In the past six months, world models and data collection have become core keywords in the field of embodied intelligence.
As a new paradigm of embodied intelligence “brain”, the world model is regarded as an important path to AGI; high-quality data is the scarce “fuel” for training models. Many companies have officially announced the construction of data mining factories and explored various data mining routes.
The competition among domestic embodied intelligence companies has expanded from ontology products to landing scenarios, ecological construction and other aspects.
Some companies are no longer limited to the iteration of their own products, but have also extended their reach to the upstream and downstream of the industrial chain through investment to build competition barriers.
In terms of technology, the “world model” has become the focus of attention, attracting much attention at the “AI Spring Festival Gala” Zhiyuan Conference and the first global developer conference held by the embodied intelligence unicorn company Xinghaitu.
This route has also received capital bets, and companies focusing on world models such as Daxiao Robotics and Xingyuanzhi have received large amounts of financing.
Previously, the mainstream technology routes in the field of embodied intelligence were VLA and VLM models, but they had the shortcomings of “understanding semantics but not physics” and insufficient generalization.
For example, when a robot receives the instruction to “get a water cup”, although it can recognize the object and understand the instruction, it may not understand that the gripping strength required for water cups of different weights is different.
World model is a virtual world constructed by AI in the “mind”, which is used to understand physical laws, predict future results and decide actions.
Wang Zhongyuan, president of Zhiyuan Research Institute, believes that the world model is the next generation artificial intelligence base model and can solve the problem of “not understanding the physical world”, but it is still in its early stages at this stage.
Recently, there have been disagreements between VLA and world models in the industry, and some companies have chosen to go hand in hand with the two technical routes and try to integrate them.
At the Global Developers Conference on June 16, Xinghaitu released the new generation VLA basic model G0.5 and announced it as open source, announcing the world model Fast-WAM and the whole body control basic model.
Xinghaitu adheres to the dual technical route of “VLA + WAM” to systematically study the generalization bottlenecks of embodied intelligence in an open environment. VLA ensures real-time operations, and WAM provides forward-looking deductions to collaboratively achieve intelligence leaps.
Gao Jiyang, CEO of Xinghaitu, said that VLA and the world model are of the same origin and symbiosis, and will be integrated in the future. The underlying similarities between the two are similar.
Guo Yandong, founder and CEO of Zhiping, also said that the world model is a core component of the VLA system, and the company has previously promoted research on the integration of the two.
In terms of data collection, VLA models and world models both rely on training data, but there is a large gap in the interaction data between humans and the physical world, which has prompted many head-embodied intelligence companies to deploy data mining factories.
Xinghaitu and Beijing Yizhuang jointly built Yishu Intelligence and launched a 1 million hours of ultra-high-quality real data plan, planning to complete one million hours this year and move towards 10 million hours in the next three years.
The plan implements three major systems: ontology real data, UMI data collection, and human POV data, and opens some data to enterprises in Yizhuang District, Beijing.
Previously, Daimeng Robots cooperated with China Mobile to launch the construction of a “data collection into the home” network, which can generate 1 million hours of data per year at full production; Luming Robot plans to deploy equipment in multiple cities for data collection, with the goal of millions of hours of data
Production capacity; MiFeng Technology, a subsidiary of Zhiyuan, released relevant platforms and hardware, announcing that it will achieve tens of millions of hours of data production capacity by 2026; JD.com announced the completion of an embodied intelligent data collection center, accumulating more than 10 million hours of high-quality data in two years, and launched an embodied intelligent data infrastructure.
In terms of industrial ecology, some embodied intelligence companies have ended up investing in upstream and downstream start-ups in the industry.
Xinghaitu has invested in 18 companies in the past year, with a comprehensive book return of more than 5 times. It has also launched the “Xingtu Plan” with Cathay Capital, planning to invest 1 billion yuan in five years.
Zhiyuan has invested in more than 20 early-stage companies in the past year and a half, and its annualized valuation has increased by 17 times. It released the “Yuansheng Ecological Development Plan” on June 10. It will first implement 100 million yuan in ecological support funds in 2026, and plans to invest 2 billion yuan in the next five years to empower industry, academia, research and developer groups.
This deep ecological binding helps leading players take the lead and take the initiative in industry development.
In addition, both Zhiyuan and Xinghaitu have proposed the concept of triple curves, involving industrial development curves and intelligence levels. The common points are related to driving productivity.
