PokeBot's home-robot bet: Tsinghua prof leaves Galaxea, raises $100M+ in months

PokeBot is a Shanghai-based general embodied-intelligence company founded in early 2026 that develops household robots capable of long-horizon manipulation: autonomously cooking, folding clothes, passing and tidying objects. Instead of the industry-standard VLA base model, it builds a video-action world model (the UAG architecture) trained with reinforcement learning across pre-training and deployment, wrapped in a WAM x RL x DATA flywheel of world-action models, real-robot RL and real-world data collection.

Founder Xu Huazhe bets homes, not factories, breed general AI: robots trained end-to-end on real-world chores become the path to AGI.

An angel round in spring 2026, then a Pre-A at the hundred-million-US-dollar level announced 2026-08-03, co-led by Shunwei Capital and Matrix Partners China with existing backers including Yunqi Capital and Xiaomi Strategic Investment increasing their stakes.

Xu Huazhe — a Tsinghua electrical-engineering graduate with a Berkeley PhD, Stanford postdoc and Tsinghua faculty post — co-founded Galaxea AI in 2023 and served as chief scientist before leaving at the company's high valuation. His stated reason: most robot makers were putting humanoids in factories to do what traditional arms already do, 'essentially using new humanoids for old-era work,' while 'true AGI should be born in the household.' He founded PokeBot in March 2026 with roughly 20 core members from his Tsinghua embodied-intelligence lab.

In July 2026 the company hired Liu Shuo, former head of Meituan's drone business commercialization, as its first commercial partner — read by press as the shift from research to market. That same month it announced the Pre-A led by Shunwei and Matrix Partners China. Technically, PokeBot abandoned the mainstream VLA foundation-model route for the UAG architecture, which it says replaces cascade pre-training with parallel pre-training and achieves five times the training efficiency, with reinforcement learning running through both pre-training and deployment.

As of 2026-09-06 PokeBot is pre-product but well-capitalized: two rounds closed within five months of a March 2026 founding, a 32B world model in its first training cycle, a June 2026 cooking demo publicized globally, and a first commercial partner hired. No consumer robot has been shipped or priced.

Home environments are messier and more random than factories, which Xu argues makes them the better training soil for general models — difficulty as data strategy.

Founder credibility concentrated capital fast: a Tsinghua professor with a Berkeley PhD and a Galaxea track record drew Shunwei, Matrix and Xiaomi Strategic Investment within months of founding.

The technical divergence (world model over VLA, RL through deployment) is a bet on a different learning curve, with five-times training efficiency claimed for the UAG architecture.

Starting with frequent, simple household tasks (passing and tidying objects) gives a nearer product wedge than full home automation, while demos like the mapo tofu task build the narrative moat.

Pick the environment by what it teaches, not by ease of deployment: Xu argues homes train more general models than factories, accepting a harder product path because the learning signal is richer.

As of 2026-09-06 PokeBot is building with strong momentum: two rounds within five months, a 32B world model in first training, a publicized June 2026 cooking demo and a commercial partner hired in July. Risks decide all embodied-AI companies: whether real-world data and RL generalize to unscripted homes, whether the non-VLA world-model bet beats better-funded rivals, and whether a consumer robot reaches useful reliability at a price households pay. Galaxea, Unitree and other well-capitalized peers push the same frontier, so PokeBot will be judged on shipped home capability, not funding.