He left a 20 billion yuan unicorn. Five months later, he raised a nine-figure dollar round.
On 3 August, PokeBot, a general-purpose embodied AI company, announced a Pre-A round at the $100 million plus level. Shunshi Capital and Matrix Partners China co-led. Zhongding Capital, Jiukun Ventures, SEE Fund and others participated. Xiaomi’s strategic investment arm and Yunqi Capital followed on.
PokeBot was founded in March 2026. It is a very young company. But its founder, Xu Huazhe, is no newcomer.
Xu was previously co-founder and chief scientist of Xinghaitu, a star embodied AI startup. He left when Xinghaitu’s valuation had crossed 20 billion yuan and its cumulative funding had reached 3 billion yuan.
The mapo tofu demo
In June, a nine-minute robot demonstration video put PokeBot on the map. A robot cooked mapo tofu from start to finish. It cut the tofu, stir-fried the mince, seasoned the dish and plated it, with no human intervention.
Tofu is fragile and requires precise force control. Ingredients change state continuously. The robot had to switch between knife and spatula, then return each tool to its place. PokeBot also showed it could fold clothes, tie pouches and thread cable ties, covering flexible object handling, fine manipulation and long-horizon tasks.
Why home, not factory
While most humanoid companies target factories and warehouses, PokeBot has aimed at the home from day one. Xu’s reasoning: factory tasks are highly standardised. Home tasks are not. Objects move, spaces change, and nothing repeats. If a robot can handle that, it has solved general manipulation.
A founder with credentials
Xu holds an undergraduate degree from Tsinghua’s Electronic Engineering department, a PhD from UC Berkeley’s BAIR lab, and completed a postdoc at Stanford’s SVL. He returned to Tsinghua as an assistant professor and founded TEA Lab, China’s first laboratory focused on robot manipulation.
He has published more than 100 papers in Science Robotics, NeurIPS, ICLR, CoRL and ICRA. He won the CoRL’23 Best System Paper and received Best Paper nominations at RSS’25 and ICRA’26.
The technology
PokeBot’s system, called WAM x RL x DATA, combines a world action model, full-pipeline reinforcement learning on physical robots, and proprietary data collection. The UAG parallel training architecture claims to preserve generalisation while boosting training efficiency fivefold.
Former Meituan drone executive Liu Shuo has joined as PokeBot’s first business partner, leading commercialisation.
The market
According to Fortune Business Insights, the global home robot market was worth $13.9 billion in 2025 and is projected to reach $107.15 billion by 2034.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience.