Translation

Author: French fries fish

Last week, a female classmate who had always adhered to the belief of “looks determine worth” asked me if the “N-Something-Something Integrated Parking and Driving System” was any good when she decided to buy a car.

At that time, there was a surge of impulse that felt like biting into a mouthful of wasabi, “The First Year of Integrated Parking and Driving System is here!”

Of course, this metaphor is somewhat exaggerated, but the concept of the “First Year of Integrated Parking and Driving System” is supported by more facts.

On the one hand, at the Chinese Electric Vehicle Hundred People Conference held earlier this year, Miao Wei, Vice Chairman of the Economic Committee of the National Committee of the Chinese People’s Political Consultative Conference, mentioned that the usage rate of L2 level driving assistance in passenger cars reached 22.2% last year, and 38% in new energy vehicles, the market trend is clear enough. On the other hand, more and more intelligent driving companies and integrated parking and driving solutions have emerged this year.

For example, Airfortune Intelligence.

Level or Scenario?

At the end of April, the first mass-produced car equipped with Airfortune Intelligence’s NOA Integrated Parking and Driving System was officially launched.

This afternoon, Airfortune Intelligence brought more details about the NOA Integrated Parking and Driving System to a media communication session.

“If we’re talking about autonomous driving, what are we really talking about?”

At the beginning of the session, Air Fortune Intelligence’s founder and CEO, Chen Yuxing, raised such a question.

In fact, automotive engineers who learned SAE have long divided autonomous driving into six levels, from L0 without automation to L5 with complete autonomous driving capabilities, and established corresponding judgment criteria. Regardless of whether this standard is still scientific today, Chen Yuxing believes that “scenario” is more practically significant than “level” based on the current mass-producible technology and the goal of promoting the wide-spread of autonomous driving technology.

Chen Yuxing gave an example.User’s commuting time is 2 hours. In the majority of driving scenes during this process, currently available intelligent driving assistive technology can help users complete the driving process and reduce the fatigue of driving. However, in some extreme scenes, such as passing through a vegetable market where vegetable leaves may cover road markings and traffic is chaotic, the complexity of the road increases. To be frank, current mass production technology is difficult to handle such complex scenes. But these situations which cannot be handled only account for a small proportion of the 2-hour driving time, probably only 5 minutes.

“The idea of the scene is to solve 95% of the time with autonomous driving first, and then have humans monitor the remaining 5% of the corner cases. Later, through technological iterations and upgrades, we can solve these remaining 5% of the problems.”

This is the problem-solving approach of EHang Intelligence.

“EHang has always been a style of doing first and speaking later, not good at releasing concepts,” said Chen Yuxing in a self-evaluation.

These words sound simple, but EHang’s every punch is aimed at the pain point of popularizing autonomous driving.

If “already in mass production” is the starting move of EHang’s NOA integrated technology, “cost-effectiveness” is their ultimate solution to promote the popularization of autonomous driving.

As for how to achieve this cost-effectiveness, EHang has its own set of methodologies on core algorithms.

“The number of model parameters and accuracy are positively correlated, but not linearly positively correlated.”

Chen Yuxing described the core algorithm model of autonomous driving in this way. This idea is similar to the “grabbing big and letting small” mentioned earlier. Before a certain level, an increase in the number of model parameters can greatly improve the accuracy of perception, but after crossing a stage, although there is still a positive correlation between the two, the gain is not so obvious.

In more simple terms, in such a state, we may only get 50 cents of the effect for the 10 yuan of medical expenses we spend, and this efficiency is obviously not a good cost-effectiveness.## Introduction

In the automotive industry, we are responsible for English translation, spelling check, and wording modification. We have developed the Scalpel model pruning method to compress the model, resulting in an efficient and compact small model. Meanwhile, we use knowledge distillation to let the NOA model learn from high-precision large models to inherit their advantages. At the same time, on the algorithm optimization level, we have connected the small model platform with high-precision multi-task training, achieving the sharing of Backbone and greatly reducing power consumption.

You may not understand the technical terms above, but as end-users, it is enough to know that we have achieved the goal of developing a NOA system for cars that can operate on only 16TOPS computing power, while other models may require dozens or even hundreds of TOPS. In the world of computing, computing power is basically equivalent to cost, and by reducing the computing power required, we have developed an excellent cost-performance solution.

To be more straightforward, for automotive manufacturers, the entire NOA equipment provided by us, including domain controllers, cameras, sensors, and other devices, can be purchased at a price of several thousand yuan. Compared with other solutions of the same capability, our NOA system can reduce costs by up to 50%. This cost-effectiveness has allowed us to implement the NOA function on cars priced at around 150,000 yuan.

First-hand Experience

So how does our NOA system perform in practice? Let’s take a look at the results of a real-world test.

Due to the current epidemic situation, we can only participate in this communication meeting online. As a result, our test drive will be conducted through a video. Please consider it as a pre-class preview for our upcoming actual test drive.

During the test drive, our NOA equipment automatically controlled the vehicle to merge into the main road after entering the ramp. Then, at the night time when visibility conditions are poor, it successfully drove the vehicle to the left lane and entered the main route.

When driving in a straight line on the circular road, there were no significant difficulties, and even the ordinary Adaptive Cruise Control (ACC) with Lane Centering Control (LCC) can complete this task very well.## Highlights of the Yihang NOA

The highlight of Yihang NOA is its ability to select the correct route on the roundabout, even when faced with multiple exits. It can automatically switch between roundabout and ramp road networks. The NOA also performs stable automatic driving on curved ramps in dim light conditions, automatic overtaking of slow vehicles, automatic merging one kilometer before entering tunnels, and automatic acceleration based on road speed limits.

From these features, it appears to be a mature navigation-assisted driving system. However, the ultimate verdict can only be given after experiencing it firsthand.

Some of you may wonder if Yihang NOA’s functions are similar to those of other navigation-assisted driving systems on the market. But beware, “no difference” is precisely the biggest difference.

It is essential to remember that this is a system with only 16 TOPS computing power and a total cost of only a few thousand yuan.

This underscores Yihang’s grand vision of “promoting the popularization of autonomous driving.” Only an inexpensive and user-friendly NOA system can promote the widespread adoption of autonomous driving by the general population. A larger user base means more road test data, which can be used to update and upgrade the NOA algorithm, leading to even better and more advanced algorithm models. Such a virtuous cycle will eventually result in ideal full autonomous driving.

At that time, it will be time for the larger computing power platforms to showcase their strength.

I am looking forward to the emergence of more players like Yihang Intelligence.

As the article stated in the beginning, the first year of integrated parking and driving has arrived.

Now that the first year has arrived, I wonder if more female students will come to me to meet up?

This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.