In the past six months, world models and data collection have become core keywords in the field of embodied intelligence.
The world model is a new paradigm of embodied intelligence “brain” and is considered 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 the exploration of various data mining routes.
The competition among domestic embodied intelligence companies has expanded from ontology products to landing scenarios, ecological construction, etc.
Some companies are no longer satisfied with iterating their own products, and have extended their reach to the upstream and downstream of the industrial chain through real money investment, trying to build competitive barriers as much as possible.
Technological turn: two routes begin to merge
Whether it is at the Zhiyuan Conference, known as the “AI Spring Festival Gala”, or the first global developer conference held at Xinghaitu, an embodied intelligence unicorn company, the “world model” is the focus of attention.
This route has also been bet on by capital. A group of companies focusing on world models have received large amounts of financing in a short period of time, such as Daxiao Robotics, Xingyuanzhi, etc.
Previously, the mainstream technology routes in the field of embodied intelligence were VLA and VLM models. Their shortcomings were “understanding semantics but not physics” and insufficient generalization.
For example, when a robot receives the instruction task of “holding a water cup”, it can recognize the object and understand the instruction, but it may not be aware of the different gripping forces required for water cups of different weights.
World model, that is, AI builds a virtual world in the “mind” to understand physical laws, predict future results, and then decide how to act.
Wang Zhongyuan, president of Zhiyuan Research Institute, believes that the world model is the next generation artificial intelligence base model that can solve the problem of “not understanding the physical world”, but it is still in its early stage.
Recently, there have been certain differences within the industry, and some views have put VLA and world models on the opposite side.
However, some companies choose to go hand in hand with the two technological routes and explore ways to integrate them.
At the Global Developers Conference on June 16, Xinghaitu officially 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 stated that it adheres to the “VLA + WAM” dual technology route and systematically studies the generalization bottleneck of embodied intelligence in an open environment.
Among them, VLA guarantees real-time operation, and WAM provides forward-looking deduction. The two work together to achieve a complete intelligent leap from “perception-decision-execution” to “prejudgment-optimization-closed loop”.
“We have never believed that VLA and world model are antagonistic. On the contrary, they are homologous and symbiotic, and they will increasingly converge in the future.” Xinghaitu CEO Gao Jiyang explained in an interview with the media that the bottom layers of the two models are somewhat similar. First, multi-modal data is turned into Token, and then multi-layer Transformer is used for encoding, and then output in some way, etc.
Guo Yandong, founder and CEO of Zhiping, also stated at the Beijing Zhiyuan Conference that the world model is not a competitive route for VLA, but a core component of the VLA system.
It is understood that the company has previously been promoting research on the integration of world models and VLA. In November last year, it collaborated with Peking University to launch Video2Act, a new generation architecture that integrates world models.
Data Mining Factory: “Million Hours” Plan Emerges
VLA models and world models are inseparable from training data, and there is a large gap in the interaction data they require from humans and the physical world.
This is also one of the reasons why many head-embodied intelligence companies have deployed data mining factories.
Yishu Intelligence, jointly built by Xinghaitu and Beijing Yizhuang, was unveiled 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.
Specifically, Yishu Intelligence plans to implement the three major systems of Ontology real data, UMI data collection, and human POV data. The company also plans to open 5% – 10% of its annual data output to enterprises in Yizhuang District, Beijing, for free or at preferential prices.
This model is similar to the plans previously announced by many personal intelligence companies, and the industry is rushing to lay out the data track.
Previously, Daimeng Robots and China Mobile formally reached a cooperation to jointly launch the construction of the “data collection into the home” network. The two parties relied on China Mobile’s hundreds of thousands of offline business outlets to jointly build an outbound data collection network covering the entire country.
In the first phase of the project, 1,000 sets of equipment will be put into operation, and at full production, it can generate 1 million hours of data every year.
Luming Robot plans to launch 10,000 backpack versions of FastUMI Pro in multiple cities this year to collect data in six real-life scenarios including industry, homes, hotels, restaurants, shopping malls, and offices. The goal is also to achieve million-hour data production capacity.
In April this year, Meifeng Technology, a subsidiary of Zhiyuan, released a one-stop physical AI data service platform and the MEgo series of ontology-less collection hardware. It also announced that it will achieve tens of millions of hours of data production capacity in 2026.
Major Internet companies are also entering the game.
JD.com officially announced that it will rely on the core advantages of the super supply chain to build the world’s largest and most comprehensive embodied intelligent data collection center, accumulating more than 10 million hours of high-quality data in two years.
JD.com also announced the world’s first embodied intelligent data infrastructure covering all links, including its self-developed ultra-high-definition collection terminal JoyEgoCam, embodied large model JoyAI-RA, embodied intelligent data trading platform, etc.
Gao Jiyang believes that the competition for the commanding heights of embodied intelligence models depends largely on the quality of data, which will affect model capabilities and application-level effects.
Wang Zhongyuan mentioned in a previous interview with the media that simulation data can make up for the lack of real data collection, but the accuracy of simulation data is not necessarily sufficient. It is doubtful whether relying entirely on this type of data can build a model that truly understands the physical world.
Locking in on “potential stocks”: building an industrial ecosystem
In addition to building data barrier advantages, some embodied intelligence companies have also invested in upstream and downstream start-ups in the industry.
In the past year, Xinghaitu has successively invested in 18 companies.
“In less than a year after starting investment, the comprehensive book return exceeded 5 times. This is a test of the market for our screening and empowerment projects.” Xinghaitu CFO Luo Tianqi revealed in his sharing that the company mainly invests in three directions: data-driven intelligence, application scenario breakthroughs and next-generation core technologies. They are not only targeting financial returns, but more importantly, voting for the future.
At the conference, Xinghaitu and Cathay Capital jointly launched the entrepreneurial partnership program “Star Path Plan” to provide capital, technology and industrial resources to early-stage entrepreneurial teams of embodied intelligence, and planned to invest 1 billion yuan in five years.
Chiyuan, which adheres to the full-stack route, also has its own investment layout. At the same time, it officially announced its ecological development plan a few days ago.
In April this year, Deng Taihua, founder, chairman and CEO of Zhiyuan, revealed that in the past year and a half, Zhiyuan has invested in more than 20 companies, all of which are in the early stages of development, and their annualized valuations have increased by 17 times.
On June 10, Zhiyuan released the “Yuansheng Ecological Development Plan”, announcing that it will first launch 100 million yuan in ecological support funds in 2026, and plans to invest 2 billion yuan in special funds in the next five years. Focusing on the four major dimensions of scientific research and academic innovation, educational talent training, ecological partner development, and development community operations, it will systematically empower industry, academia, research and developer groups to jointly build an open, collaborative, and symbiotic embodied intelligence ecosystem.
Zhiyuan said that relying on this development plan layout, it aims to cover thousands of scientific research institutions, link up with 10,000-level ecological partners, cultivate 100,000-level developers, and reach millions of end-users.
This type of in-depth ecological binding will not only provide financial and resource support for start-up companies, but will also help leading players take the lead and take the initiative in industry development.
It is worth mentioning that both Zhiyuan and Xinghaitu have proposed the concept of triple curves. One is the industrial development curve, and the other is the level of intelligence. Their common points involve driving productivity.
