HardKrypton Exclusive | A 1999-born associate professor at Tongji University and former core employee of DJI has founded a world model startup, which has raised tens of millions of yuan in financing less than a month after its establishment.
Hard Krypton has learned that WorldMind, a world model enterprise, has recently completed tens of millions of yuan in angel round financing. The investor is Yuan Sheng Capital. The financing will be mainly used for team building and expansion, as well as infrastructure construction and foundational model training. Dots Capital acts as the exclusive financial advisor of the company.
WorldMind was founded in September 2026. Its core team focuses on the interdisciplinary integration of embodied intelligence, world models and brain science. Ren Yunfan, founder and CEO, was born in 1999, who used to be a key algorithm pre-research backbone at DJI, obtained his doctorate from MaRS Lab of the University of Hong Kong, and then carried out postdoctoral research at RPG of the University of Zurich. He was selected as a national-level young talent at the age of 26, and currently serves as a doctoral supervisor and tenured associate professor at Tongji University.
Zhang Qi, CTO of the company, holds a doctorate in artificial intelligence from University of Technology Sydney, is a national-level young talent, and serves as a young scientist for the major national special project "Brain Science and Brain-Inspired Intelligence" under the Ministry of Science and Technology. He has long been focused on research of brain visual decoding and brain-inspired intelligence, and has pioneered the Mind-X series of methods for brain visual encoding and decoding. The team also has a number of doctors graduated from top universities at home and abroad, who have long been engaged in research related to world models, embodied intelligence and multimodal large models, and have in-depth technical accumulation in the field of world models.
In the past two years, world models have gradually become one of the hottest keywords in the field of embodied intelligence. The industry is facing a more fundamental problem than "making robots able to perform actions": how exactly can machines understand the real world. The Scaling Law of language models has proved that as long as data, computing power and model scale are continuously increased, capabilities can emerge continuously. But the real world is not a set of internet texts that can be crawled infinitely. It is continuous, multimodal and open, and there are always new objects, new scenarios and new changes.
WorldMind's judgment is that what world models really need to find may not be a simple Scaling path that replicates language models. In the interview, Ren Yunfan, the founder, said that they have not seen the Scaling Law of world models being truly verified. So he raised another question: do robots really need to understand the whole world?
Ren Yunfan believes that the answer is no. What a kitchen robot needs to understand first may only be a kitchen of more than ten square meters. WorldMind calls this kind of limited, specific environment directly related to the agent a "small world", which borrows the concept of "Umwelt" proposed by Jakob von Uexküll. This also turns a problem that cannot be exhaustively enumerated into a problem that can be trained, deployed and obtain real feedback.
"But a small world is not an isolated island", Ren Yunfan emphasized. The contact and friction in family scenarios may appear in another form in factory and warehousing scenarios. Therefore, the problems solved in one scenario are not only for the scenario itself, but also for accumulating knowledge for a larger world model. WorldMind hopes to first build verifiable models in specific scenarios such as families, factories, warehousing and outdoor areas, then extract the physical laws that coexist in different small worlds, and gradually form a larger world model base.
This is also their another understanding of "Scaling". Scaling does not necessarily mean having massive amounts of data from the very beginning, but can start from a small world that is good enough. Ren Yunfan believes that if a specific industry problem is solved to a sufficiently high level first, it is equivalent to finding a very good Scaling starting point. Then as scenarios, data and industries expand in the future, the model can truly grow upward along a curve.
Based on this understanding, WorldMind divides the world model into four parts: knowledge, memory, deduction and calibration. Knowledge corresponds to relatively stable objective laws, such as Newtonian mechanics, fluid, materials and illumination. Memory is more like the agent's own experience, which records what has happened recently, how objects move, and what the robot has just done.
In traditional intuition, if a model wants to face more and more tasks at the same time, it seems that it has to become larger and larger. But WorldMind believes that what really needs to grow continuously is the knowledge of the world, rather than the neural deduction module responsible for understanding "what will happen next". "Just like the traditional physics engine, no matter the robot is predicting the position of a car in the next second, or judging how a bottle will move after being pushed, the bottom layer can be attributed to dynamics and kinematics problems." In addition, the model also needs to continuously receive feedback from the real world, and use new data to correct its own predictions.
At present, WorldMind has independently developed two core technologies: memory-augmented world model and brain-inspired world action model. Among them, the memory-augmented world model introduces the fast and slow system mechanism into the training and reasoning process of the world model for the first time. Through the dynamic memory mechanism and spatio-temporal dual-stream guidance mechanism, it realizes high-fidelity environment modeling, significantly improves the utilization efficiency of real machine data, and reduces the training cost of the world model. The relevant model ranks first in the authoritative world model evaluation benchmark WorldArena1.0 open source list, and stays firmly in the top three in multiple evaluation tracks of the newly released WorldArena2.0.
The latter constructs a hierarchical joint architecture and a unified mixture-of-experts model of "world deduction - spatial planning - action generation" by introducing the brain output copy mechanism and the egocentric 4D world knowledge architecture, ranking first in the global simulation open source list of world action models. Moving further to reality, it is the online self-calibration of the world model on real machines such as drones, allowing the model and strategy to learn synchronously in the real world and evolve continuously.
Therefore, the product they are developing now is not a "robot model" in the simple sense, but a complete chain from world prediction to action execution, and then to real world feedback.
The following is an excerpt of the interview with Ren Yunfan, founder of WorldMind:
Hard Krypton: You emphasize starting from small worlds, how is the sharing between different small worlds realized specifically?
Ren Yunfan: We believe that the long-tail problems existing in every small world are essentially a projection of the unified physical laws in this scenario.
For example, home service robots may only cover 80% of the situations through data from family scenarios, and the remaining 20% of long-tail problems may involve physical laws such as contact mechanics and friction. But these laws may be a more common data distribution in factories, warehousing or other scenarios. We can learn these physical laws from other scenarios, and then use them to make up for the long-tail problems in family scenarios.
Therefore, small worlds are not isolated islands. All small worlds can jointly maintain a shared large world model base, and the large world model can distill more complete models and physical laws to every downstream small world, forming uplink and downlink data and knowledge interaction.
Hard Krypton: What is the biggest difference between your world model architecture and other world models?
Ren Yunfan: The standard world model usually predicts the next state based on the current state and the action to be taken. Our model separates knowledge and memory from the state independently. That is to say, our prediction of the next moment not only depends on the current state, but also depends on the model's cognition of the world, that is, world knowledge, as well as the tasks just executed and historical memory.
After this decomposition, one advantage is that the physical laws learned by the model are no longer implicitly entangled in the state space, but have relatively clear physical meaning. We can explicitly maintain knowledge and memory, better carry out knowledge sharing and memory extraction, and also use a smaller model to achieve better results in a specific vertical field.
Hard Krypton: What scenarios do you hope to implement landing from at the initial stage?
Ren Yunfan: In the early stage, from the perspective of practicality and company development, we may pick some "hard nuts", such as the home service scenario that is difficult to solve at present, and the kitchen scenario with strong diversity. On the one hand, it can demonstrate the company's R&D strength, and on the other hand, we hope to realize real landing.
The specific landing direction depends on several industry co-creation partners we are currently communicating with, which will roughly include some industry scenarios, industrial scenarios, and To B factories and production lines. The equipment in these scenarios is often very non-standard. We hope to use a unified model to enable robots to complete tasks in different scenarios and predict changes in scenarios. For example, in the construction site scenario, the world model has two purposes. First, as a training ground for robots, it provides training capabilities for "large embodied" equipment such as excavators, and can also be used for VLA data collection or reinforcement learning to solve some problems that have not been solved by embodied intelligence at present; second, it predicts the changes of the entire construction site, such as the impact of concrete drying and weather on the scenario, so as to help with construction period or design planning.
We hope to first define one or two closed scenarios, find industry co-creation partners, and run through the entire closed loop of data collection, model training, practical application and data return in real scenarios. This is a down-to-earth landing direction that we are focusing on.
Hard Krypton: What is the biggest engineering challenge for current landing?
Ren Yunfan: This problem is definitely difficult, otherwise there would not be no general solution up to now. But our way to solve the problem is to narrow down the scope, not to try to solve the problems of the whole industry at once, but to complete a specific scenario and task well first, and establish the data return system.
At present, we have done some verification in relatively controllable scenarios such as laboratory desktops. After entering the real scenarios in the next step, we will definitely encounter various corner cases. The To B scenario is relatively controllable as a whole, but it is much more complex than the laboratory. What really requires engineering capabilities is how to make the entire pipeline of knowledge, memory and data return run through. We are very confident in the algorithm, but more time and energy need to be invested in the engineering aspect.
As long as the environment is defined to be sufficiently compact and the scenario is defined to be sufficiently closed, the upper and lower limits of this matter are very clear. The worst case is to complete environment deduction and action execution in one task, which can be achieved by current methods. What we really want to explore and break through is the upper limit, that is, the world can gradually expand from a table and a production line to the entire factory.
So we will not doubt whether the entire pipeline can be landed. As long as the scenario is defined clearly enough, it is achievable to achieve relatively good results for one production line or one task. What we really care about is whether it can achieve better generalization from small to large.