Huawei's high-end assisted driving, the map shows the "three-point force"

Cover source: The movie "Matchless" 2018

In the public opinion pool where high-definition maps are "abandoned", the head team is effectively using maps. Of course, there is also a strong push for "data closed loop".

This is not an advertisement, if you think it is, please call Huawei to hit money.

Artist|Utada

Appreciation from peers, sometimes objective; And denigration may be the opposite.

For the question of "who does the best in the end of high-end assisted driving", a consensus has barely been reached in the industry at present——

"Huawei? Rounding counts as one. ”

From a technical point of view, they have indeed been recognized by the ADAS and L4 engineers of the big manufacturers -

"The basic functions are average, but the high-level is excellent", "the algorithm is advanced, the toolchain is very good";

But on the other hand, from the perspective of the market, the scoffing answer is simple and crude but hits the heart:

"Well done, but so what? It's not that the high match can't be sold."

Take NOA, a certain fox model city NOA with more than 400,000 as an example. A car rental company in Shanghai originally bought a high-end car and later bought 5 more.

The reason is interesting, including the United States, Germany and Japan and overseas Tier1 system engineers have come to test drive, after sighing "lying in the groove, doing well", they immediately went back to write reports.

"Shanghai's urban NOA function experience is indeed good, at least better than ours."

One Tier1 engineer working on the L2++ system acknowledged the gap, joking that he didn't know if his company needed to. "The bosses have never experienced the car, and they don't know where the superiority of developing high-end comes from."

However, some people said bluntly: "Although Huawei's overall strength of high-end assisted driving is the top 2 in China, the investment is also unmatched by all competitors." ”

The cost is indeed high.

Comparing the configuration of several hundred people in car manufacturers, startups, and Tier1, it is rumored that Huawei's entire team size (software and hardware, maps, road tests, etc.) related to autonomous driving is as many as 7,000 people;

According to a product manager's calculation, the cost of Huawei's assisted driving module on a car of the same price is twice that of Xpeng.

In addition to expensive hardware configurations such as MDC810, 13 cameras, and 3 lidars, Huawei's "mid-layer capabilities" are customized and optimized from algorithms, middleware, and chips. At the bottom level, there is also the invisible help of road mining, map and data closed-loop.

In addition, some L4 engineers pointed out that when the Transformer model was just blown on the market, Huawei actually used it. "You have to admit that they are the fastest Chinese companies that imitate Tesla at the moment."

However, this part of the cost is implemented in the price of the terminal, which is the gap between the low-end and high-end versions of nearly 200,000. Interestingly, a certain fox low configuration is an ADAS configuration provided by the head overseas manufacturer, which has been evaluated as "rotten but really cheap".

This is the cruelest thing about car intelligence - consumers really don't spend hundreds of thousands to buy a good high-end driving assistance system.

However, as a quasi-car factory, Huawei's autonomous driving capability construction logic is worth learning.

01 The "value" of maps

From 2022, the seemingly cost-driven assisted driving "high-definition map revolution" has swept the automotive market, and Huawei is one of the shouters.

However, at the same time, it has been rumored in the industry that Huawei has an extremely large urban road data acquisition team to provide trusted "road priori data" for intelligent driving systems.

In response to this rumor, a high-definition map industry source pointed out that maps do play an important role in Huawei's improvement of assisted driving capabilities.

In October 2020, Huawei released a petal map called Petal Maps at the Mate40 mobile phone conference, providing mobile phone users with positioning, driving navigation, and real-time road condition services.

But at that time, this map APP team mainly served overseas, first integrating the data of an overseas map dealer, and then building a navigation and POI engine based on this.

Later, the same team cooperated with NavInfo in China to put the map into the car (cockpit); Combined with his own map Grade A qualification, he was finally taken to support the needs of smart driving.

"So Huawei has made a complete 'takeaway', which is a serious approach, combining mobile phones, car machines and smart driving maps into one." So do it deeply. ”

And at the end of 2022, the merger of the car map and the mobile map team also confirmed this.

Some technical people speculate that as a map Grade A qualification holder, Huawei is fully capable of forming a team and driving the Huawei Inside production vehicle to collect data on the field——

A model equipped with three lidar can indeed achieve the high accuracy required for surveying and mapping.

The data collected back can supplement its three businesses - mobile phones, car machines and smart driving. This approach is similar to building a huge data middle office and is also considered part of a "full data closed loop".

The benefits are simple: one map is multi-purpose, the data is accurate, and the iteration is rapid.

However, the premise that Huawei can do this is that on the one hand, it has strong financial resources, and on the other hand, it has a decisive Grade A surveying and mapping qualification. Most car manufacturers have obvious shortcomings in this point and can only rely on the picture vendor.

For example, new forces such as Xiaopeng, Ideal and Xiaomi have companies such as AutoNavi and Meituan behind them, doing some data compliance or road procurement work for the former.

Image from Here

Therefore, the trend of "going to high-precision maps" in today's automotive intelligent market echoes with "getting rid of map vendors". But using Huawei as a frame of reference is not just about cost.

It's "bad experience".

According to data from the internal feedback of a new force car factory, the main reason for the downgrade and takeover of highway NOA is not "perception problem", but "wrong map data".

"The real high-speed NOA average, about 30 kilometers to take over or downgrade once.

In other words, the mileage of more than 100 kilometers from Suzhou to Shanghai will have to be taken over three or four times. The core reason is that the map data is wrong. ”

A technical leader of a technology company pointed out that in practical applications, the absolute accuracy of high-definition maps can reach 50 centimeters, which is "thank God";

The 10 cm and 5 cm promised by the picture dealers in many advertisements are "international jokes".

In addition, problems such as "solid lines becoming dotted lines", dashed lines being drawn as solid lines, and missing lane lines in HD maps are still extremely common after "getting on the bus".

"Many car manufacturers who have been on high-speed NOA are now talking about high-definition maps, not just because of cost. After all, everyone has long bought the highway data of the map dealer, and the update cost per kilometer is actually not expensive.

But why do the new forces have to apply for qualifications themselves, form hundreds of collection fleets, and do dynamic updates?

It's because the experience is not good. Even if HD maps are used as a set of 'a priori data', they cannot be trusted unconditionally. ”

Therefore, less reliance on maps and more trust in perception refer more to "making the map format lighter, but updating faster".

And the way to improve the speed of updating not only requires the formation of hundreds of thousands of cars for road mining, it will definitely involve a very hot concept this year -

Data closed loop.

02 The "closed loop of data" that is not seen

As we pointed out in the previous Big Model article, this is an extremely important task that has no obvious results.

For example, if there is a positioning drift on the vehicle end, or the perception recognition is wrong, the system will send back the original video and pictures, which will flow into the computer room (on the cloud) after compliance processing, and the engineer will do individual training or draw inferences according to this scenario.

When such "difficult scenarios" accumulate to a certain extent, the corresponding algorithm model will be further iterated and upgraded, and then fed back to the car end.

This is an application of "data closed loop".

In a broad sense, as a full-link closed-loop system involving data collection, cleaning, storage, annotation, training, iteration and feedback, it not only includes deep learning, reinforcement learning links, but also maps.

As a result, "data closed-loop" works within many companies, either into the mapping team or into the perception team.

"It serves both the iteration of perceptual models such as Transformer and the iteration of maps."

An industry insider speculated that a car factory's lightweight map strategy is "based on the underlying road network of the map, and then generate a layer that can be dynamically updated in real time". Over time, the graph quotient is no longer needed.

And he believes that Huawei's logic is also the same.

At the beginning, Huawei also relied on high-definition maps from a certain photo vendor. But they have their own cloud and data closed loops, eventually perfecting their own set of "maps."

"Huawei's closed-loop data tools do a good job, especially training and labeling." An external engineer noted.

Compared with the rapid iteration in IDC, most car companies just count on letting the map dealers help them do millions of kilometers before mass production.

"In the face of the endless traffic environment of cattle, ghosts, snakes and gods in China, this is far from enough."

In this way, the reason why new forces such as Xiaopeng build data centers and try to upgrade from Grade B to Grade A qualifications through various means is also the same purpose.

Of course, the price of data closed loop is "burning money" -

The survival place of data closed loop and large model must be the computer room, and there are only a few car manufacturers willing to build huge data centers (IDCs).

In addition, due to the billions of dollars invested but the output is slow, teams engaged in data closure are usually not well received in car factories. As a "supporting team", its role is similar to that of the security department in the Internet factory -

There is no real business. If you don't reach a certain node, you can't perceive the meaning of existence. And a lot of compliance is involved.

"It's a matter of patience and funding. However, in view of the current market situation of autonomous driving, most car manufacturers are unwilling to do Huawei's investment. ”

Therefore, some capital, local governments and cloud vendors have begun to ponder the business of "third-party operation data closed-loop and large-model IDC". But whether this is the direction of the market or not, it is also confusing with the "automatic driving becomes cold".

"Car manufacturers only care about whether this year's car can be sold, and most of them don't care what will happen in three years.

Now you go and tell them to engage in IDC self-research, and after reporting to the boss, you find that this is more expensive than when you didn't do it before?

So the final answer to this question, let time tell the market. ”