A startup founded by a former Huawei engineer shipped its 1,000th general-purpose robot and then did something its competitors mostly avoid: it posted a massive real-world robot dataset online for anyone to use. AgiBot (智元机器人), also known as Zhiyuan Robotics, is betting that the bottleneck in robotics is not hardware alone but the data to train it — and that giving data away can build the ecosystem that buys its robots.
Key takeaways
- AgiBot began mass-producing general-purpose robots on December 16, 2024, and rolled out its 1,000th unit on January 6, 2025.
- The 1,000 units split into 731 bipedal humanoids and 269 wheeled robots.
- It open-sourced the Lingxi X1 humanoid's full hardware and code in October 2024, and released the AgiBot World dataset on December 30, 2024.
- AgiBot claims its dataset is roughly ten times larger in long-horizon tasks than Google's Open X-Embodiment — a company-stated comparison.
A fast rise from a famous founder
AgiBot was founded in February 2023 by Peng Zhihui (彭志辉), widely known online as "Zhiyuan Jun" (稚晖君), a former member of Huawei's "Top Minds" programme. The company moved quickly from announcement to volume. It unveiled its first AI robot, the Expedition A1, in August 2023, and by August 2024 it launched two product families — the Expedition (远征) and Lingxi (灵犀) series — with five commercial humanoid models entering mass production.
The production milestone is concrete. On December 16, 2024, AgiBot announced it had officially opened general robot mass production. Three weeks later, on January 6, 2025, it said its 1,000th mass-produced general-purpose embodied robot had rolled off the line — 731 of them bipedal humanoids (the Expedition A2 and Lingxi X1) and 269 wheeled general-purpose robots (the Expedition A2-D and A2-W). Reporting from the company's Lin-gang (临港) mass-production factory in Shanghai describes more than 200 workers assembling robots, with most core components developed in-house, and an annual capacity described as in the thousands of units.
The open-source bet
What sets AgiBot apart from many hardware-first rivals is how aggressively it open-sourced its stack. In October 2024 it released the Lingxi X1 — a high-degree-of-freedom, modular humanoid about 133 centimeters tall and 33 kilograms — as a fully open-source design, with hardware drawings and code published on GitHub and a development guide on its site. It also open-sourced AimRT, a high-performance robot communication framework, at the end of September 2024.
Then came the bigger move. On December 30, 2024, AgiBot released AgiBot World, a million-scale real-robot dataset, planning to upload it in batches to Hugging Face, GitHub, and its own project page. According to AgiBot, the dataset covers long-horizon tasks at a scale roughly ten times that of Google's Open X-Embodiment, with about one hundred times the scene coverage, and at what it describes as industrial-grade quality rather than lab-grade. Those multipliers are the company's own claims, and they are best read as positioning rather than independently verified benchmarks.
Building the software layer too
AgiBot did not stop at data. In February 2025 it open-sourced AgiBot Digital World, a large-scale robotics simulation framework for generating simulation data and evaluating models. In March 2025 it released Genie Operator-1 (启元大模型), which it describes as its first general embodied base model, combining a multimodal large model with a mixture-of-experts architecture. Days later it unveiled the Lingxi X2, a bipedal interactive humanoid with an emotional-computing engine.
The investor roster reinforces the strategic weight: AgiBot has drawn funding from Hillhouse, CDH Investments, and BYD, which holds a 2.51 percent stake, alongside iSoftStone.
Why give it away?
The logic is circular by design. Robots in real factories generate training data; that data improves models; better models make the robots more useful; more useful robots sell to more customers who generate more data. AgiBot's public framing is that open data lowers the entry barrier for the whole field and pulls the ecosystem toward its hardware and software.
According to Li An, chief scientist at BrainNet (脑机网), the companies that win the humanoid race may be the ones that treat data as the product and hardware as the distribution channel, not the other way around.
Honest limitations
This article draws heavily on AgiBot's own announcements, the company's timeline, and Chinese media reporting that cites the company directly; the production count, open-source releases, and dataset launch are consistent across these but are not independently audited. The dataset size comparisons (ten times, one hundred times) are AgiBot's own claims and should be treated as marketing positioning rather than verified benchmarks. AgiBot World's actual usefulness depends on independent researchers reproducing results on it, which is still unfolding. The "1,000 units" figure mixes bipedal and wheeled robots, so it is not a clean count of humanoids alone. Long-term reliability and the economics of the Lin-gang factory at full capacity are not yet publicly documented.
What readers can do now
- If you build or fund robotics, pull the AgiBot World dataset from Hugging Face or GitHub and benchmark it against Open X-Embodiment on your own tasks before trusting any vendor's multiplier claims.
- Watch whether AgiBot's open-source hardware (Lingxi X1) spawns a third-party accessory and software ecosystem — that, not the unit count, is the real moat.
- Track the data-flywheel claim over time: does AgiBot's model quality visibly improve as its deployed fleet grows, or does the loop stall at pilot scale?