Humanoid Robots Move From Dance to Work, but Generalization Gap Remains

Past years, humanoid robots at exhibitions and galas mainly sang, danced, boxed, greeted guests, and made coffee. Some online commenters dismissed them as remote-controlled toys.

Unitree founder Wang Xingxing responded that performance and locomotion are easier commercial scenarios, but Unitree aims for robots to do more valuable work. At the August 2026 World Robot Conference, the focus shifted.

Unitree showed a nearly 3-meter-tall manned mech; Zhicheng robots did household chores; Zibianliang robot folded clothes and watered flowers; and Hikvision's robot wore an apron to simulate a fast-food worker taking orders and placing meal boxes. Hu Luhui, founder of Zhicheng AI, said the most direct impression was a clear change: robots focused on doing things, completing tasks, and improving real productivity.

Turning robots into all-purpose butlers remains difficult. In his WRC speech, Wang said the industry's biggest bottleneck is insufficient generalization in embodied intelligence.

Generalization means a robot cannot only handle the standard answer. If trained to see plate A and put it in the dishwasher, it may perform beautifully. But if given plate B, or the dishwasher moves, kitchen lighting changes, or an extra cup appears, it may suddenly fail.

Hu breaks the problem into embodiment, task, and environment generalization. Robots still lack true understanding of the physical world. The bottleneck is insufficient alignment between AI model inputs and outputs and the robot itself. It looks like a small gap, but it is a huge divide between being able to do something and doing it well.