China's most experienced company in mass production of autonomous driving shows off its strength

This may be the fastest-growing autonomous driving company in China.

At the just concluded Q3 Brand Day of Momenta, the company regarded by industry insiders as "the company that best understands mass production of autonomous driving in China" released its latest report card:

With rapid revenue growth, Momenta, which was established less than two years ago, has achieved revenue of billions this year.

This in itself is incredible.

In terms of product implementation, 10 mass-produced products were launched in the second quarter of this year, with both hardware and software in place. The latest product is the unmanned terminal delivery vehicle "Xiaomanlu" developed in cooperation with Alibaba DAMO Academy.

What is even more beyond industry experience is the rapid iteration of technology.

Since January this year, more than 1 million kilometers of intelligent driving data have been accumulated, and these are not simulations, but actual mileage on real roads in China.

With such amazing progress, what is Haomo Zhixing’s killer feature?

All of Haomo’s senior executives made an appearance, sharing technical details rarely seen before and revealing all their secret weapons.

1 million kilometers of data in hand, but the problem is not that simple

Let’s first explain where Momenta’s 1 million kilometers came from.

Momenta was incubated in Great Wall Motor Group and was born out of the self-concerns of this traditional automobile giant. It also shoulders the core task of Great Wall's intelligent transformation.

The Wei brand Mocha model launched this year is the debut of the Haomo intelligent driving system.

The 1 million kilometers, counted from the start of testing in January this year, also includes data contributed by ordinary users after the car was launched on the market, all come from this car.

However, with 1 million kilometers of data in hand, it is not so easy to find problems.

At the Brand Day, Gu Weihao, CEO of Baidu's driverless car team, started with the questions and revealed the efforts and explorations behind Haomo's report card.

First of all, among the massive amount of data, there is not much key data that can improve the capabilities of the assisted driving system.

For example, in a video image of a city expressway, more than 60% may be straight sections with no emergencies.

No matter how much such data there is, it will not play a decisive role in improving system capabilities.

On the contrary, image data with low frequency of occurrence and small targets are the key to filling the system's shortcomings.

How to pick out these valuable data is the first challenge

However, in valuable images, insufficient model capabilities may also lead to missed detection of key small targets and reduce data utilization.

In addition, there may be "data bias problems" in model capabilities. For example, it can recognize white passenger cars but cannot recognize white passenger cars obscured by plants.

These two problems appeared in the early stages of data collection.

After obtaining processed and valuable data, the system still has to overcome other challenges.

The first is how to iterate quickly, or more generally, how to speed up the training of new models after adjusting the parameters.

Secondly, the continuous influx of data has led to rapid iteration of model versions. It is also a challenge to verify these different models in a short period of time.

Autonomous driving companies are worried about the lack of data, but companies like Haomo, which have no shortage of data and scenarios, face another level of difficulty.

Just now, at the third brand day of Momenta, CEO Gu Weihao revealed in detail for the first time how Momenta would respond.

Data and training go hand in hand, says the company that understands mass production of autonomous driving best

The problems in the development process have been found, and how to solve these problems has become the core focus of Momenta's Brand Day.

From the technical solutions to these challenges, we can also understand why Haomo, a company that understands the mass production of autonomous driving the most.

Data diagnosis method of "big leading small"

Let’s first talk about finding valuable scene data. Haomo calls this process diagnosis.

There are currently two methods of diagnosis.

The first approach is to obtain diagnostic results through clear system failure signals, such as through manual takeover signals.

That is to say, if the user discovers the system's capabilities are insufficient during use and takes over, the system will capture the data within a period of time before and after the takeover and upload it to the scene library for analysis and learning.

The second method is to diagnose errors in the vehicle-side model through a more powerful rear-end server-side model.

The vehicle-side model is limited by computing power, transmission delay, and parameters, so its initial capabilities are naturally insufficient and limited. Generally, a small model is responsible for a part of the perception task.

Therefore, in actual testing, Haomo discovered that previous system versions often missed detections of small targets at long distances.

The large model deployed on the server is called the Fundamental Model, which is a full-task perception large model based on Transformer.

It requires high computing power and consumes a lot of resources, but it has superb capabilities and can detect errors such as missed detections or false detections by small models, or reduced recognition capabilities in bad weather.

△The upper part shows the missed detection of the vehicle-side model, and the lower part shows the correction of the large model

After finding the problem, the results are returned to the vehicle-side model for retraining and learning, so that effective data can be captured to the greatest extent possible.

“Learn from one example and apply it to other cases” to solve data bias

After finding the problematic scenario, you need to supplement it with sufficient sample data for this scenario, that is, find enough other similar data of the same type.

Only by using this method to adjust samples can a better AI model be made.

Through the Great Wall Wei Mocha model that has been launched on the market, Haomo has accumulated a huge database of road scenes.

Faced with massive scenes, Haomo's approach is to first vectorize the image using an unsupervised learning method, convert the image data into feature vectors, and then cluster similar images together through spectral clustering.

After obtaining the clustering results, for the required target scene, we can find a large amount of related data of the same category as positive samples, as well as similar and easily confused data of other categories as negative samples.

And among the categories, only the data near the class center and class boundary are selected to improve the labeling efficiency.

This method can also effectively mix heterogeneous data in an appropriate way to improve the effect of the final model.

Data diagnosis relies on “the big leading the small”, while “data bias” is solved by analogy.

Parallel training reduces the time required for alchemy by half

Now that we have obtained the key data for improving model capabilities, the next step is to "refine the elixir".

Transformer has strong capabilities, but its training speed is also slow.

Even on a server with 360GB RAM and 4 V100 GPUs, standard data parallel DDP training of the Swin-Transformer network takes more than 100 hours.

If Haomo’s engineers slightly change the network structure, parameter configuration, or replace the data, the iteration cycle to see the results is nearly a hundred hours.

But these operations occur frequently, which seriously slows down technology iteration.

Therefore, in order to improve the training speed, in addition to the common data parallelism, more sophisticated model parallelism methods are also needed.

The first is data parallelism. The complete network is trained on each GPU, and the data is cut into pieces to fit the GPU. At the same time, the gradient of each layer will interact with other GPUs. This can further improve the convergence speed of the model and achieve the same training effect with fewer epochs.

This hybrid approach of data parallelism and model parallelism is called pipeline parallelism.

For swin-transformer, a pipeline parallel solution is adopted, which can increase the overall speed by 50%-80%.

Production line test scenario

With the improvement of training efficiency, new problems arise:

The model iterates quickly and has multiple versions. How to verify its effectiveness?

The mainstream approach is of course to put the model into a simulation environment for testing, but traditional simulation is a very inefficient method.

From scene design, to setting up road models, setting up vehicle models, setting up traffic flow models, to final simulation testing... each person can only do 30 per day.

Therefore, Haomo has developed an automatic conversion tool for semantic scenes and a parameter generalization tool, which can automatically convert the description text of the scene library in CSS into simulation test scenes, and obtain a huge number of simulation test cases through discrete sampling within an appropriate range.

At the same time, through parallel operation in the cloud, more than 10,000 simulation test cases can be automatically generated every day.

To put it simply, the previous data diagnosis is actually a kind of data labeling automation, and the automatic conversion of cloud-based semantic scenarios is a tool for assembly line production test scenarios.

All martial arts in the world are invincible except speed.

The same is true for autonomous driving, and Momenta has the most in-depth understanding of it.

The data is large and abundant, and the processing is fast and accurate. Everything is based on this, which makes the incredible speed of mass production possible.

A look at Momenta from a technical perspective

Within Momenta, Chairman Zhang Kai and CEO Gu Weihao are called testing maniacs, as they spend a lot of time every week personally testing smart driving products.