Summary of Key Points
Recently, the "world models" approach in the field of embodied intelligence (simply put, AI with a physical form, such as robots) has suddenly become a hot topic in the investment community. CrossDimension Intelligence just received a $1 billion Series B financing, with a valuation exceeding tens of billions; companies like Ziyuanxi and Zhifang also saw their valuations soar past $20 billion, indicating rapid funding growth. This approach differs from previous mainstream technologies by focusing on enabling robots to understand the laws of the real physical world. They use virtual synthetic data for training, with the goal of making robots capable of performing actual tasks in real-world scenarios, rather than just appearing to be able to do so.
1. Why has the "world model" track suddenly become so popular?
In the primary market (where unlisted companies seek investment), "world models" have become one of the hottest tracks:
- CrossDimension Intelligence raised $1 billion in its Series B financing, with a valuation exceeding tens of billions;
- Ziyuanxi completed four rounds of financing in just two months, raising nearly $5 billion; both companies' valuations have surpassed $20 billion;
- Other companies such as Jijia Shijing and Qianxun Intelligence also completed multiple rounds of financing within a few months.
The reason for the massive influx of capital is simple: Previous robot technologies faced limitations (for example, they could only perform fixed tasks), while world models are seen as a solution to this problem—allowing robots to truly understand the physical world and adapt to various scenarios, making them the next big opportunity in the industry.
2. Who are the investors? Why have both national entities and industrial capital joined in?
The list of investors in CrossDimension Intelligence is quite impressive:
- National-level funds (large financial institutions at the state level);
- State-owned venture capital firms (investment organizations from local governments);
- Real-industry capitals (such as Lens Technology, which produces smartphone glass and wants to enter the robotics sector);
- Local innovation investment platforms (Nanshan Zhanxin Investment, Chengdu Kechuang Investment, etc.);
The involvement of these diverse investors is not just about providing funding; national entities represent policy support, while industrial capital can provide essential resources such as factories and real-world scenarios, indicating that this technology is not just an idea in the air but has practical applications (such as smart manufacturing and commercial services) with potential for profit.
3. What makes world models better than previous technologies?
The previous mainstream technology was the VLA model (vision + language + action), but it had significant issues after two years of development:
- Robots could only perform fixed tasks in specific scenarios (for example, picking up cups from a particular table) and would struggle in new environments due to poor generalization.
World models, on the other hand, aim to enable AI to understand the real physical world—things like the weight and shape of objects, how to pick them up without dropping them, and how to move them in accordance with physical laws. Every step from modeling to perception, understanding, and execution corresponds to the real world, allowing robots to adapt to various scenarios (e.g., moving different items in different factories).
4. The secret to data training: Synthetic data is crucial
Robots need a large amount of data for learning, but collecting real-world data is challenging:
- Letting robots learn through trial and error in real scenarios can be costly, as they may damage equipment (e.g., breaking cups);
- The amount of real-world data available for training is limited.
CrossDimension Intelligence uses "synthetic data" to overcome this issue. Their DexVerse engine can generate simulated data in virtual environments (e.g., a scenario where a robotic arm moves boxes). This approach is safe, efficient, and allows for the creation of an unlimited amount of training data.
5. The biggest question in the industry: Can models trained with virtual data actually perform tasks?
The main concern is whether robots trained with synthetic data can truly perform well in real-world situations.
For example, a robot may appear to be able to move boxes in videos, but will it really do so in reality? Jia Kui, the CEO of CrossDimension Intelligence, emphasizes that the goal is not for robots to just look like they are working, but for them to actually complete tasks within a real coordinate system—only then can these technologies be applied in practical scenarios (such as smart manufacturing in factories or service robots in shopping malls).
In summary, the surge in interest in world models reflects the industry's and investors' desire to address the issue of robot inflexibility and make AI more integrated into real-world applications. The next step will be to see whether these companies can successfully implement this technology, transforming robots from laboratory prototypes into practical industrial tools.