NeuroAI NEUROAINEUROAI.SITE
ESC

Peng Zhihui's AgiBot bets on open-source humanoids and a robot data engine

How Huawei 'Genius Youth' Peng Zhihui built AgiBot around open hardware, a foundation model and a robot data factory.

2026-10-27 · 939 words · NeuroAI
Peng Zhihui's AgiBot bets on open-source humanoids and a robot data engine

A hardware maker who once built a self-balancing bicycle and a robot arm that stitches grape skin left a stable job at Huawei to start a company most people had never heard of. Within months, Peng Zhihui (彭志辉), known online as "Zhihui Jun," had a walking humanoid and a factory plan.

The company, AgiBot (智元机器人 / Zhiyuan Robotics), is now one of Shanghai's flagship embodied-intelligence (具身智能) startups. Peng, a former member of Huawei's selective "Genius Youth" (天才少年) program, has built AgiBot on three unconventional bets: open-source robots, a homegrown foundation model, and a dedicated data factory feeding it.

Key takeaways

  • Peng Zhihui founded AgiBot in Shanghai's Lingang area in February 2023, after leaving Huawei's "Genius Youth" program in December 2022.
  • The first humanoid, Yuanzheng A1, debuted on 18 August 2023: 1.75 m tall, 53 kg, 49 degrees of freedom, 7 km/h walking speed.
  • AgiBot open-sourced its Lingxi X1 humanoid in October 2024 and released the GO-1 (启元) embodied foundation model in March 2025.
  • It says it produced more than 10,000 robots by early 2026, after crossing 1,000 units in January 2025.
  • Backers include Tencent, BYD, HongShan and Hillhouse; press reports have valued it around 10 billion yuan (≈ US$1.4 billion).

From viral maker to company builder

Peng's reputation predates AgiBot. As "Zhihui Jun" he posted detailed build logs of self-driving bikes, miniature TVs and a robotic arm nicknamed "Dummy" that could sew grape skin — projects that drew millions of views and established him as a maker's maker. In 2020 he joined Huawei through its "Genius Youth" recruitment drive, working as an AI algorithm engineer on edge computing for the Ascend platform.

He left in December 2022 and, two months later, co-founded AgiBot with partners including a Shanghai Jiao Tong University professor. The speed was deliberate: rather than spend years on research papers, the team moved straight to a product launch.

A fast-moving robot family

AgiBot's hardware cadence has been unusually dense:

  • Yuanzheng A1 (Aug 2023) — the first bipedal humanoid, built for industrial tasks like bolt tightening and inspection.
  • Five-model expansion (Aug 2024) — the Yuanzheng A2, A2-Max and A2-W, plus the Lingxi X1 and X1-W, covering interaction, flexible manufacturing and developer use.
  • Lingxi X1 open-sourced (Oct 2024) — AgiBot released most design files and software, inviting developers to build on it.
  • Lingxi X2 (Mar 2025) — a smaller, nimbler humanoid shown riding scooters and hoverboards, with Peng overseeing its development.

The open-source move mirrors a belief that a robotics ecosystem, not a single locked product, is what scales adoption.

The data engine

Peng's core argument is that robots fail not for lack of hardware but for lack of data. AgiBot built a roughly 2,000-square-meter data-collection center in Shanghai where around 100 robots generate tens of thousands of motion records each day. Late in 2024 it open-sourced "AgiBot World," a manipulation dataset it says contains more than one million entries and exceeds Google's Open X-Embodiment in scale and real-world relevance for humanoid scenarios.

That data feeds GO-1 (启元大模型), the generalist embodied foundation model AgiBot released on 10 March 2025. Built on a Vision-Language-Latent-Action (ViLLA) architecture, GO-1 is designed to learn from human videos and generalize to new tasks with minimal specific training — the kind of capability a factory or warehouse actually needs.

Scale and the "358" plan

AgiBot says it began mass production in late 2024, reaching nearly 1,000 units by mid-December and crossing the 1,000-unit mark in January 2025. The company has stated it produced more than 10,000 robots by early 2026. Revenue has been reported to have climbed from about 300,000 yuan (≈ US$42,000) in 2023 to over 1 billion yuan (≈ US$140 million) in 2025, though those figures come from media summaries of company disclosures rather than audited filings.

Peng has publicly framed an ambitious roadmap — internally called the "358" plan — targeting 1 billion yuan in revenue at year three, 10 billion at year five and 100 billion at year eight. It is a stated ambition, not a forecast.

Funding and recognition

AgiBot has attracted a broad investor base, including Tencent, BYD, HongShan (formerly Sequoia China), Hillhouse Investment, Warburg Pincus, CDH Investments and Lanchi Ventures, alongside government-backed funds. Press reports have placed its valuation around 10 billion yuan (≈ US$1.4 billion), with talk of a higher round.

Peng was named to Forbes China's Most Innovative Companies lists in 2023 and 2024, and to Fortune China's 40 Under 40 in August 2024. He is a frequent, plain-spoken presence on Chinese developer platforms, where he documents engineering trade-offs rather than polished marketing.

Honest limitations

  • Production counts (1,000 by Jan 2025; 10,000+ by early 2026) are company-stated; independent verification of installed base and utilization is limited.
  • Revenue figures (≈ US$42,000 in 2023 to ≈ US$140 million in 2025) come from media reports of company disclosures, not audited public filings.
  • The AgiBot World "exceeds Google Open X-Embodiment" claim is the company's own comparison; dataset quality is best judged by downstream task results, which vary.
  • The "358" revenue targets are internal ambitions, not guidance.
  • Valuation around 10 billion yuan is based on press reporting; exact terms are not public.

What readers can do now

  1. If you are a robotics developer, the open-sourced Lingxi X1 design files and AimRT middleware are a rare chance to study a full humanoid stack, not just a black-box API.
  2. If you train embodied-AI models, examine AgiBot World's structure as a template for how real robot interaction data is collected at scale.
  3. If you follow the sector, track whether AgiBot's stated production and revenue numbers are later confirmed in audited filings — that is the line between demo momentum and a real business.

Related coverage

Cite this article

NeuroAI (脑机网), "Peng Zhihui's AgiBot bets on open-source humanoids and a robot data engine," 27 October 2026, https://neuroai.site/a/na-peng-zhihui-agibot-embodied-ai

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Builders” → · Back to home · Markdown version