NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

AgiBot open-sourced a million-robot dataset — then built a world model to make sense of it

How a Shanghai startup paired a million-scale real-robot dataset with a world model and a base model, and why its manufacturing scale may matter more than its demos.

2026-11-15 · 866 words · NeuroAI
AgiBot open-sourced a million-robot dataset — then built a world model to make sense of it

Imagine a researcher in a Zurich lab downloading 1 million robot demonstrations recorded in Chinese factories, homes and offices, then trains a model that can predict what a robotic arm should do next. None of those robots were built in Zurich. They were made by AgiBot (智元机器人), a Shanghai company that decided the bottleneck in humanoid (人形机器人) development was not the hardware — it was the data.

AgiBot's bet is that embodied intelligence (具身智能) will advance the way large language models did: through shared, massive datasets and open competition. The company has backed that bet with an unusual combination of manufacturing scale and open release.

A dataset before a product

In December 2024, AgiBot published AgiBot World, which it describes as the first million-scale dataset of real robot trajectories collected across full real-world scenarios with industrial-grade quality control. The company contrasts it with Google's Open X-Embodiment dataset, claiming roughly 10 times the long-horizon data scale and 100 times the scene coverage, lifted from lab-grade to industrial-grade quality.

The claim is bold and partly unverifiable without independent benchmarking, but the direction is clear: AgiBot is trying to do for robot behavior what ImageNet did for vision.

According to Li An, chief scientist at BrainNet (脑机网), "A million real-robot trajectories only matter if the model can generalize beyond them — the dataset is necessary, not sufficient."

Building the brains

In March 2025, AgiBot released Genie Operator-1 (智元启元大模型), its first general embodied base model, built on a vision-language model (VLM) fused with a mixture-of-experts (MoE) architecture. A day later it unveiled the Lingxi X2 (灵犀X2), a bipedal interactive humanoid with an emotional-computing engine aimed at shows, receptions and companionship.

In October 2025 it went further, hosting the "AgiBot World Challenge" at the IROS robotics conference in Hangzhou, with two tracks — Manipulation and World Model — a total prize pool of US$560,000, and 431 teams from 23 countries across five continents. The World Model track asked participants to build systems that can anticipate physical outcomes, using AgiBot's open data and its EVAC baseline.

The world-model ambition sits on top of this stack: rather than hand-programming each task, AgiBot wants a base model that, given a goal, plans and adapts across different robot bodies — what it calls "one brain, many forms" (一脑多形). A world model, in this framing, is the part that predicts how the physical environment will react before the robot acts.

Beyond the dataset and the base model, AgiBot has pushed an open developer stack. Through its Lingqu (灵渠) community it releases Lingqu OS, the AimRT middleware and dexterous-hand SDKs, mirroring the open-source playbook of the large model (大模型) world. The aim is to make AgiBot hardware the default target for outside researchers, so that improvements compound across the ecosystem rather than staying inside the company. It is a quieter strategy than the flashy demos, but arguably the one most likely to outlast a single product cycle.

Shipping hardware at the same time

AgiBot is not only a software story. On January 6, 2025, its 1,000th mass-produced general embodied robot rolled off the line — 731 bipedal humanoids and 269 wheeled general-purpose units. On November 20, 2025, its Yuanzheng A2 (远征A2) set a Guinness World Record for the longest humanoid walking distance, 106.286 km. By December 8, 2025, the 5,000th robot had been produced at its Shanghai factory.

Commercially, the clearest signal came from China Mobile. In a 2025–2027 humanoid procurement, AgiBot won the full-size package worth 78 million yuan (≈ US$11 million) of a total 124 million yuan (≈ US$17 million) project, beating out rivals for the larger slice. And in March 2026, a robot co-developed with SAIC-GM, the "Nengzai-1" (能仔1号), went to work on a Buick battery production line — among the first humanoids deployed on a real mass-production line in China's auto industry.

The harder questions

Scale and openness do not by themselves prove a robot can work. AgiBot's deployments are concentrated in flexible manufacturing, research, education, performance and reception — scenarios where a failure is embarrassing but not costly. The leap to unreliable, unstructured homes or hospitals remains unproven. Its pace of announcements — a walking record here, a 5,000th unit there — can also obscure how little is independently verified about day-to-day reliability once a machine leaves the demo floor.

Honest limitations

Key figures come from AgiBot's official channels (its WeChat announcements, Baike entries summarizing them) and Chinese state-affiliated business media (China Economic Net, CNR). The "10x / 100x" comparisons to Google's dataset are the company's own and were not independently benchmarked by this writer. The China Mobile win is corroborated by multiple outlets including China Securities Journal, but unit counts and delivery status were not separately verified. I have not assessed the real-world reliability of Genie Operator-1 or the Nengzai-1 deployment beyond announced milestones. Currency conversions are approximate.

What readers can do now

  • If you build with robot data, examine AgiBot World's licensing and task coverage before assuming it fits your domain.
  • Judge embodied-AI claims by generalization across bodies and scenes, not by dataset size alone.
  • Watch whether AgiBot's line deployments move from "piloted" to "routine" over the next two quarters — that is the real test of the world-model thesis.

Related coverage

Cite this article

NeuroAI (脑机网), "AgiBot open-sourced a million-robot dataset — then built a world model to make sense of it," 15 November 2026, https://neuroai.site/a/na-robot-agibot-world-dataset

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Humanoids & Robotics” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories