Dialogue with Huang Qingqiu, former top leader of Huawei's Intelligent Driving AI Division: Data quality comes first, model architecture comes second.

"Big data is the driving force of machine learning." This conclusion from Fei-Fei Li, known as the "Godmother of AI", is being accepted and recognized by a growing number of AI practitioners, including Huang Qingqiu, former "Genius Youth" of Huawei and CTO of Moqi Intelligence.

After Moqi Intelligence completed a series of angel round financing totaling more than 1 billion yuan, 36Kr met Huang Qingqiu in Shanghai. He told 36Kr: "Assuming the data quality is infinitely high, even the simplest model can be trained to produce excellent results."

Among the dazzling stories told by numerous embodied intelligence startups, Huang Qingqiu's data-centered problem-solving approach seems very simple and highly pragmatic.

Huang Qingqiu was the second batch of "Genius Youth" at Huawei. In 2020, he joined Huawei's Automotive BU under this program and served as the head of the Autonomous Driving AI Department. During his tenure at Huawei, Huang Qingqiu successively led the breakthrough of AI-based LiDAR perception, AI-based perception fusion system, and the full AI transformation of the entire assisted driving system in the ADS 4.0 stage.

In addition, he led the team to build Huawei's autonomous driving data engineering system from 0 to 1. This work was launched in the ADS 2.0 era, and "was already relatively mature in the 3.0 stage".

In the ADS 2.0 era, the data system led by Huang Qingqiu only served GOD (General Obstacle Detection Network). By the 3.0 era, in addition to GOD, it also served all static perception, traffic lights, and even prediction modules. In the 4.0 era, this data closed-loop system was able to serve the entire end-to-end process.

It is precisely this experience of deeply participating in data engineering that has accumulated valuable methodologies for his current entrepreneurial career in embodied intelligence.

Almost all embodied intelligence entrepreneurs recognize that the ultimate goal of embodied intelligence is to enter households on a large scale. An important reason why current embodied robots cannot achieve this is that the generalization ability of embodied models is insufficient, which means the "brain" of the robot is not intelligent enough.

To solve the problem of intelligence, Huang Qingqiu regards data engineering as the key to breaking the deadlock. In Huang Qingqiu's view, "the model architecture is as simple as possible", and "assuming the data quality is infinitely high, even the simplest model can be trained to produce excellent results".

He explained to 36Kr that the essence of a model is information compression and modal conversion, and the upper limit of a model's capability depends on the number of parameters. Most improvements made to model architectures mainly improve the efficiency of the model in absorbing data, and have relatively low impact on the upper limit of the model's capability. Such efficiency gap can be compensated by more data and longer training time.

Therefore, Huang Qingqiu proposed that at the current stage where high-quality embodied data is scarce, data engineering should take higher priority than model architecture innovation.

Based on this cognition, Huang Qingqiu built a data flywheel for Moqi Intelligence:

In the data collection phase, Moqi Intelligence has independently developed lightweight wearable devices, and hired many cleaners to collect data in various commercial scenarios such as hotels and commercial-residential apartments.

"After the data is collected, strict data quality screening is required. Since some data may have problems such as inaccurate movements or non-standard wearing of devices, we need to filter out all valid data."

"After screening, the data needs to be categorized. This ensures that even with massive data, we can accurately retrieve the exact data needed for training each specific model."

After categorization, the company also performs automatic annotation on the data, labels the ground truth required for training, and finally uses the data for model training. At the same time, Moqi Intelligence has designed multiple layers of defense to conduct large-scale evaluation of the models. "For scenarios that perform unsatisfactorily after evaluation, we will go back to recollect and mine data for them."

Huang Qingqiu admitted that his concept of data engineering is not particularly rare, and in the near future, more and more practitioners will realize that embodied intelligence "requires a data closed-loop system similar to that of autonomous driving".

The data collection methods in the industry are also evolving from teleoperation to portable collection, because teleoperation is suitable for quickly building demos in specific scenarios but has limited generalization performance.

However, this does not mean that every company can build a good data closed-loop system.

Huang Qingqiu told 36Kr that what determines the performance of the data closed-loop system is sometimes "some minor details".

He took the example of "re-localization" in autonomous driving. "Re-localization" means that when labeling ground truth, if two cars pass the same place, enterprises hope that the data labeled by the first car can be reused by the second car, but this cannot always be achieved. This requires that the positioning accuracy of both the second car and the first car is high enough to ensure strict matching between the two, and many companies "cannot achieve centimeter-level re-localization".

In the data closed-loop system, "there are thousands of such details, each of which needs to be optimized to the extreme, and each detail may lead to gaps between different companies". Huang Qingqiu's in-depth understanding of data engineering precisely benefits from his experience in the autonomous driving field. In his technical roadmap, data closed-loop is the underlying foundation, and developing a self-developed native embodied model is an inevitable step to maximize the value of data.

In terms of the choice of model technical roadmap, Huang Qingqiu does not want to label Moqi Intelligence's model as a world model or VLA. He believes that two points deserve more attention: first, what the input and output are, and second, whether the model follows the end-to-end training paradigm. As long as these two foundations are firmly laid, the model training methods can be replaced at any time. Whether it is VLA or world model, "they are essentially just phased tools and plugins".

Huang Qingqiu's philosophy is that Moqi Intelligence will definitely develop its own pre-training in the future to build a native embodied model.

But before reaching this stage, Moqi Intelligence will first accumulate real machine experience and system capabilities quickly through post-training. "This stage has already passed." At present, Moqi Intelligence is conducting pre-training based on open-source models. Around the end of this year, when Moqi Intelligence has accumulated sufficient data, it will start training its own native model from scratch.

In the autonomous driving industry, which is widely recognized as highly close to embodied intelligence, leading autonomous driving companies such as Momenta have to a certain extent confirmed the feasibility of the path to intelligence via data engineering. But in the embodied intelligence industry, can this path still be replicated? This remains to be verified by more embodied startups.

The following is the transcript of the conversation between 36Kr and Huang Qingqiu, Co-founder and CTO of Moqi Intelligence, which has been edited:

Talking about autonomous driving experience: Huawei tried VLA and suspended it after finding it inferior to the VA system

36Kr Auto: What kind of process did Huawei's ADS go through to achieve full AI transformation?

Huang Qingqiu: The first modules Huawei transformed with AI were visual perception and LiDAR perception. I joined Huawei in 2020 to take charge of this work, which was around the end of 2020 to the beginning of 2021. The final implementation was applied on the Arcfox vehicle.

At that time, we replaced the mechanical LiDAR with semi-solid LiDAR, which is the small "brim" shape LiDAR seen on vehicles now. At first, everyone used the rotating mechanical LiDAR mounted on the top of the vehicle. After switching to semi-solid LiDAR, we found that the performance of the original algorithm dropped very significantly, so I was assigned to solve this problem.

Therefore, in ADS 1.0, my main responsibility was to fully transform LiDAR perception with AI. After the AI transformation, the performance was really good, which made us determined to develop multi-sensor fusion and self-developed LiDAR later.

For ADS 2.0, we started to develop AI-based fusion perception, which integrated the AI-processed LiDAR and vision modules to turn the entire target perception into an AI model.

At that time, we called it BEV front-fusion perception, which has two key concepts: front-fusion and BEV. Front-fusion means that all vision and LiDAR data are input into the neural network, and the neural network automatically decides which data to trust. BEV refers to finding a unified representation space for perception.

This completed the AI transformation of the entire target perception. At the same time, static perception also followed the BEV framework, so we had two separate networks for dynamic and static perception. In the 2.0 era, we also removed the high-precision map for the first time, realizing that the system can be driven across the whole country.

When it came to ADS 3.0, GOD was fully launched. After upgrading BEV to GOD, the original system could only perceive objects on the whitelist such as cars and pedestrians, while the GOD system can perceive all obstacles, even a sack or an iron sheet dropped on the ground. This further improved the AI capability of the entire perception system.

Then in ADS 4.0, we went a step further, integrating perception, planning and control together to start the end-to-end development, so the entire system was fully transformed into an AI model.

It started with a small point -- AI transformation of LiDAR and vision, then expanded to AI transformation of the entire fusion perception, and finally incorporated planning and control, turning the whole system into an AI model.

36Kr Auto: That is, from AI transforming a single point, then expanding to a two-stage structure, and finally to a one-stage structure. You just mentioned that the system can be driven across the whole country, which was promoted publicly by Richard Yu, Huawei's consumer business CEO. Did Huawei's external product promotion campaign affect the pace of your subsequent R&D?

Huang Qingqiu: I think it definitely had a relatively large impact. For a product to succeed, technology is only one aspect, and business performance is also very important.

For example, the core reason why we could develop subsequent GOD and end-to-end functions is that we had sufficient data, the premise of having data is that we had enough vehicles on the road, and the premise of having enough vehicles is that the vehicles can be sold well. So all links are closely connected.

Huawei's intelligent driving business began to explode in the ADS 2.0 era. At that time, in addition to the technical breakthrough of nationwide drivability, front-fusion implementation and excellent safety obstacle avoidance performance, the AITO M7 vehicle was indeed a bestseller at that stage. Therefore, both technology and commercialization made great progress in that period.

36Kr Auto: I often use Huawei ADS, and I can feel that the experience of each generation of Huawei's system is improving linearly. The experience of some other car companies' systems may fluctuate, and the performance will regress when the technical generation is updated. What determines this behind the scenes?

Huang Qingqiu: I think there are two reasons behind this. The first is that Huawei's release standards are very strict. No matter what technology is used, the final release standards are user experience and safety performance.

We will go through multiple internal test processes including unit test, simulation test, generalized real vehicle test, and beta user test. The test process may last for several months. Only versions that have passed strict tests will be officially released, ensuring that each released version is reliable in quality and better than the previous generation. This is the first point, the quality control system is very strict.

Second, Huawei is very pragmatic as a whole. We will not use a certain technology just because it looks fancy.

For example, when everyone was hyping the VLA concept, we did not fail to try it internally. We had pre-research teams experimenting with VLA, and kept comparing it with the VA-centered end-to-end system.

The final comparison results showed that although VLA used a larger number of parameters, it did not bring additional benefits. Or under the same computing power, its performance was even worse than the VA system.

Based on this, we judged that although the term VLA was very popular in the industry and academia, it may not be what autonomous driving really needs at this stage.

Talking about data engineering: Assuming the data quality is infinitely high, even the simplest model can be trained to produce excellent results

36Kr Auto: I noticed that you also mentioned in previous interviews that you attach great importance to data. Why are people now talking less about model architecture and technical routes, and focusing on data? Is data the key to solving current problems?

Huang Qingqiu: First of all, the workload of data processing is greater than that of modifying model structures. Because data requires more detailed analysis. Even for autonomous driving, we are facing 10 million clips or 100,000 hours level of data, which contains a huge number of details. So we need to conduct in-depth analysis, which takes much more time.

But for models, it may only take one hour to modify a structure, and the rest of the time is spent running experiments. After the experiment is completed, we analyze the results. The manual workload required is not particularly large, and what is needed is inspiration and sufficient GPU (computing power) to support the experiments.

Secondly, I think data is indeed more important at the current stage. Assuming the data quality is infinitely high, even the simplest model can be trained to produce excellent results. Our improvements on model architecture are mostly to improve the efficiency of the model in absorbing data, and their impact on the upper limit of model capability is not that large, far less than the impact of the number of parameters. At this stage, high-quality embodied data is far from sufficient, and relatively speaking, the upper limit of model capability is redundant.

In other words, data is the current short board. If we want to improve the capability of the entire system, we naturally need to put more effort into making up for this short board.

36Kr Auto: How is the entire data team and data process of Moqi Intelligence built now?

Huang Qingqiu: We have built a complete data closed-loop system. First of all, the data collection phase is different from that of vehicles. We developed this kind of wearable device by ourselves. Our first stop was hotels and some commercial-residential apartments, where we hired many cleaners to collect data.

After the data is collected, quality screening is required. Because some actions may be non-standard, or the devices may be worn incorrectly, there are all kinds of abnormal situations, and we need to filter out all valid data from them.

After screening, we need to categorize the data. This ensures that even if we have massive data, we can accurately retrieve the exact data needed for training each specific model. For example, if we need 100,000 pieces of data of folding clothes today, and 1 million pieces of data of placing cups tomorrow, we need a data mining system to label these data, so as to ensure that we can mine the required data at any time.

After mining, the next step is called AutoLabel. This technology has been applied in autonomous driving, which means automatically labeling all data to generate the ground truth required for training, and finally input the data into the training process. After the training, we set up multiple layers of defense to conduct large-scale evaluation of the models. For scenarios with unsatisfactory performance after evaluation, we will go back to recollect and mine data. This is the entire data flywheel.

36Kr Auto: Is your data concept relatively rare in the current embodied intelligence industry?

Huang Qingqiu: I don't think it is rare now. More and more people will gradually realize that a data closed-loop system similar to that of autonomous driving is needed. Everyone is building such a system, and the only difference is how well each company's system is built. In addition, the data collection method of the entire industry is basically evolving from teleoperation to portable collection, and it only depends on who can move faster.

However, even if everyone is doing this, there are still gaps between different systems. For example, everyone in the autonomous driving industry knows that a data closed-loop system needs to be built, but there are still obvious gaps between different companies' data closed-loop systems.

36Kr Auto: Where do these gaps mainly come from? Resources?

Huang Qingqiu: I don't think it's just resources. The data closed-loop system may contain thousands of modules, and the performance difference of these thousands of modules is sometimes caused by some minor details.

For example, we have a function called re-localization. When labeling the ground truth, suppose a car passes a certain place and we have completed the labeling, AutoLabel processing, and manual check. When the next car passes the same place, we definitely hope that the already labeled data can be directly reused instead of being labeled again, since it is the same place.

This requires that the positioning accuracy of both the second car and the first car is high enough to ensure strict matching between the two. I think many companies cannot achieve centimeter-level re-localization. This is one of the differences.

In the data closed-loop system, there are thousands of such details, each of which needs to be optimized to the extreme, and each detail may lead to gaps between different companies.

36Kr Auto: It actually depends on whether the company has a deep enough understanding of the entire data closed-loop system. Where does your understanding come from? Is it because you have deeply participated in the