Qualitative:
- First in China, second in the world (Tesla FSD is the strongest).
- However, the technology roadmap cannot support nationwide deployment and internationalization, so it must undergo transformation.
- However, XPeng has a transformation plan.
Fast progress in mass production:
Where is CNGP’s strength? First and foremost, its engineering practice has progressed rapidly.
XPeng’s ADAS has a basic selling point: from high-speed NGP to VPA parking, and then to CNGP… their products were basically the first to be mass-produced in China and could be used.
Otherwise, from theory to laboratory to demo, there are countless examples of urban ADAS in the industry, but without engineering practice, it is meaningless–
Huawei’s ADS based on the Haval Hi version amazed the whole of China last year, but this year, how many media outlets have dared to tout the Haval Hi version mass-produced car?
The existing LCC and high-speed NOA developed by Haval are garbage, do I trust that they will have OTA urban NOA in the future?
A few days ago, Horizon Robotics’ AI Day hyped up the “heavy perception, light map” route. I agree, but please let Horizon Robotics provide me with a real-life experience of mass production on the Great Wall before I can believe it.
Without mass production, there is nothing.
Even if XPeng only opened up one city in Guangzhou, it was still the first to be mass-produced. A full Chinese banquet touted is not as good as a hot steamed bun in reality.
Connecting the Three Major Scenarios:
In addition to mass production, XPeng’s CNGP’s biggest significance is that it has minimally sewn together the three major driving scenarios of “garage, city, and highway” for the first time in China.
In theory, it is possible to find a route that realizes “driving out of the garage, passing through a busy city, getting on the highway, getting off the highway and entering the city, and automatically parking in another garage.”
Even if this route is very rare, it at least shows us the possibility of universal mass production, rather than just a technology demonstration demo that cannot handle routes away from company headquarters and research institutes.
Implemented on P5:
This has been touted countless times. Who would have thought that the strongest assisted driving in China would be achieved on a $200K converted oil car made by a junk product planner, Cai Peng?
The autonomous driving department gritted its teeth, carved a laser radar first mass production, brand new visual perception architecture, new-generation control architecture… into this junk car. (At this point, it should have been G9’s 400V version that completed this mission.)
The highest configuration of P5 costs just over $200K, which also gives us hope for the widespread use of advanced assisted driving technology–Lidar and computing power must follow the semiconductor industry’s law of continuous decline; algorithm development has only marginal cost, not vehicle cost.Yes, you can say it’s not appropriate for P5 to offer this kind of ADAS at this price, but there are no cheap P5s – how can you believe that advanced ADAS can be widely accessible?
In other words, if some brands cannot achieve the ADAS experience of P5 for their new models with more than 200,000 RMB, there is no reason except for the fact that their development and product planning are stupid.
What, is there an absolute gap in intelligence between the development and leadership teams? Can Peng Gang’s supply chain be so poor that he can get things that you can’t?
If you can’t achieve it, it’s just because you are stupid, that’s all.
Only open in Guangzhou:
However, the problem still exists. This time, it’s only opened in Guangzhou, which is much less than the expected three and a half cities (the top three cities of Shanghai, Guangzhou, and Shenzhen plus half of Beijing).
Combined with the overall scarcity of the high-end version of P5, XPeng CNGP is only a “thousand-unit” scale, which is much lower than my expectation.
And these small roads in urban villages can’t be accessed; during the whole experience process, there will definitely be a lot of problems that need a long time to be fixed and reviewed.
Although it will be opened in Shenzhen later, it cannot fundamentally increase the fleet size.
Strong dependence on high-precision maps:
Moreover, because XPeng is at the forefront of engineering practice, it must be highly dependent on high-precision maps of cities; which is exactly what many fans summarized in previous comments – “light rails.”
This description is very accurate. P5’s CNGP is basically using high-precision maps as a virtual track, and the perception system is at best a supplementary function for obstacle avoidance.
Without high-precision maps, it is nothing; even because of freshness issues, strong dependence on high-precision maps can cause more bugs.
For example, urban road maintenance is a very common thing, but the HD map update time of the current map supplier in China is at least one quarter. How can you keep up with that?
You are driving on the road, and the perception system detects that the road ahead is blocked due to construction, and your eyes also see it, but the high-precision map tells the vehicle, “Go ahead boldly!” Who does the car trust? If you don’t trust the high-precision map, how do you deal with it?
You said you want to detour, but many things in your navigation information are tightly bound to the high-precision map. Once you detour, how do you deal with the delay in re-planning the route?
These are all very tricky problems and the reasons why I am not optimistic about the heavy HD Map route.
Terrible surveying and mapping qualifications:
But the reason why I am not optimistic about the heavy HD Map route is the super strict surveying and mapping qualifications in China.
According to national regulations, automakers can obtain Grade B qualifications for scanning and mapping.
This is not a big deal for XPeng’s new generation of vehicles: everyone is equipped with high-pixel surround view system + high-precision Lidar, which is better than the scanning vehicles of Gaode Maps and Baidu Maps.
Everyone is eager to kick out the garbage high-precision maps from the map vendors and use their own larger-scale and better-equipped mass-produced fleets to establish private HD Map databases with much better quality.But the relevant department said no!
You’ve created a private HD Map, but in order to use it in cars, you need Grade A qualifications. And I won’t give you guys Grade A qualifications!
Why? I can’t believe how reckless you are. Let me figure out how to manage it first, don’t create any loopholes that can cause big problems in the future!
Why do Baidu and Amap have Grade A qualifications? They’ve been cooperating with the relevant departments for decades. They know where the boundaries are and won’t cross them. They won’t take the initiative to make disruptive innovations. Both sides are at ease.
Why does Yikatong have Grade A qualifications? It’s obvious that the relevant department knows their lack of ability. Although it’s also part of the automotive industry, with such poor capabilities, what kind of waves can they create?
Tesla and WeRide, on the other hand, have the ability and ambition to create disruptions. The relevant department likes and fears them. They like their strong abilities and good performance, but fear that their capabilities are too strong and exceed their imagination, making them difficult to control.
A Dead-End of Policies and Technology:
Therefore, I didn’t hype up Xpeng’s CNGP much before. I believe that the concerns of China’s mapping department are also shared by governments around the world.
The high-precision map data loaded in cars is so rich that it demands extremely high management capabilities from any government. It’s like pushing the legal boundaries based on crazy explorations.
Not to mention China, I think Europe and the United States will also strictly regulate HD maps when they come to their senses.
So I don’t think this is the right way to go. The policy risk is too high. It’s not only difficult to be widely deployed in China, but also a huge problem if you want to sell overseas.
Luckily, There’s LCC-L
Why did I hype up Xpeng’s LCC-L so much? Because it’s a killer in response to policies. I don’t need HD Maps anymore. I just blindly perceive forward!
Xpeng’s LCC-L is essentially the same as Tesla’s FSD: enhance the perception system to see things and enhance the decision-making system to understand things. As long as there’s lane-level navigation map data, the car can drive like a human being.
The technical principle of CNGP, on the same road, is that the high-precision map tells me that the left lane is for turning left, and I choose that lane based on the data logic. Then based on the high-precision map’s designated best route, I complete the left turn.
But if we look at the technical principle of Tesla’s FSD or Xpeng’s LCC-L, it’s like this: I first identify the ground symbols of the three lanes and verify them with the navigation map data. Good, we’ve confirmed that the furthest left lane is for turning left. Then, using the perception system to detect the opposite lane that matches the “Left Turn” action, we internally plan the best route and complete the left turn.
In plain language, high-precision map routing is like machine operations, while FSD and LCC-L are like human drivers.So the strength of XPeng lies here: In addition to the preemptive HD Map technology route, they also have a super-strong, equally mass-produced backhand.
“Besides the fancy martial arts routine, I also have real fighting skills!” This is where any other car maker, so-called self-driving company… can’t catch up.
Once it’s solved, it’s a knockout:
However, no matter what, XPeng’s mass production is still worth recognizing; because the LCC-L route is extremely difficult, even if it is used as a shadow system of CNGP, it still requires a lot of time to collect data, train models… and even undergo hardware updates.
Human technological progress has never been achieved overnight, successful industrial innovation is always achieved through small, quick iterations.
Even if we know that the heavy and high-precision map route is wrong, it does not prevent us from using this route to validate hardware, algorithms, and policy management in advance…
Moreover, traditional car companies, especially Germany and Japan, have to temporarily use HD Maps when they want to do high-level ADAS; because their development system is too outdated and can only rely on suppliers.
And traditional automotive suppliers cannot adapt to the rapid semiconductor, software-based trend of car companies; good things are all in the private technology system of enterprises like Tesla and WeRide, and they are playing vertical integration, how do you play?
When they are still struggling in HD Maps, we have already completed this route and developed the heavy sensing route.
Then it’s a knockout, one punch kills an old aristocrat-“My lord, eat sh*t!”
This kind of thing has already happened once from fuel to electric transformation, and it will happen countless times in the future.
Time will prove everything!
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.