HomeArticle

Feng Shuo from Miji Intelligence: How Many More Nines Are Still Missing for Robots to Step Off the Demonstration Stage? | Xinglian Portfolio Updates

星连资本2026-11-16 15:36
Feng Shuo Leads Miji Intelligence to Tackle Physical AI, Boasting High Success Rate and a Long Growth Trajectory.

If a robot can successfully serve plates 9 times out of 10, would you let it help out at home?

That sounds pretty good. But that one failed attempt might mean a dropped plate, or spilled hot soup. If the task is changed to assembling an instrument, or even performing a surgical operation, a 90% success rate is far from sufficient. 90% only has one 9, and there are 99%, 99.9%, all the way up to 8 nines — 99.999999%. The more nines you get, the less room there is for the robot to make mistakes.

In the view of Feng Shuo, Associate Professor in the Department of Automation of Tsinghua University and Chief Scientist of DenseAI, what Physical AI truly needs to follow is this long slope of continuously adding more "nines".

According to his judgment, the task success rate of current robots has not even reached one 9 yet. Factory production requires about 4 nines, safety-critical scenarios such as entering households may need 6 nines, and performing surgeries may require 8 nines or even higher. Moving from 1 nine to 8 nines may take a very long time, but every step forward can unlock a commercial scenario that was previously inaccessible.

Feng Shuo, born in 1991 and a Tsinghua alumnus, has been walking on this road for ten years.

When AlphaGo defeated human players in 2016, Feng Shuo, who was pursuing his doctorate at Tsinghua at that time, thought of a different problem: the chessboard is just a toy that allows models to try and make mistakes without scruples; real robots have to face expensive error costs, and any single failure may lead to huge losses. For robots to enter thousands of households, they must first be nearly 100% fail-safe. Since then, he has been working on this goal: first, to make autonomous vehicles reach a level exceeding human drivers, then make drones, unmanned ships, and quadruped robots learn to handle edge scenarios, and keep approaching more "nines".

As he moved forward, he found a problem that no one could avoid, wrote it into a paper, and it was featured on the cover of *Nature* in 2023. He named this problem the "curse of rarity": Truly valuable data is extremely scarce, scattered in massive volumes of data, making it extremely difficult for AI to learn from it. The fact that autonomous driving is always stuck by rare road conditions, and the task success rate of robots is still below 90%, all stems from this root cause. The solution he proposed is called "Dense Learning" — "concentrate" the useful data before learning.

In January 2026, Feng Shuo founded DenseAI together with his senior fellow apprentice Shi Mengkai, bringing this methodology to a broader battlefield. His idea is very simple: robots themselves know best about their own conditions, and they can select the most valuable batch of data for themselves from massive datasets through historical experience. What DenseAI aims to achieve is to enable robots to iterate on their own.

01 Build a brain for Physical AI

DenseAI was registered in January 2026, and its name shares the same pronunciation as "Dense Intelligence". Did the idea of starting a business come even earlier?

Feng Shuo: I had the idea of starting a business a long time ago, but I did not take practical action until 2025, when I found my senior fellow apprentice Shi Mengkai to start the business together.

Why did you choose this point in time?

First, technology has reached an inflection point — environmental modeling is truly critical for Physical AI;

Second, there is strong market demand: after two years of development of Embodied Intelligence, this problem has remained unsolved, many leading companies have not made breakthroughs, and there has been no major progress at the brain level of models. We believe this is a great opportunity;

Third, the research paradigm has changed — intelligence needs to be integrated with data, scenarios and capital at an earlier stage. In the past, people would first complete the 0-to-1 progress in the laboratory, then move to the market to achieve 1-to-10 and 10-to-100. This is not the case in the AI era. Everyone may need to work together from the very beginning, otherwise there will be no data, no computing power, and no way to retain talents.

In the simplest terms, what you are building is a brain that can be connected to various types of robots, right?

Feng Shuo: Exactly, we are building a brain for Physical AI.

In current world model research, about half of the researchers are working on "next-frame video generation", while you are working on "next-state prediction". What is the difference between the two approaches?

Feng Shuo: What we are building must be physical intelligence — video generation still stays in the digital world. The core of the world model is to provide physical intuition.

For example, when I pick up this cup, there is no clear picture in my mind before I do it, nor precise calculation based on Newton's laws. What I have is only physical intuition — I know it will probably be picked up, but the trajectory of the movement cannot be calculated, nor does it exist as a clear picture in my mind. Providing this kind of physical intuition for machines is the core goal of the world model.

Many people would think that your data lacks labels for "actions" and "forces", aren't you worried about that?

Feng Shuo: Let's think about this: what we are working on is physical intuition — do people have the concept of "force" in their mind? No, they don't.

There are two types of prediction in physics: one is to look at the current state and predict what will happen next, which is kinematics; the other is to first know what action is applied and how much force is exerted, then calculate what will happen next, which is dynamics. We believe what the world model does is the former — physical intuition is not dynamics, dynamics is physical calculation. Before Newton, people lived in this world entirely relying on intuition; even now that we have learned Newton's laws, do you calculate the force when you pick up a cup? No. Humans don't need to do that, and machines don't necessarily need to either.

Then is VLA still necessary?

Feng Shuo: VLA (Vision-Language-Action model) is only a transitional product — it is a spliced work that grafts an action output onto the Vision-Language Model (VLM). The future robot brain is unlikely to be such a simple architecture: it certainly includes language capabilities, and it must also include the world model and physical intuition, and in some cases, precise calculation and memory are also required. The world model provides exactly the physical intuition, which is a very important part of the entire system.

How did you get to know Shi Mengkai?

Feng Shuo: We got to know each other when we were undergraduates. He is one year senior to me, majored in Automation in the 2009 intake, and ranked second in the Hebei Province college entrance examination. We both took on both academic and student work responsibilities at Tsinghua — I was the president of the department student union, and he was the secretary of the department youth league committee; later I took over the position of youth league committee secretary, and he served as the president of the university association for science and technology.

Ten years ago, the nation advocated mass entrepreneurship and innovation, and the entrepreneurship boom was high at Tsinghua. The university association for science and technology was in charge of innovation and entrepreneurship affairs at that time. Although we did not start a business directly back then, we were both deeply involved in those activities, and planted an entrepreneurial seed in our respective hearts.

Now the company has more than 30 employees, all with Tsinghua backgrounds. The other three co-founders are our junior fellow apprentices we invited to join the team. We got to know each other back in school, and after graduation they went to different industries. When we invited each other to start a business together, we found that we shared the same ambition and goals.

Why do you have to start a company to follow the technical route you trust?

Feng Shuo: We are definitely a technology-driven company, but we will not be a pure research lab like Neo Lab. The capabilities of Physical AI cannot be achieved overnight, this is a long slope — it may take a very long time to raise the success rate from 1 nine (90% success rate) to 8 nines, but every time we unlock a new nine, we may unlock a new commercial scenario. Just like large language models: every improvement in capability suddenly enables it to write code, then after another improvement, it suddenly can solve math problems and build agents.

The core logic for us to enter this field is: to solve key bottleneck problems in the track that the country attaches great importance to. Our goal is at least to run alongside the world's top level, and ideally to lead it. I also hope to use this platform to attract like-minded people, like-minded capital, even like-minded competitors, it doesn't matter who can carry this flag forward.

No matter what commercial achievements the company makes, the key measure for us Tsinghua people to evaluate this matter is: leave our mark on the future landscape of Physical AI. I do not have to be the one to achieve the final success, but I must be part of the effort that leads to it — this is the sentiment and mission of our team.

02 In the past ten years, every step has been preparing for the world model

You said the world model should let machines "understand the world". When did you first have this idea?

Feng Shuo: In 2016, I was pursuing my doctorate at Tsinghua. That year, my research direction shifted from autonomous vehicle platoon control to traffic environment modeling, and the direct trigger was AlphaGo — the shock AlphaGo brought to our generation was no less than that brought by ChatGPT. The fact that AI surpassed humans in the game of Go made me realize that the AI era has truly arrived: how powerful would such a strong technology be when it enters the physical world?

So I chose autonomous vehicles as the carrier, but I did not study how the car should drive. Instead, I studied what might happen around the car after it hits the road.

At that time, most people in the autonomous driving circle were working on modules such as perception, prediction, planning and control, which is equivalent to letting the car "know itself", and treating the external environment as an opaque black box; what I did was to let the car "understand the world", comprehend and predict the environment it is in — which is essentially the world model we talk about today.

Was this direction very niche at that time?

Feng Shuo: Very few people worked on it. At that time, the term "world model" had not even appeared. Everyone thought we were doing traditional traffic simulation, but they are completely different. Traditional simulation has very coarse granularity, we need to model human behaviors, vehicle behaviors, and even unexpected events such as flying foam and falling tires, otherwise the vehicle cannot handle those situations.

Our conclusion at that time was: the reason why autonomous driving has not been realized for a long time is that the understanding of the world is insufficient, and the system crashes as soon as the corner case (rare extreme scenario) appears. Now the autonomous driving industry has also started to work on world models — if the car's brain has no understanding of the world, it is impossible to achieve fully autonomous driving by only modifying decision-making and control modules.

Later you went to the University of Michigan and developed NADE (Naturalistic Adversarial Driving Environment). What capabilities did you build through this step?

Feng Shuo: NADE was published in 2021, and it is the first Nature sub-journal paper in our field. The E in NADE stands for Environment, which is exactly the world model today. There is a classic framework for intelligence research: the interaction between Agent and Environment — Agent means knowing yourself, and Environment means understanding the world. AlphaGo has pushed the Agent to the extreme, but the chessboard is fixed and closed; autonomous driving faces an open and random environment, so it is particularly important to build a good Environment module.

N stands for Naturalistic, which means real — the constructed environment must be realistic. But realism alone is not enough, the environment must also have sufficient training intensity. You can understand it as building a real world plus a "reinforcement training camp", where the intelligent agent is surrounded by coaches, and it can make up for its shortcomings wherever it does not perform well. At the same time, mathematical methods ensure that the capabilities trained in the training camp can be calibrated back to the real world, so the training will not go off track.

Later you proposed the "curse of rarity" and got it featured on the cover of *Nature*. How should we understand this concept?

Feng Shuo: The industry has been talking about corner cases and long-tail problems, but there has never been a clear definition of what the long tail is and why it is so difficult. In 2023, we proposed the scientific problem of the "curse of rarity", which is the counterpart of the classic "curse of dimensionality" in the AI field.

To put it simply, truly valuable data is too scarce, scattered in massive volumes of data, making it very difficult for AI to learn from it. AI performs very well in common scenarios, but as soon as the situation changes slightly, it breaks down.

It's like I found a valley, completed the geological survey, proved that all people coming from the grassland must pass through this valley, and explained why it is so hard to walk through. Dense Learning is exactly the method to build bridges and pave roads in this valley. Our core contribution is adding the word "dense", that's why our company is named DenseAI.

How did you figure out this problem?

Feng Shuo: At first, I only saw the phenomenon — the later the training goes, the less useful data there is, and the harder it is for the model to improve. I went back to read the classic works of deep learning from the very beginning, and found that in principle, deep learning is built on statistics, which is naturally unfriendly to rare high-value data, and even treats them as noise and discards them. In massive datasets, a small number of data points will be regarded as "outliers", but they are exactly the ones that may be the most valuable.

I received the news that the paper was selected as the cover story when I was on a business trip in China. I got up at more than 5 a.m. to catch a flight, I was so happy after receiving the email that I lost my suitcase in a daze, and later contacted the airport to get it back. When AlphaGo was featured on the cover of *Nature* in 2016, that achievement felt completely out of reach for me. In the entire Physical AI field, our paper is probably the first one to be featured on the *Nature* cover.

"Dense Reinforcement Learning for Safety Validation of Autonomous Vehicles"

What is the relationship between the curse of rarity and today's robots?

Feng Shuo: The core indicator to measure Embodied Intelligence is task success rate, which has not even reached one 9 (90%) at present. How many nines are needed for real deployment? Factory production requires about 4 nines (99.99%); for safety-critical scenarios such as serving the elderly and entering households, it may need 6 nines; for robots performing surgeries, it may need 8 nines or even 9 nines.

The journey from 1 nine to 9 nines is exactly facing the curse of rarity: the higher the requirement for success rate, the less and sparser the available data is, and no one knows what these data should look like.

It's like a student who has already scored 94 points, no one knows exactly what exercises he needs to do to get a full score of 100. Giving him competition-level or university-level exercises may not be useful, and the weak knowledge points may also be hidden in the exercises he already knows how to do. We have only reached the stage of 1 nine now, and this problem will become more and more critical as we move forward.

Then how to find out these "exercises that should be done"?

Feng Shuo: At first, we invited experts to write rules to judge what scenarios may be dangerous; the latest research no longer requires humans to write any rules, and relies entirely on learning.

Let's still use