Unitree's Wang Xingxing names AI-to-robot alignment as main hurdle for embodied AI
In short
Speaking at the 2026 World Physical AI Conference in Hangzhou on 22 October, Unitree's Wang Xingxing said the biggest challenge for embodied intelligence to pass its "GPT moment" is aligning AI models' inputs and outputs with real robots. He expects robots could follow instructions to tidy a room by the end of this year or next year, if things go fast.
Read the full story 1 min read
Wang Xingxing, who leads Unitree (宇树科技), said at the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference that the biggest challenge for embodied intelligence to pass its GPT moment is aligning the inputs and outputs of AI models with real robots. The conference was held in Hangzhou on 22 October, according to Yicai and Jiemian. [ 1 , 2 , 3 ]
Jiemian and 36Kr reported that Wang repeated his definition of the GPT moment for embodied intelligence: robots completing about 80% of main tasks in about 80% of unfamiliar scenes, from voice or text instructions. They said he believes large models are already strong at robot planning and navigation. He expects that, if things go fast, robots could follow instructions to tidy a room by the end of this year or next year. [ 2 , 3 ]
Yicai, citing Securities Times, reported other parts of his keynote. Wang said that, in a sense, top large-model companies have basically achieved self-evolution of models internally, with models writing programs and collecting data and people taking part with appropriate guidance. He said the effect is not widely felt because it takes 5 to 10 years for a technical breakthrough to reach everyone, and that signs of a breakthrough in physical AI are already visible. Yicai described the challenge as passing the "ChatGPT moment", while Jiemian and 36Kr used "GPT moment". [ 1 , 2 , 3 ]
Why it matters
The remarks come from the head of a robot maker at the company's own partner conference, so they show how Unitree frames progress in embodied AI. The documents give no independent assessment of the forecast.
Key facts
- The 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference was held in Hangzhou on 22 October. [ 1 , 2 ]
- Wang Xingxing of Unitree (宇树科技) said the biggest challenge for embodied intelligence passing its GPT moment is aligning AI model inputs and outputs with real robots. [ 1 , 2 , 3 ]
- Wang defined the GPT moment as robots completing about 80% of main tasks in about 80% of unfamiliar scenes from voice or text instructions. [ 2 , 3 ]
- Wang said large models are already strong at robot planning and navigation. [ 2 , 3 ]
- Wang expects robots could follow instructions to tidy a room as soon as the end of this year or next year. [ 2 , 3 ]
- Yicai quoted Wang as saying top large-model companies have basically achieved self-evolution of models internally, with models writing programs and collecting data and people guiding as needed. [ 1 ]
Confirmed by several sources
Still unclear
- Whether top large-model companies have basically achieved self-evolution of models internally. Only Yicai, citing Securities Times, reports Wang's statement, and no document offers evidence for it.
- Whether robots will follow instructions to tidy rooms by the end of this year or next year. This is Wang's own forecast, made conditionally ("if fast"), and no document gives supporting evidence.
- Wang's exact titles. Yicai calls him chairman, general manager and chief technology officer; Jiemian calls him chairman, founder and CEO; 36Kr calls him founder.
- Wang described the GPT moment as robots completing about 80% of main tasks in about 80% of unfamiliar scenes via voice or text instructions, and expects room-tidying robots as soon as the end of this year or next year. Reported by a single source so far