software failures, repeated recalibration, accuracy degradation, overheating and damage to transmission components. Visual, joint, force and tactile data still need to be synchronized; data cannot readily be shared between different hand designs, and some on-site customer data cannot be retrieved for training. Until reliability and productivity improve, high-DoF solutions will also struggle to meet industrial customers' return-on-investment requirements. Model approaches continue to evolve rapidly. The report believes vision-language-action models, world models and reinforcement learning will coexist: vision-language-action models understand instructions, objects and task context; world models characterize physical dynamics and predict outcomes; and post-training, reinforcement learning and control models translate general capabilities into reliable execution of specific tasks. Some applications previously required 6–12 months of pre-training, while better foundation models can reduce the incremental data needed for relatively simple new tasks to dozens of hours. Therefore, compared with a single benchmark score, the time needed to complete a new task and the amount of data required per new task may become more effective indicators of model quality. Several companies expect embodied intelligence's “GPT moment” to arrive in 18–24 months, or around 2028, but the report regards this expectation as optimistic. Industry definitions of this inflection point include a 70–80% zero-shot success rate on unfamiliar tasks, near-production performance after limited post-training, the ability to execute long-horizon tasks continuously, resilience to dynamic disturbances and stochastic environments, and substantial reductions in the incremental data and engineering work required for each new task. The report identifies real-world performance, commercial ramp-up and whether data scale can translate into generalization as key validation criteria, and believes rapid adaptation to technological change is more important than betting on a single architecture. Data remains the most frequently cited bottleneck, but raw hours are becoming an increasingly weak indicator of capability. Different tasks require different data recipes: locomotion and performance tasks rely more heavily on simulation; navigation, grasping and placement can use more egocentric data; and household tasks depend more on teleoperation. Data quality varies considerably. Approximately 80% of Galbot's internal teleoperation data remains usable after screening, whereas the usable rate for low-cost outsourced egocentric data may be below 10%. UBTECH aims to obtain approximately 1 million hours of usable data in 2026 but estimates that this may require approximately 4 million hours of raw data. Outsourced egocentric data can cost as little as approximately US$2 per hour in low-cost countries, while high-quality multimodal hand data costs approximately US$90–100 per hour. Value may therefore shift from basic collection toward task design, multimodal synchronization, quality control, cleaning and alignment, and model-linked data selection. Galbot targets approximately 1 million hours in 2026 and approximately 10 million hours by the first half of 2028. Galaxea targets approximately 1 million effective hours by the end of 2026, approximately 90% from egocentric sources or UMI and approximately 10% from physical robots. UBTECH currently has approximately 300,000 hours and targets approximately 1 million usable hours in 2026. PsiBot currently has approximately 300,000 hours and views approximately 1 million hours as a potential threshold for more pronounced generalization. Longer-term barriers are more likely to arise from failures, human interventions and edge cases generated by deployments. Fleet scale alone is insufficient: data rights, scenario diversity and the speed of the feedback loop determine whether data can improve models. For customer data that cannot leave the factory, the report also discusses distributed training solutions involving local storage and on-site training, with only the improved model transmitted back. Compute is increasingly looking like a consequence of data expansion rather than an independent technology race. Several companies have reached the scale of thousands of accelerators. UBTECH has cloud training resources at the thousand-accelerator level and plans to reach the ten-thousand-accelerator level within the next 3–5 years. Galbot is also at the thousand-accelerator level, and its computing power may expand alongside data and parameter scale, potentially reaching 100,000 accelerators by 2028. AI² Robotics uses leased compute at the thousand-accelerator level and plans to build its own cluster, reaching the ten-thousand-accelerator level over the long term. OneRobotics has approximately 1,000 H100-equivalent resources. Training compute itself may not be a durable moat, but it can be used to cross-check a company's data scale, model capabilities and iteration speed. Integrators currently rely almost entirely on overseas chips while actively testing domestic alternatives. Some companies estimate that one Nvidia chip is equivalent to approximately four domestic chips, indicating a substantial performance gap, although domestic compute has progressed from nonexistence to availability. As models, context windows and multimodal sensory inputs expand, requirements for memory bandwidth, data transfer, latency and power consumption rise, making edge compute another area for vertical integration. UBTECH and MetaX are jointly developing a dedicated chip for humanoid robots, with potential deployment in 2–3 years. Galbot is exploring compute-in-memory, domestic edge chips and cloud-edge collaborative inference. Hardware challenges have shifted from whether a robot can move and perform demonstrations to whether it can operate continuously at production intensity. The main current constraints include reliability, battery life and joint thermal limitations. Even if a model can decide what action to take, the physical system may not be able to execute it efficiently and repeatedly. Some large manufacturing customers require compensation when robots cause production-line downtime, meaning the economic loss from a single failure may greatly exceed the value of the robot itself once robots are embedded in production systems. Key metrics are therefore shifting toward continuous operating time, failure probability, calibration drift, recovery time, thermal and energy efficiency, overall weight and integration, and the durability of sensors and end effectors. Cost reduction is also relying increasingly on design simplification, material and process substitution, module standardization and production consistency rather than merely lowering component prices. Magnesium alloys, metal injection molding, carbon fiber and materials such as PEEK are being used to improve weight, wear resistance and heat resistance. Lingyi iTech is promoting battery and joint module standardization, which it estimates could reduce duplicative design and development costs by 10–20%, while simplifying complex structures could lower manufacturing costs by another 10–15%. At the stock level, Morgan Stanley believes that as deployments begin, the threshold for component suppliers will rise from passing product certification to achieving reliability, consistency, yield, mass-production scale and cost control. It therefore prefers leading suppliers with mass-production capabilities and selects Leader Harmonious Drive Systems, Hengli Hydraulic and Shuanghuan Driveline as preferred names, all rated OW. Appendix valuations show that Lingyi iTech is valued using a multi-stage residual income model with an assumed cost of equity of 11%, medium-term growth rate of 16% and perpetual growth rate of 5%. Geekplus is valued at 5.6x 2026E P/S, with 50% based on the 8.3x average P/S of humanoid robotics and autonomous-driving peers after a 20% discount and the other 50% based on the 4.6x average P/S of logistics and warehouse automation peers. Hengli Hydraulic's non-humanoid robotics business is assigned a target P/E of 30x for 2027, below the approximately 40–50x peak in the previous cycle because the domestic sales peak in this cycle is expected to be weaker, profit growth is normalizing and the pump and valve business has passed its hypergrowth phase. Its humanoid robotics business is valued using DCF with a WACC of 11% and a perpetual growth rate of 4%.

Analysis framework

The report first compares on-site observations at WRC 2026 and discussions with Geekplus, Lingyi iTech, OneRobotics, UBTECH and multiple unlisted robotics companies with WRC 2025. It then analyzes commercialization bottlenecks across six main themes: application deployment, dexterous hands, models, data, computing power and hardware. Validation metrics include actual productivity, repeat purchases and deployment scale at individual sites, data usability rates, adaptation speed for new tasks, continuous operating reliability and cost. Finally, the report maps industry trends to component suppliers with mass-production capabilities and values selected covered companies using residual income, P/S, target P/E and DCF methodologies.

Methodology notes

• (Method Outside the Vocabulary)

  Conference Research and Annual Exhibition Comparison

  The report compares on-site demonstrations and corporate discussions at WRC 2026 with WRC 2025 to identify the industry's shift in focus from basic motion capabilities to productivity, reliability and commercial deployment.

• (Method Outside the Vocabulary)

  Deployment Productivity and Data Yield Metrics

  Rather than observing only robot counts or raw data hours, the report examines deployment scale at individual sites, repeat purchases, task speed, the proportion of usable data, and the time and data required for new tasks to determine whether scale expansion can translate into actual capabilities.

• Valuation MethodRIM Residual Income Model

  Multi-Stage Residual Income Valuation

  Lingyi iTech's base case uses a multi-stage residual income model to estimate equity value with an 11% cost of equity, 16% medium-term growth rate and 5% perpetual growth rate.

• Valuation MethodPS valuation

  Peer-Weighted P/S Valuation

  Geekplus uses a 2026E P/S multiple of 5.6x, derived by combining the valuations of humanoid robotics and autonomous-driving peers with those of logistics and warehouse automation peers at a 50% weighting each.

• Valuation MethodPE/PEG valuation

  Target P/E Valuation

  Hengli Hydraulic's non-humanoid robotics business uses a target P/E of 30x for 2027, compared with a peak valuation of approximately 40–50x in the previous cycle.

• Valuation MethodDCF Discounted Cash Flow

  DCF Valuation of the Humanoid Robotics Business

  Hengli Hydraulic's humanoid robotics business is valued using a discounted cash flow model, assuming an 11% WACC and a 4% perpetual growth rate.

Asset mapping & comparison

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

• Geekplus (2590.HK)

  The report rates the company OW and views it as an investment related to the commercialization of logistics and warehouse automation and embodied-intelligence products.

  Strengths

  The cumulative shipment target for embodied-intelligence products over the next three years is 10,000 units, approximately two-thirds of which will be Gino wheeled humanoid robots. Opportunities exist for market-share gains, margin expansion and new-customer acquisition.

  Weaknesses

  Valuation must reference peers in still-early-stage humanoid robotics, autonomous driving and logistics automation, and investment in embodied intelligence may exceed expectations.

  Comparison

  The base case uses 5.6x 2026E P/S, based on a 50/50 weighting of the 8.3x valuation for humanoid robotics and autonomous-driving peers after a 20% discount and the 4.6x valuation for logistics and warehouse automation peers.

  Risks

  Expiration of lock-up periods, weaker-than-expected order growth due to geopolitics, higher-than-expected investment in embodied intelligence, and deteriorating sentiment toward the robotics or AI markets.

• Lingyi iTech (002600.SZ)

  The report rates the company UW and focuses on its standardization of key robotics components and AI hardware-related businesses.

  Strengths

  Standardization of battery and joint modules is expected to reduce duplicative design and development costs by 10–20%, while structural simplification could further lower manufacturing costs by 10–15%.

  Weaknesses

  Growth and customer qualification for new businesses such as AI infrastructure, foldable smartphones, AI glasses and XR remain uncertain.

  Comparison

  The report uses a multi-stage residual income model consistent with its technology hardware coverage framework, assuming an 11% cost of equity, 16% medium-term growth rate and 5% perpetual growth rate.

  Risks

  Slower-than-expected AI infrastructure ramp-up or customer qualification, weak smartphone and foldable-screen demand, and delays in AI glasses, XR or other new businesses.

• OneRobotics (6600.HK)

  The report lists the company as NC and uses it as an industry observation target for investment in embodied-intelligence models, data and computing power.

  Strengths

  The company targets an approximately 50% revenue CAGR from 2026 to 2028 and has approximately 1,000 H100-equivalent compute resources.

  Weaknesses

  The report notes that data from the same scenario may become rapidly redundant after several hours, making data diversity and utilization efficiency more important than simply increasing hours.

  Comparison

  Its compute scale is within the thousand-accelerator range described in the report for several leading embodied-intelligence companies.

• UBTECH (9880.HK)

  The report lists the company as NC and views it as an important observation target for industrial humanoid robot deployment, model transfer, data and edge-chip development.

  Strengths

  The company has secured repeat purchases from Foxconn, targets approximately 10,000 units in 2026 and 20,000–40,000 units in 2027, and plans to obtain approximately 1 million hours of usable data.

  Weaknesses

  Depalletizing currently reaches 60–80% of the required production cycle time, leaving a gap before labor can be fully replaced. Approximately 1 million hours of usable data may require approximately 4 million hours of raw data.

  Comparison

  Industrial deliveries focus mainly on semi-structured use cases such as depalletizing and material handling, consistent with the early deployment path described in the report.

  Risks

  Industrial customers may demand compensation when robots cause production-line downtime, amplifying the economic losses arising from insufficient reliability.

• Leader Harmonious Drive Systems (688017.SS)

  The report rates the company OW and lists it as a preferred leading robotics component company.

  Strengths

  The report believes that once deployment begins, reliability, consistency, yield, mass-production scale and cost control will raise barriers for component suppliers, benefiting industry leaders.

  Comparison

  The company is included among the report's preferred component names alongside Hengli Hydraulic and Shuanghuan Driveline.

• Hengli Hydraulic (601100.SS)

  The report rates the company OW and lists it as a preferred leading robotics component company.

  Strengths

  Potential upside scenarios include stronger-than-expected demand for excavators, pumps and valves; entry into more overseas brands' supply chains; and higher humanoid robot penetration and supply-chain share.

  Weaknesses

  The pump and valve business has passed its hypergrowth phase, and the report believes the domestic sales peak in this cycle may be weaker than in the previous cycle.

  Comparison

  The non-humanoid robotics business uses a target P/E of 30x for 2027, below the approximately 40–50x peak in the previous cycle. The humanoid robotics business uses a DCF with an 11% WACC and 4% perpetual growth rate.

• Shuanghuan Driveline (002472.SZ)

  The report rates the company OW and lists it as a preferred leading robotics component company.

  Strengths

  The report believes that as the industry shifts from product certification to reliable mass production, leading component suppliers with manufacturing scale, yield and cost-control capabilities will be better positioned.

  Comparison

  The company is included among the report's preferred names alongside Leader Harmonious Drive Systems and Hengli Hydraulic.

Key data

• Average Number of Customers per IntegratorDozensUp from approximately 10 at the beginning of 2026, but individual customers generally still deploy only dozens of robots.
• Galaxea's Logistics CapabilitiesApproximately 10,000 SKUs; 70–80% of human speedCan handle unfamiliar objects across fulfillment, sorting, tote handling, picking and packaging.
• UBTECH Depalletizing Productivity60–80% of the required production cycle timeIndustrial deliveries in 2026 are focused mainly on depalletizing and material handling.
• Mech-Mind Shipments and GrowthApproximately 10,000 systems in 2025; approximately +70% YoY revenue growth in 1Q26Order growth was similar in the first half of 2026.
• Deep Robotics Industrial-Grade Quadruped Robot TargetApproximately 10,000 units in 2026Compared with approximately 3,000 units in 2025; grew by more than 100% YoY in the first half of 2026.
• Major Companies' Shipment TargetsGeekplus: cumulative 10,000 units over the next three years; UBTECH: approximately 10,000 units in 2026 and 20,000–40,000 units in 2027Approximately two-thirds of Geekplus's target comprises Gino robots and one-third robotic arm workstations.
• Dexterous Hand Price RangesLinkage: RMB3,000–6,000; direct drive: RMB20,000–50,000; tendon driven: up to RMB90,000–100,000Higher degrees of freedom and greater flexibility generally entail higher cost, thermal load and integration difficulty.
• Dexterous Hand Reliability TestingMore than 1 million laboratory grasps; longest continuous powered operation under six monthsContinuous industrial operation still faces issues involving calibration, accuracy, overheating and transmission-component service life.
• New-Task Adaptation CycleReduced from the historical 6–12 months to dozens of hours of incremental data for simple tasksFoundation models and end-to-end data pipelines can accelerate vertical-model development and transfer across use cases.
• Expected Embodied-Intelligence “GPT Moment”In 18–24 months, around 2028Several companies expect this timing, but the report considers the timeline optimistic.
• Differences in Data Usability RatesApproximately 80% for high-quality internal teleoperation data; below 10% for low-cost outsourced egocentric dataData yield may be more informative than raw data hours.
• UBTECH Data TargetApproximately 1 million hours of usable data in 2026May require approximately 4 million hours of raw data; currently approximately 300,000 hours.
• Data Collection CostsOutsourced egocentric data as low as approximately US$2/hour; high-quality multimodal hand data approximately US$90–100/hourValue may shift toward task design, synchronization, quality control, cleaning and model-linked selection.
• Gap Between Domestic and Overseas Training ChipsApproximately four domestic chips equivalent to one Nvidia chipThis is an estimate of the performance gap provided by some companies; integrators still rely mainly on overseas chips.
• Lingyi iTech Module Standardization Cost Reduction10–20% reduction in duplicative design and development costs; further 10–15% reduction in manufacturing costsDerived respectively from standardization of key modules and simplification of complex structures.
• Geekplus Valuation5.6x 2026E P/SBased on a 50/50 weighting of an 8.3x peer valuation discounted by 20% and a 4.6x valuation for logistics and warehouse automation peers.
• Hengli Hydraulic Valuation Assumptions30x 2027 P/E for the non-humanoid business; 11% WACC and 4% perpetual growth for the humanoid businessThe target P/E and DCF valuation methods are used, respectively.

Impact & implications

The report believes that assessments of value in the robotics industry are shifting from demonstrations of motion and announced capacity targets toward verification of productivity, repeat purchases, reliability, closed-loop data feedback and unit economics. Semi-structured logistics, industrial material handling, inspection and upgrades to conventional automation may scale earlier than general-purpose household humanoid robots. As customers begin demanding compensation for production-line downtime, the quality, consistency, yield, durability and mass-production cost of complete robots and components will become higher barriers. The report therefore favors leading component suppliers with reliable mass-production capabilities. Model architecture and training compute may not independently constitute durable moats; the rights, quality and diversity of deployment data and the speed of feedback are more likely to determine long-term competitiveness.

Risks

• Individual customers generally still deploy only dozens of robots, leaving the industry far from large-scale adoption.
• Speed and accuracy on complex or high-precision tasks remain materially below human levels, and insufficient productivity may delay labor substitution.
• The continuous operating reliability of dexterous hands, tactile-data synchronization, cross-architecture data reuse, service life and high-DoF costs still do not meet the requirements for scaled industrial applications.
• On-site customer data may be unavailable for training because of ownership and security requirements, limiting closed-loop deployment feedback.
• A significant performance gap remains between domestic and overseas training chips, and integrators currently rely mainly on overseas chips.
• Battery life, joint thermal constraints, calibration drift and system failures may cause production-line downtime, with the resulting economic losses potentially exceeding the value of the robot itself.
• The commercial value, efficiency and safety issues of household humanoid robots will take time to resolve.
• Lingyi iTech faces risks that AI infrastructure and customer qualification proceed more slowly than expected, smartphone and foldable-screen demand remains weak, and new businesses such as AI glasses and XR are delayed.
• Geekplus faces risks from the expiration of lock-up periods, geopolitical impacts on orders, higher-than-expected investment in embodied intelligence, and deteriorating robotics and AI market sentiment.

What to watch

• Track whether deployment volumes per customer can expand beyond dozens of units and whether proof-of-concept projects can continue converting into repeat purchases.
• Monitor robot productivity, commercial ramp-up, and speed and accuracy on complex tasks in real-world scenarios.
• Track whether adaptation time for new tasks and the amount of data required for each new task continue to decline.
• Verify whether zero-shot success rates on unfamiliar tasks can reach 70–80% and approach production performance after limited post-training.
• Observe whether data yield, scenario diversity and closed-loop deployment feedback can translate million-hour-scale datasets into stronger generalization.
• Track continuous operating time, failure probability, calibration drift, recovery time, thermal management, battery life and end-effector durability.
• Monitor progress in testing domestic training chips and the potential deployment of UBTECH's dedicated humanoid robot edge chip over the next 2–3 years.
• Track whether companies can deliver on their shipment, revenue, capacity, data and compute targets for 2026–2028.