宇树王兴兴把物理AI突破预期缩短至两三年内，并宣布世界物理AI大赛

NextFin News — Wang Xingxing is moving his timeline. In August, speaking the day after Unitree’s STAR Market debut, he placed a true “ChatGPT moment” for embodied machines somewhere between a few years and a decade—defined as a robot that could enter an unfamiliar home and finish most daily tasks on voice or text alone. In Hangzhou on October 22, at the World Physical AI Conference and Unitree partner summit, he struck a sharper note: physical AI already shows signs of being “broken through,” with a decisive step possible in one to two years, or two to three at the outside.

The bottleneck he named did not change. Models must align what they predict with what motors, fingers and imperfect sensors actually do in the real world. That alignment problem—inputs and outputs that drift by centimeters and milliseconds—still separates choreography from chores. What shifted was his reading of the curve. Frontier software labs, he said, have in some internal sense already reached model self-evolution: systems that write code, gather data and improve with only light human steering. The lag the public feels is diffusion, which he still put at five to ten years from lab breakthrough to everyday use.

Unitree is trying to own more of that stack for robots. Wang said the company is pre-researching a “self-evolution V1.0” large model for physical AI and pushing smaller models fitted to its own hardware. Beside software, he listed two industrial bets: raising mass-production capacity toward the volumes a true mass market would demand—far above the roughly ten to twenty thousand G1 units he cited as already built—and building the compute base for a closed loop of autonomous development, simulation, real-world test and feedback. In his framing, the scarce inputs of the next era are data, compute, chip supply, energy and minerals as much as clever demos.

Deployment order matters as much as model version. Commercial service roles, he suggested, will likely move faster than either the household or the heavy industrial floor. That hierarchy matches where humanoid shipments have already concentrated: structured logistics and factory cells, not living rooms. Aspirational language about homes in 2027–2028 sits beside that narrower near-term path.

The conference also gave the claim a public instrument. Organizers announced a World Physical AI Competition open to teams and individuals worldwide, with annual prize money in the tens of millions of yuan and eight tracks—among them free movement, athletics, combat, dexterous manipulation, production, sports events, multimodal interaction and elder care. Contests do not invent generalization. They do force comparable tasks, attract outside talent and turn supplier days into a recruitment funnel.

Read against the August speech, the October message is less a contradiction than a recalibration under listing-season scrutiny. Then, the priority was cooling fantasies that Unitree could already replace labor at scale. Now, with partners in the room and a competition on the calendar, the priority is to argue that the technical wall is cracking—and that Unitree intends to hold the hardware, the model loop and the ecosystem prizes when it does. Investors and factory buyers will still ask the same rude questions: success rates when the scene shifts, cost per completed task, and whether “self-evolution” shows up in shipped software or only in keynote nouns.

Breakthrough talk is cheap until a robot finishes the job in someone else’s building. Wang has shortened the clock he is willing to put on that test. The next two or three years will show whether the crack he sees becomes a door.

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App