At the World Robot Conference 2026 main forum, five executives from China's leading embodied-AI companies gave a strikingly honest answer to the industry's loudest question: the barrier to mass production is not model capability, it is delivery — getting from one robot to 10,000 units. With UBTech vowing to build 10,000 industrial humanoids this year and Unitree announcing 18,000 cumulative humanoids off the line, the panel named the real walls: brain reliability that must climb from 99% to 99.99%, batch consistency that still has no industry standard, and products that have not yet found a proven demand match.
The Bottleneck Is Delivery, Not Demos
WRC 2026 was a show of strength — crowds like a subway rush hour, robots everywhere. But underneath, the industry is quietly repositioning. Wheeled dual-arm makers are building humanoids (Galaxy General's ET1 "Xingzai"); humanoid makers are building wheeled dual-arm robots (UBTech's Cruzr Y1 and S2, demonstrated moving boxes and palletizing); world models have become the phrase in every forum; and Fourier showed a centaur-style robot for stability-first deployments. These shifts point to one question the industry must answer this year: how do you actually mass-produce embodied robots?
For context, many in the industry treat one million units as the true mass-production bar. Last year's fewer-than-20,000 humanoids was "mass production" only in quotation marks. This year UBTech set a 10,000-unit industrial humanoid target, and freshly listed Unitree announced 18,000 cumulative humanoids produced. The WRC panel — Zhao Tongyang (CEO, Zhongqing Robotics), Cheng Hao (founder and chairman, Jiasu Jinhua), Liu Yulong (co-founder and president, Zhishen Technology), Jiao Jichao (VP of UBTech, head of its Embodied Intelligence and Humanoid Robotics Research Institute) and Xu Lei (head of JD Smart Robotics) — spent the session explaining why the gap between that ambition and reality is so wide.
The Five Walls Between One Robot and 10,000
1. Brain reliability: from 99% to 99.99%. Zhao Tongyang put the math bluntly: at a 99% task success rate, one out of every 100 missions fails — and in the real world, a robot that does not come back is unacceptable. "The brain needs to go from 99% to 99.99% before robots can truly enter the real world," he said. Cheng Hao agreed that the embodied large model has not converged on a technical route and is far from engineering maturity, even though hardware consistency (hundreds of robots moving in sync) and "cerebellum" skills (locomotion control, vision-motor coordination, whole-body control) have made real progress.
2. Hardware specs and batch consistency. Liu Yulong said the first wall in actual delivery is hardware: sufficient payload within limited self-weight, tolerance for high and low temperatures, sustained operating hours — all must clear the customer's bottom line. The second wall is batch consistency: 10,000 units means consistent yield, process, evaluation and final delivery state across batches. Unlike cars, drones and home appliances, embodied robots still lack clear, complete industry standards, so companies must build them jointly with upstream and downstream partners.
3. Manufacturing and the two-to-three-year payback. Jiao Jichao emphasized that 10,000-unit production challenges the production line itself: efficiency cannot grow by adding people. From R&D to NPI to batch procurement and quality control, the industry needs standardized processes — an engineering change found at unit 100 or 200 must flow into the factory quickly or consistency suffers. And industrial customers ultimately compute one number: can the robot replacing a workstation pay back its BOM, delivery and maintenance costs within two to three years?
4. Demand match. Xu Lei pointed to sales data showing embodied products convert far worse than consumer electronics and phones — product and demand have not converged. "Once you find the true demand match, delivery of 10,000-plus units will not be the biggest problem," he said.
Is the Endgame Winner-Take-All?
The panel also debated the endgame. Will embodied AI be winner-take-all? Xu Lei argued that today's diversity (bodies, brains, models, full-stack players) will converge as the industry scales, because standards are what enable scale — but long-term segmentation by scenario and region will keep many companies alive. Zhao Tongyang expects concentration: "Robots fuse large models, hardware, ecosystems and software; eventually one or two Apple- or Microsoft-like platform companies will emerge," just as smartphones concentrated profits in a few hands. Jiao Jichao sees concentration at the supply-chain level (motors, reducers, embodied-AI chips) and in hardware platforms, with a rich application layer on top — like the smartphone model, except embodied robots couple hardware and software so deeply that body makers may need full-stack, platform-level capability. Liu Yulong counters that education, industrial, commercial and consumer markets will each produce their own leaders.
Cheng Hao offered the cleanest frame: three phases — hardware, agent, model. We are in the hardware phase, so products bloom; in the agent phase, bodies and systems converge into a few platforms with increasingly rich applications; in the model phase, embodied large models absorb applications and applications themselves consolidate. General and specialized devices run in parallel — the smartphone did not kill the camera.
On the brain-body question, the panel split along the same lines: Jiao Jichao argues data is deeply body-specific (motor parameters and configurations differ, so models do not transfer across bodies) and deep hardware-software matching is the more effective route; Xu Lei and Cheng Hao expect eventual decoupling — Cheng notes that even BASIC differed across computers in the late 1970s. Zhao Tongyang's verdict: if you want to be Apple, you combine body and soul.
What the Embodied Large Model Is Missing
What does the embodied large model most need? Liu Yulong: real data, in both quantity and quality — the gap between the virtual brain and the physical body is large, and effective real-world data is scarce; the problem grows for humanoids that must move, manipulate and run long-horizon tasks. Recent progress such as GEN-1.5, a robot foundation model that learns tasks in seconds, shows capability gains — but data and generalization remain the hard constraints. Jiao Jichao: data categories (tactile and force data are nowhere near text and image abundance) and model architecture (current multimodal architectures migrated from NLP may not fit embodied data). Zhao Tongyang: generalization — you cannot train a separate model for every customer request; if you try, you become a tool maker, not a general robot company. Cheng Hao's framing: "We are in the dial-up era while users expect 5G," but the brain does not need to be finished for robots to sell — dial-up computers sold in huge volumes.
What to Do Now: Start With Boring Work
The panel's concrete advice for companies: pick the easy tasks first. Zhao Tongyang listed three entry points — patrol, security and property management (walking and vision are already good enough); industrial dirty, heavy or dangerous work (risk-bearing and handling tasks); and ecosystem plays (open bodies and underlying capabilities so third parties build applications — a few hundred engineers cannot cover every industry). Home remains the late market. Jiao Jichao expects industrial demand to scale next, since industrial customers are urgent and relatively price-insensitive; he also sees a real market in emotional companionship that goes beyond dialogue into actions and micro-expressions. Liu Yulong's path: research, then entertainment, then quadruped training, patrol, emergency and firefighting — dangerous, repetitive, tiring jobs that are often non-negotiable needs — then industrial and commercial B2B, and finally consumer scale. Cheng Hao's rule: both ToB and ToC start from simple tasks and work up — the industry's shift from WAM to WTM metrics is the same coin as demand matching; the winners are those who do simple tasks thoroughly first.
The takeaway: there is no moment when all technology is ready. Brain, hardware, production, standards, cost and scenario demand pull each other forward. The companies that win will find the tasks robots can do reliably today that create real value, build a data loop through real delivery, and climb into harder tasks from there. Ten-thousand-unit delivery is only the starting line — the harder question is whether those robots actually get used.
Frequently Asked Questions
Q: Why is a 99% task success rate not enough for robots?
A: At 99%, one out of every 100 missions fails — in real deployments a robot that fails to come back is unacceptable. Industry leaders at WRC 2026 say reliability must reach 99.99% before robots can broadly enter the real world, and that gap is currently the biggest bottleneck to 10,000-unit delivery.
Q: Which robot scenarios will scale first?
A: The panel's consensus: patrol and security, industrial dirty/heavy/dangerous work, and B2B scenarios before consumer. Industrial customers are urgent and less price-sensitive, and a robot replacing a workstation must pay back within two to three years. Home service ranks as the latest market.
Q: Will embodied AI end in winner-take-all?
A: Probably platform concentration, not single-player victory. Expect one or two Apple- or Microsoft-like platform companies, concentration in supply-chain components (motors, reducers, embodied chips), a rich application layer, and a three-phase evolution from hardware to agent to model.
