The opening of this closed-door meeting was not a vision but a verification question. On the morning of September 17, six parallel closed-door sessions of the 2026 ITValue Summit Digital Value Annual Conference were held in Sanya, among them "Physical AI Enters Reality: Let Real Needs Meet Frontier Technology," co-organized by TMTPost and the Zhiyuan Research Institute, bringing the scenario side and the technology side to the same table. Moderator Wan Ning mentioned that across more than 40,000 hotels nationwide, small carts deliver meals and supplies every day, and the head of Home Inn told him delivery robots have become standard equipment in economy chain hotels—real demand happens every day, and someone is willing to pay for it.
But the people paying did the math first. Yang Hui, General Manager of Yuexiu Property's Digital Intelligence Development Center, placed one of the most mature cleaning robots on the market into her own residential underground garage: two levels, over 21,000 square meters of work area, mechanical parking spaces it cannot reach, circular ramps it cannot enter, actual operating area less than half, capable only of cleaning the flat lane surfaces, actual efficiency only 35% of theoretical value. From this she distilled four bottlenecks to scaled deployment: data cannot be assembled—a single skill requires thousands of hours; generalization does not pass—real scenarios cut performance in half; environments cannot be reused—changing sites means starting over; costs do not add up—three or four units to replace one person.
The answer can only grow out of engineering. Yuexiu iterated to its latest generation, and by restructuring scheduling, retrofitting elevator controls, optimizing routes, and connecting fault work orders, 80% of operations can be completed independently, only then reducing one person on site. Yang Hui emphasized this cannot be solved by swapping in a better machine, but requires systematic restructuring of scenario, body, algorithm, and operations, running "Yuexiu builds the scenario, algorithms train the model, robots execute, Yuexiu does acceptance" as a closed loop, replicating across garages after both ROI and utilization rate meet targets. Yuexiu has done sixteen years of digitalization, with over one hundred projects precipitating more than six thousand model master libraries—these computable spaces are precisely the venues prepared for robots.