China's robot chiefs say big models are still in the 'dial-up era'
At the World Robot Conference in Beijing, the founders actually shipping robots — UBTech, Robot Era, Booster Robotics, Jushen Technology and JD's robotics arm among them — delivered a rare public reality check: embodied AI's bottleneck is no longer demos, it's reliability, manufacturing standards, and cold ROI math. The consensus: chase the boring, high-demand industrial jobs first.
Robotics veteran Wang Tianmiao put the question bluntly at a WRC 2026 roundtable: if large models are still in their "dial-up internet" era, what exactly is blocking robots from real orders? The Beihang professor framed 2026 as the industry's "year one of scaled deployment," yet China's humanoid deliveries this year total only around 15,000 units — while traditional industrial robots keep growing at roughly 20% with exports already past ten billion yuan a year. His three questions to five embodied-AI executives drew a line between viral backflip videos and what actually survives a factory floor.
The first fault line was whether the blocker is the brain or the assembly line — and the room split cleanly. On the tech side, Robot Era CEO Zhao Tongyang argued that the roughly 96% task success rates labs love to cite are disqualifying in the physical world: out of 100 tasks, four failures can mean a destroyed production line or a robot that doesn't come back, so To-B buyers demand two more nines of reliability before "ten-thousand-unit delivery" means anything. Booster Robotics CEO Cheng Hao agreed that no end-to-end model has yet survived engineering-grade testing, calling today's fragmented approaches — hierarchical stacks versus VLA fusion — far from converged.
But the operators who have shipped volume disagreed. Jushen Technology co-founder Liu Yulong said his company has mass-produced about 15,000 robots as of June — including quadrupeds for inspection and firefighting — and in narrow scenarios like moving goods from A to B, "good enough" brains already work; his real pain is yield consistency, batch quality control, and the absence of mature industry standards for things like load-to-weight ratio and mean time between failures. UBTech VP Jiao Jichao, whose company is ramping toward ten-thousand-unit capacity this year, added that customers never ask for perfect success rates — they ask whether replacing a workstation pays for itself within two to three years. And JD's Xu Lei disclosed that conversion rates for embodied robots run dozens of times worse than consumer electronics, because too much of what's on sale is spectacle rather than ROI-positive utility.
On reliability itself, the panel converged on data and architecture. Traditional industrial arms run 60,000–80,000 hours between failures; a probabilistic large model nowhere near that bar simply cannot be trusted where tolerance is zero. Jiao pointed to a hidden culprit: cross-embodiment transfer — models trained on one robot body break on another because motor parameters and sensor data shift — plus a shortage of real touch, force, and vision data that internet-pretrained Transformers were never built to close. He expects an entirely new embodied-native architecture to emerge. Cheng Hao's counter was patience: dial-up had commercial value before broadband arrived, and embodied AI will climb the same gradual slope — which is why the panel's closing advice was to start in high-demand, semi-structured business niches like inspection, material handling, and sorting before ever touching homes.
New Zealand will require social platforms to verify users' ages or face fines of up to 10% of global revenue, under a bill announced by the government. Per Bloomberg, "high risk" platforms would have to check whether users are under 16 — or pay penalties scaled to worldwide revenue. The bill follows Australia's landmark under-16 ban, which already forced Meta to remove hundreds of thousands of teen accounts, and lands amid a global scramble over kids' social feeds — France's top court recently blocked an under-15 ban on free-expression grounds.
What to watch: whether Wellington's revenue-scaled penalty — borrowed from the EU playbook — proves more effective than Australia's fine-based regime at forcing platforms to build real age verification.
Do you think a 10%-of-global-revenue threat will finally make platforms take age verification seriously — or just produce another round of checkbox compliance? Tell us in the comments.
