Xpeng's end-to-end intelligent driving is implemented, XNGP welcomes an epic update: L3 autonomous driving reaches the critical point

Xpeng AI Tianji system XOS 5.2.0 has gotten rid of the concept of "opening the city" and realized "opening the whole country". At the same time, with the support of the first mass-produced end-to-end large model in China, its high-end intelligent driving system has been upgraded to the level of "easy to use in the whole country". It can be said that Xpeng XNGP has become the ceiling of mass-produced intelligent driving systems.

After He Xiaopeng experienced Tesla's FSD in the United States, there were rumors that he would adopt pure vision solutions for all of Xiaopeng's high-end intelligent driving systems.

However, what really convinced He Xiaopeng was the powerful end-to-end capabilities.

What is the difference between smart driving that is usable and smart driving that is good to use?

In short, it is the difference between end-to-end intelligent driving and non-end-to-end intelligent driving.

In simple terms, traditional intelligent driving solutions adopt a modular model, which divides perception, prediction, and planning into three independent models. It requires a large number of engineers to set driving rules for various corner cases so that intelligent driving vehicles can recognize different scenarios and adopt corresponding driving modes.

The disadvantage is that the technology stack of each model is quite different, and the downstream planning model needs to rely on engineers to write a lot of code to formulate driving rules. Under the modular technology architecture, the transmission of information will be impaired, the system maintenance is difficult, and it cannot calmly deal with complex road conditions.

The ultimate end-to-end capability is to achieve scene understanding, allowing the vehicle to judge intelligent driving behavior.

Strictly speaking, the vehicle is given an intelligent driving brain.

On May 20 this year, Xpeng Motors' end-to-end large-scale model was put into mass production; on July 30, the end-to-end XNGP was officially upgraded from "can be driven throughout the country" to "easy to use throughout the country". He Xiaopeng announced at the "Xpeng Motors AI Intelligent Driving Technology Launch Conference": the AI ​​Dimensity system XOS 5.2.0 version will be pushed to all global users.

The improvement of the core capabilities of the AI ​​Dimensity system XOS 5.2.0 version lies in the "three no's": it is open nationwide to all cities, routes and road conditions.

Among them, "no city limit" means that the coverage of XNGP will be expanded to all cities in the country, truly breaking free from the constraints of high-precision maps and introducing high-level intelligent driving into the end-to-end large model era.

"Unlimited routes" means that XNGP can be applied to all public roads in the country, and the latest public beta version is already able to automatically pass through ETC toll stations.

"No restrictions on road conditions" means that XNGP can adapt to complex road conditions including U-turns, roundabouts and narrow roads.

Among the many urban NOA solutions tested by Zhijia.com, roundabouts and U-turns are unsolvable problems. Some models can successfully make U-turns by coincidence, but it sometimes works and sometimes doesn't.

Without end-to-end, this scenario cannot be solved with current technology.

Not to mention, from being usable to being easy to use.

He Xiaopeng said that these “three no limits” define the standard of “easy to use across the country”.

The so-called "easy to use" means, for a user, that it can be used and trusted.

There are currently various debates in the industry about end-to-end.

There is a distinction between true and false, and there is a dispute over the route.

The core is to transform the original four modules of perception, planning, control and decision-making into overall thinking and judgment end-to-end, and finally achieve understanding of the scene.

End-to-end, to put it simply, means that intelligent driving is defined by data and then by AI.

This dimensional upgrade makes the concept of "opening the city" that has been debated in the industry over the past year seem rather useless.

It allows high-level intelligent driving to break free from the constraints of high-precision maps and corner cases, and enter the level of full-scene understanding and full-scene intelligent driving, gradually approaching autonomous driving and breaking through to L3.

As the concept of “opening the city” becomes less important, there is no need to distinguish between highway NOA and urban NOA.

So what is the route of Xiaopeng's end-to-end big model?

It consists of the neural network XNet + the large control model XPlanner + the large language model XBrain. Xiaopeng's end-to-end large model integrates the three models of perception, prediction, and planning. Programmers do not need to write lengthy codes to formulate rules. Instead, they use massive data to train the system, giving machines the ability to learn, think, and analyze autonomously.

The end-to-end model does not suffer from information transfer loss and can better handle complex driving tasks.

With the support of the end-to-end large model, XNGP has enhanced its "brain-supplementing" learning ability, and can achieve "driving while watching", and can respond to complex and changing scenarios in real time.

According to information released by Xiaopeng Motors, in the fourth quarter of 2024, XNGP will connect ETC toll stations, parking lot gates, internal roads of the park, etc., to achieve a true "door-to-door" experience.

The so-called door-to-door can be further refined into parking space to parking space, that is, one-click opening from the departure point to the destination, achieving autonomous driving beyond L3 level.

According to the implementation plan announced by He Xiaopeng, XNGP will be open to all cities across the country in July this year.

However, before the official announcement that it can be driven nationwide, XNGP already had the ability to be driven nationwide at the beginning of this year. But in order to ensure the safety of intelligent driving, Xiaopeng Motors still spent more than half a year conducting large-scale real-car tests. Its goal is to upgrade from "able to drive" to "easy to use" before pushing it into the market.

To ensure reliability, XNGP has become the only real-vehicle test and verification system in the industry that has been carried out in more than 2,595 cities (including prefecture-level cities and county-level cities in China), with a cumulative test mileage of more than 7.56 million kilometers, and AI training has been generalized to expand the intelligent driving system in all cities across the country.

Of course, end-to-end application does not mean that the technology is fully mature. It still needs iteration, but the application of end-to-end large models can significantly shorten the evolution cycle of intelligent driving technology and experience, from half a year to a year to 5 months.

According to He Xiaopeng, after achieving end-to-end large-scale model mass production, Xiaopeng Motors' OTA updates have achieved "a version iteration every two days and an experience upgrade every two weeks." Since the AI ​​Dimensity system was first released globally on May 20, a total of five full updates have been pushed within 70 days, achieving at least 35 version iterations.

He Xiaopeng said that Xiaopeng Motors' OTA iteration speed set a record in the industry.

According to the phase plan announced at the press conference, Xpeng Motors' intelligent driving is currently in the first stage of "easy to use across the country", and it is expected to achieve the second stage of "easy to use across the country" in the fourth quarter of 2024. By the end of 2025 to the beginning of 2026, the number of users' safety and efficiency takeovers will be greatly reduced, improving the smoothness of the user experience and achieving "user love".

In addition to XNGP, the AI ​​Dimensity system XOS 5.2.0 version also implements 484 function upgrades, including the newly added "camera sensing leaving the car and parking" function and "100m tracking reversing" function.

Among them, the "100-meter tracking reverse" can automatically return along the original driving trajectory in extreme scenarios such as dead ends and narrow roads where U-turns are impossible, realizing "one-click return to the original position"; the "camera sensing leaving the car and parking in" can trigger automatic parking by covering a specific external camera after the owner gets out of the car, thereby effectively reducing parking pressure in common scenarios such as taking items from the trunk.

In terms of the smart cockpit, the AI ​​Dimensity system XOS 5.2.0 version also realizes interconnection with the Apple ecosystem, and adds practical functions such as remote voice control, gesture control, and versatile NFC touch control. In order to highlight personalization, 80 new custom instruments, 1,620 new driving modes, and all-round projection functions have also been added.

End-to-end is definitely a complete evolution for Xiaopeng Intelligent Driving.

However, Xiaopeng Motors claims to have become the world's second automaker after Tesla to apply the end-to-end big model, and also the first end-to-end intelligent driving big model to be mass-produced in China. This statement will definitely cause different automakers to raise objections.

This year's end-to-end production battle has reached a fever pitch, with Huawei, Ideal, and NIO all vying to be the first to put their vehicles into mass production.

He Xiaopeng concluded: "After computer networks, the Internet, mobile Internet, new energy vehicles and cloud services, AI will begin to lead a new era trend and technological wave after 2023, and will bring four new directions: chips, large models, driverless cars and robots. A new batch of leading companies have been born under this AI wave, and Xiaopeng Motors is one of them."