A post-2000s founding team has secured two consecutive rounds of financing, betting on the "brain-like architecture" for machines.

Large models are rapidly entering the physical world.

From VLAs, world models to embodied foundation models, robots are gaining increasingly strong capabilities in visual understanding, language reasoning and motion generation. As the outside world increasingly recognizes the importance of embodied brains, this field is becoming a track that entrepreneurs are flocking to.

Wuhan Qizhi Mars Landing Technology Co., Ltd. (abbreviated as "Mars Landing") is a new player in this field. It was officially established in April 2025, and its founder Zhu Yuhan is a member of Generation Z born after 2000.

Different from continuing to train a larger embodied foundation model, Mars Landing has chosen another direction: building a brain-like intelligent architecture for the physical world based on spatial intelligence.

In Zhu Yuhan's view, if models solve the problem of "what can a machine do", a real brain needs to solve a more difficult problem: how to organize these capabilities to enable the machine to continuously and autonomously complete tasks in the real world.

This kind of thinking and judgment beyond his age quickly attracted investments. Now, the financing progress of Mars Landing has been unveiled: it has successively completed two rounds of financing totaling tens of millions of yuan, led by Lihe Venture Capital, Optics Valley Financial Holdings, Ruijiang Investment and Wuhan Hi-Tech, with additional investment from existing investor Qichuang Venture.

"Model" Does Not Equal "Brain"

The reason why it targets the spatial intelligent brain-like track stems from an observation of Zhu Yuhan: the industry may have equated "model" with "robot brain" prematurely. In his view, "Essentially, a model provides one or a set of capabilities. No matter how powerful the capabilities are, they do not equal a complete brain."

This difference is particularly prominent in long-horizon tasks. Today's robots have performed well in single tasks, and even some cross-task and cross-scenario tasks: from grasping, sorting, and folding clothes to opening doors, transporting, and moving operations, models are enabling robots to "do things" more and more skillfully.

However, "being able to do many things" and "being able to complete a complex task autonomously for a long time" are two different issues. Once the task is extended from a few minutes to several hours, requiring the robot to operate continuously and autonomously across scenarios and tasks, the problem will quickly become complicated. The robot not only needs to judge what to do next, but also keep remembering where it is, what it has done, and which step the task has reached; when the environment changes or execution fails halfway, it must be able to adjust itself and continue to complete the task.

These problems can hardly be solved naturally only by relying on a stronger model. Long-term autonomy requires continuous spatial memory, task state management, multi-capability collaboration, feedback and error correction. What needs to be solved is no longer just "whether the model is strong enough", but "how these capabilities can be organized for a long time".

As Zhu Yuhan said, "Models solve the problem of capabilities, while the brain solves the problem of how to organize capabilities into intelligence." This is also the "brain-like" concept that Mars Landing refers to: instead of imitating neurons, it builds an intelligent organization architecture outside the model to form a continuous closed loop of memory, tasks, skills, actions and feedback.

The Path Has Not Converged

Focus More on Effective Experience

In the past few years, the success of large model Scaling has been fully verified, which also makes the industry naturally hope to replicate this path to embodied intelligence. But if the model is not equal to the "brain", then simply following the path of model Scaling may not be able to scale out a real machine "brain".

Language models have formed a relatively clear Scaling path, but embodied intelligence is far from reaching this stage. What data is truly valuable, how to collect it, how to organize it, and how to train it have not yet been fully finalized.

Embodied data is highly bound to the robot body, sensors, motion space and model architecture. Once the technical paradigm changes, the value of past data may be re-evaluated.

Therefore, Zhu Yuhan believes that it is far from the time to judge the moat simply by "who has more data". "The industry now likes to talk about data scales of millions or tens of millions of hours. But if the data paradigm itself has not converged, then having 1 million hours or 10 million hours of data today cannot directly explain how deep the moat is."

On the contrary, when the technical paradigm is changing rapidly, more data does not necessarily mean a deeper moat, but may also mean higher migration costs. Today's data assets may even become a historical burden in the future.

This perception is also reflected in Mars Landing's technical route: compared with simply accumulating data scale, they pay more attention to whether the machine can continuously form experience in the interaction with the real world. Data records "what happened", while experience further includes actions, results, failures and corrections.

In the view of Mars Landing, if the important foundation of language model Scaling is Token, then what embodied intelligence is worth paying more attention to may be Experience Scaling — whether the machine can continuously accumulate effective experience in actions, feedback and corrections.

Starting from Open Scenarios

Regarding the current hot trend of "robots entering factories", Zhu Yuhan also has a different judgment: "Robots entering factories" is a real demand, but "robots entering factories" does not equal "the implementation of embodied intelligence". A considerable part of the demand is essentially an automation problem.

For a factory, what it really buys is not the "intelligence level" of the robot, but productivity. Enterprises calculate whether it can be cheaper, faster, more stable, and how long it takes to get the return on investment. If traditional automation, machine vision combined with an appropriate amount of AI can already solve the problem stably, then replacing it with a more "intelligent" robot will not naturally create additional value.

Of course, in flexible manufacturing, complex assembly, non-standard operations and stock environments that are difficult to completely transform, more general robot capabilities are showing their value.

Zhu Yuhan believes that two issues need to be distinguished here: where robots are easier to land and where machine intelligence is easier to evolve are not necessarily the same issue.

Industrial production pursues minimizing uncertainty through standardization, while what general machine intelligence really needs to solve is exactly uncertainty. The more a problem can be completely standardized, the closer it is to an automation problem; the more it cannot be standardized, the greater the value of intelligence itself.

This constitutes the direction that Mars Landing truly focuses on: complex environments such as underground spaces, tunnels, mountain forests, and emergency scenarios, as well as open scenarios such as future households, elderly care, and commercial services. These scenarios have one thing in common: The world will not be completely standardized for robots.

In these scenarios, the network will be interrupted, positioning will be lost, the environment and tasks will change, and the machine must deal with a large number of problems that have not been defined in advance. Zhu Yuhan believes that this kind of complex, open and long-tail environment is closer to the real world that machine intelligence ultimately needs to face.

In the more distant future, true embodied RSI (Recursive Self-Improvement) is inseparable from the continuous interaction between machines and the real world. This is not difficult to understand. Simulation, synthetic data and world models can reduce training costs and improve learning efficiency, but machines ultimately need to accept feedback from the real world, forming a closed loop of "perception - memory - decision-making - action - feedback - error correction - re-learning" in tasks.

A failure should not only leave a log, but also change the machine's behavior when facing similar problems next time. In Zhu Yuhan's view, "The real world should not only be the final test set for embodied intelligence, it itself should become part of the machine's continuous learning."

This is another meaning of Mars Landing's understanding of "brain-like": It not only focuses on how intelligence is organized, but also on how intelligence is formed in the continuous interaction with the world.

When AI Enters the Physical World

Based on these judgments, Mars Landing has not chosen to retrain an embodied foundation model, but cuts into the underlying architecture of machine intelligence from the perspective of spatial intelligence.

When a machine enters the physical world, it needs to continuously understand where it is, what is around it, where it has been, what changes have taken place in the environment, and what the relationship between tasks and space is.

Focusing on this direction, Mars Landing has formed a technical system from spatial understanding and memory, task organization and collaboration to end-side skill execution, and has built a product matrix including Xingqing M1, Xingqun M2 and Xingmang M3.

This is not intended to replace VLA, world models or stronger foundation models in the future. On the contrary, Zhu Yuhan believes that models will definitely become more and more powerful.

However, the stronger the model and the more capabilities the machine has, the more important it is to "organize intelligence".

When AI truly enters the physical world, the problem in the next stage may no longer be just "how much the machine can do", but whether it can form continuous memory, organize complex tasks, accumulate experience from real feedback, and finally achieve long-term autonomy in a constantly changing world.

Models endow machines with capabilities, and what the brain-like architecture needs to solve is how these capabilities eventually form complete machine intelligence. This is the direction that Mars Landing has chosen to bet on.

This article is from the WeChat official account "AI of Investment Circle", and is published by 36Kr with authorization.