PokeBot Secures 9-Figure Pre-A: Embodied AI's Battleground Shifts from 'Walking' to 'Manipulating'
Category: industry · Embodied AI / robot manipulation Published: 2026-08-03 08:40 (Beijing Time) Source: AI Tech Review exclusive; public materials from Tsinghua TEA Lab and Huazhe Xu
Where the Money Goes
According to AI Tech Review, PokeBot (破壳机器人), founded in April 2026, has completed a 9-figure-level Pre-A round. Shunwei Capital and Matrix Partners China led the round, with participation from Zhongding Capital, Ubiquant Ventures, Junshan Capital, SEE Fund, Liepin Investment, and Yuannuo Capital. Existing investors — including Yunqi Capital, Xiaomi Strategic Investment, Honghui Fund, Innoangel Fund, and Eastern Jiafu — also increased their stakes.
The report did not disclose the exact amount, post-money valuation, or closing documents, so the accurate description remains a "9-figure-level Pre-A" rather than any specific dollar figure.
Why It Keeps Emphasizing 'Manipulation'
Robot navigation and locomotion already have mature engineering pathways. But at the kitchen counter, problems suddenly get hard: cups slip, tofu crumbles, states inside the pot change, and knives and seasonings must be switched constantly. One wrong action can cascade into dozens of subsequent failures.
PokeBot's publicly stated roadmap is WAM × RL × DATA:
- WAM (World Action Model): predicts how actions will change the environment and helps plan the next step;
- RL (full-pipeline real-robot reinforcement learning): lets the robot adjust based on rewards and failures during actual physical interaction;
- DATA: real-world household data captured via proprietary collection devices, feeding back into the model and RL loop.
- https://www.163.com/dy/article/L3D5EC2I0511DPVD.html
- http://hxu.rocks/
- https://new.qq.com/rain/a/20260622A0ATAF00?refer=cp_1009
- https://c.m.163.com/news/a/L2GTC5U40511C4AA.html
The most striking public demo is a roughly 9-minute fully autonomous mapo tofu cooking video. The report breaks down the challenges as "long, dexterous, dynamic, cluttered, precise": long-horizon steps, deformable ingredients, real-time state changes, multi-tool coordination, and millimeter-level placement — precisely the hardest parts to standardize in home environments.
My Take
The real value of this news is that it pulls embodied AI back from "looking like the right action" to "controlling the consequences of the action." Vision-language-action models let a robot understand "go grab the cup," but they don't automatically ensure it knows what happens when the cup receives force, how to recover from a bad grasp, or how to re-plan mid-task.
That said, a single demo cannot prove generality. When household items, lighting, counter heights, or ingredient states change, success rate, recovery time, and human takeover rate are the hard metrics. PokeBot's next step should be publishing statistics on cross-household, cross-object, long-duration operation — not more beautifully edited demos.
Sources and evidence: