A self-taught inventor who filmed himself building a grape-stitching robot arm on Bilibili is now running one of China's most watched embodied-AI (具身智能) companies. The gap between those two facts is smaller than it looks: both are about turning algorithms into things that move.
Peng Zhimin (彭志辉), online handle Zhi Hui Jun (稚晖君), is the co-founder, president and CTO of AgiBot (智元机器人).
The make-things background
Peng was born in 1993 in Jiangxi. He earned a bachelor's degree in 2015 from the University of Electronic Science and Technology of China (UESTC), College of Life Sciences and Technology, and a master's in 2018 from UESTC's School of Information and Communication Engineering.
His public reputation came from hardware videos — an autonomous bicycle, a tiny Linux "TV," a robotic arm that stitched a grape's skin. Before robotics he was an algorithm engineer at OPPO's research institute AI lab.
Huawei's Genius Youth
In 2020 he joined Huawei through the Huawei Genius Youth Program (华为天才少年计划) at its top salary tier, reported at about 2.01 million yuan (≈ US$280,000 / HK$2.2M). There he worked on Ascend (昇腾) AI chips and AI algorithms.
In late 2022 he left Huawei. In February 2023 he co-founded AgiBot with Deng Taihua (邓泰华), a former Huawei executive.
From one robot to a production line
The first embodied robot, Yuanzheng A1 (远征A1), appeared in August 2023. From there the company moved fast:
- December 2024: AgiBot began mass production of general-purpose embodied robots.
- January 2025: the 1,000th general embodied robot rolled off the line — reported as a record for the category in China.
- March 2025: it released GO-1 (Genie Operator-1), a general embodied foundation model built on a Vision-Language-Latent-Action (ViLLA) architecture — a VLM plus a mixture-of-experts (MoE) design.
GO-1's pitch is practical: learn from human video, generalize from few samples, run on different robot bodies ("one brain, many forms"), and improve from a data-feedback loop. AgiBot said the "pour water" task dropped from tens of thousands of collected samples to about a thousand once the model was attached.
Open data, closed loop
In December 2024 AgiBot open-sourced AgiBot World, a real-robot dataset described as over 1 million trajectories across 217 tasks and five scenarios. A broad data pool plus a foundation model is the same flywheel ByteDance uses for Doubao — just for arms and legs instead of text.
Peng also sits as vice-chair of the Ministry of Industry and Information Technology's Humanoid Robot Standardization Technical Committee, and in 2026 took a listed-company board chair role — a sign the embodied-AI sector is colliding with capital markets.
Why he is a builder worth watching
Peng is unusual because the hardware fluency and the model strategy come from the same person. Most robotics firms buy or borrow their "brain"; AgiBot trains its own. That vertical stack — body, data, model — is exactly what the 具身智能 (embodied AI) race is about.
The technical bet behind the story
Peng's path matters because it collapses three roles — hardware designer, model trainer, and company builder — into one person who can hold the whole stack in his head. The GO-1 claim, that a foundation model cuts sample-collection needs by an order of magnitude, is the kind of result that decides whether humanoids scale past demo videos. If "learn from human video, generalize from few samples" actually holds across tasks, the cost of teaching a new skill drops from weeks of teleoperation to a day of data.
What to watch next
The open question is whether AgiBot's vertical stack stays ahead of rivals who split the work — Unitree on the body, a lab on the brain. Peng's advantage is integration speed; the risk is that a specialized brain vendor out-runs an integrated one. Either way, the MIIT committee seat means the rules for the whole category will bear his fingerprints, so the个人 (individual) and the institution pull in the same direction.
Honest limitations
Biographical facts (UESTC, OPPO, Huawei Genius Youth, AgiBot founding, GO-1, 1,000-robot milestone) are corroborated by China News Service, Securities Times and China Securities Journal reporting. The salary figure is widely reported but is a media-reported tier, not a contract we can verify. Production and "first mass-produced" claims come from the company and Chinese financial press; independent unit-shipment audits are not public. The technical claims about GO-1 (sample efficiency, cross-body transfer) are the company's stated results, not independently benchmarked here.
What readers can do now
- If you follow embodied AI, study AgiBot World as an open real-robot dataset — it is a rare public asset for manipulation research.
- Compare GO-1's "one brain, many forms" claim against Unitree and UBTECH roadmaps to see which vertical-stack bet wins.
- Watch the MIIT humanoid-robot standardization committee — Peng's vice-chair seat signals where China's robot safety and interface rules will be written.