AINEWS
At Beijing's World Robot Conference 2026, a Beihang professor asked five humanoid robot CEOs to explain the gap. They split into two camps, and ~15,000 cumulative units shipped is the proof.
At the World Robot Conference 2026 in Tianmiao. If large language models are still in the "dial-up internet" era, he asked, what is actually stopping humanoid robots from shipping at real-factory scale, and when will those models peak at solving physical-world problems? The five humanoid-robot CEOs on stage did not have one answer. They had two, and the room did not reconcile them.
The numbers behind the question are the reason it landed. Chinese humanoid-robot cumulative delivery in 2026 is roughly 15,000 units, Wang said in his opening. Traditional Chinese industrial-robot exports are already past 100 billion yuan (about $14 billion at recent exchange rates) and growing around 20% a year. The demo-to-delivery gap is not a vibe; it is a roughly 100-fold difference in installed base against an adjacent industry that has been quietly compounding for a decade. The panel's job was to explain why.
The first camp answered: the brain is not reliable enough. Today's embodied models succeed about 96% of the time on a task. "In 100 tasks, 4 times the robot cannot come back alive, or it will damage the customer's line," he said at the WRC 2026 roundtable. That is fine for a research demo. In a real factory, a one-in-twenty-five failure rate is the difference between a pilot and a contract. Humanoid robots will not cross the 10,000-unit-per-year line until the field pushes reliability from two nines (99%) to three or four nines (99.9% or 99.99%). That is a software problem before it is a hardware problem, and it is not close to solved.
Cheng Hao sharpened the claim. The demand is there, the product definition is not the issue, and the real choke point is that the technology is "not ready," he said, invoking the venture term TPMF, meaning technology–product–market fit, the stage where a product is technically capable but not yet a sustainable business. There are "many schools, not yet converged," he said, and no end-to-end large model has cleared engineering validation. For Cheng, the bottleneck is upstream of the factory floor: the field has not agreed on the right way to train a brain that has to touch the world.
The second camp answered: the brain is already good enough for the jobs that exist. 具身科技 had delivered roughly 15,000 robots by June 2026, almost all of them quadrupeds, the four-legged workhorses of warehouse and inspection fleets, not humanoids. In a narrow vertical scene, the brain only needs to be "good enough." The actual blockers, he said, are hardware metrics: payload-to-weight ratio, mean time between failures under temperature extremes, and cross-batch yield on the assembly line. There is no equivalent, in robotics, of the auto industry's standardized NPI, or new product introduction, the staged process that takes a design from prototype to mass production, and the field's quality-control systems are still being invented. Liu's bet: humanoid robots will not reach scale because their AI gets smarter, but because their manufacturing gets more boring.
Real customers, he said, care about return on investment inside two to three years, not about whether a robot succeeds 99.99% of the time in a lab. UBTech is approaching 10,000-unit annual capacity. The industry's missing piece is a standardized NPI process and quality-control system; the 100-to-200-to-10,000 step exposes every gap. JD's internal data shows embodied-robot conversion rates are "tens of times worse" than for traditional consumer electronics, he said.
Both camps agreed on one thing: the 15,000-unit 2026 number is the falsifier. If the algorithm-reliability camp is right, that figure will not move until a four-nines embodied model ships in a customer line. If the engineering-standards camp is right, it will move the moment a real NPI process and a credible ROI case land in a single vertical. Watch which companies spend 2027 publishing four-nines benchmarks, and which publish factory-yield curves. The two paths are now visible in the same room, and the WRC 2026 cohort will inherit the result.