JIN Yuzhi of Huawei: We won't follow the VLA route. WA is the ultimate solution for autonomous driving.

At the node when Huawei ADS 4 is about to be installed in a large number of vehicles, Jin Yuzhi, CEO of Huawei's Intelligent Automotive Solution Business Unit, announced a set of data to show the phased achievements of Huawei's automotive business to the outside world:

As of July this year, the number of vehicles equipped with Huawei's Qiankun Intelligent Driving system reached 1 million, the shipment volume of Huawei's lidars exceeded 1 million units, and the cumulative mileage of Huawei's assisted driving reached 4 billion kilometers;

By the end of August this year, a total of 28 models in cooperation with Huawei were launched, including models from "Five Kingdoms" and brands such as Avatr, Deepal, Voyah, Mengshi, GAC Trumpchi, Fangchengbao, and Audi.

In Jin Yuzhi's view, the core reason why Huawei has achieved this series of achievements in the automotive field is Huawei's long - term strategic vision.

Jin Yuzhi said that Huawei's automotive BU started investing in 2014. It took more than a decade and a large amount of R & D resources to achieve current - period profitability. After achieving current - period profitability, Huawei did not set very clear commercialization goals for its automotive business.

According to Jin Yuzhi, Huawei's consideration behind this is that focusing only on commercialization often backfires. On the contrary, as long as Huawei adheres to technological R & D and meets user needs, Huawei's automotive business will surely achieve current - period profitability one day and cumulative profitability one day.

When it comes to the choice of assisted - driving technology routes, while leading automakers are betting on VLA (Perception - Language - Action Model) and have achieved rapid catch - up in assisted - driving capabilities, Jin Yuzhi believes that WA (World - Action Model) is the ultimate solution to truly achieve autonomous driving.

In his view, VLA, against the backdrop of the relatively mature development of large models relying on LM (Language Model), converts videos into language tokens for training and then into actions to control the vehicle's movement trajectory. This approach may seem like a shortcut, but it is not the ultimate solution for autonomous driving.

Therefore, Huawei is more willing to try WA, a technology solution that currently seems difficult but, in their view, is more likely to achieve true autonomous driving.

Jin Yuzhi believes that WA is a large model that directly controls the vehicle through information input such as vision without converting various information into language and then using a large language model to control the vehicle. Here, the "vision" he mentioned is just a representative. It may come from sound, vision, or touch.

Huawei has also launched the WEWA (Cloud World Engine, World - Action Model) based on the WA architecture, which will be deployed in ADS 4.0.

Regarding the issue of whether assisted driving should be charged, Jin Yuzhi is quite straightforward: There is no such thing as a free lunch in the world. What is called "free" is just a transfer of the payment method.

He pointed out that current automakers advertising free assisted driving either offer it for free only for a limited number of years, or the price of assisted driving is already included in the vehicle price, or their assisted - driving systems are not good enough and they use users as guinea pigs. Apart from these, there is no free lunch in the world.

Moreover, from the perspective of commercial returns, it is reasonable to charge for assisted - driving systems.

During the entire usage cycle of a vehicle, assisted - driving suppliers need to continuously iterate, maintain, and perform OTA updates, all of which require cost input. For users who bought the initial version of ADS, Huawei has been iterating and upgrading it for them every year. Although users pay more at the beginning, the longer they use it, the better the experience. Therefore, the cost of Huawei's assisted driving is not high when calculated over time.

This concept of full - lifecycle management is applied not only to Huawei's Qiankun Intelligent Driving system but also to the Hongmeng cockpit.

The Hongmeng cockpit not only horizontally integrates various vertical domains based on the MoLA architecture but also vertically integrates the application ecosystem, hardware, and devices.

Jin Yuzhi observed that currently, some automakers are trying to decouple cockpit hardware and software to reduce vehicle production and R & D costs. However, this leads to a decline in user experience and high difficulty in later maintenance. Therefore, whether it is the Hongmeng cockpit or the Qiankun Intelligent Driving system, Huawei firmly adopts the full - stack model to fully couple software and hardware, thus ensuring user experience and later maintenance and upgrades.

Looking to the future, Huawei's plans for assisted driving and intelligent cockpits are as follows:

In 2026, assisted driving will have high - speed L3 capabilities and pilot urban L4 capabilities. In 2027, pilot operations of unmanned trunk logistics will be carried out, and large - scale commercialization of urban L4 will be achieved. In 2028, strive to achieve large - scale commercialization of unmanned trunk logistics;

Meanwhile, Huawei is also working hard to turn the intelligent cockpit into a "digital nanny" and AI into an AI agent.

The following is an edited exchange between 36Kr Auto and Jin Yuzhi, CEO of Huawei's Intelligent Automotive Solution Business Unit:

Question: Currently, some automakers believe that VLA is the ultimate technology route for the assisted - driving industry and can even achieve true L4 through it. What does Huawei think about this issue?

Jin Yuzhi: Companies following the VLA technology route believe that currently, people learn knowledge through various large language models such as OpenAI, converting all kinds of learning into the form of LM after learning information from the Internet.

VLA is trying, against the backdrop of the relatively mature development of large models through LM, to convert videos into language tokens for training and then into actions to control the vehicle's movement trajectory.

Huawei will not follow the VLA path. We believe that such a path, although seemingly a shortcut, is not the path to true autonomous driving. Huawei values WA, that is, world action, skipping the language link in the middle.

This path currently seems very difficult, but it can achieve true autonomous driving.

WA directly controls the vehicle through the action end or directly through information input such as vision, rather than converting various information into language and then using a large language model to control the vehicle. Here, "vision" is just a representative. It may come from sound, vision, or touch.

Question: How many players with true L3 and L4 capabilities do you think will emerge globally in the next few years?

Jin Yuzhi: Currently, we don't know how many companies will ultimately be able to achieve autonomous driving, but it must be a small number. Just like the embodied - intelligence industry was very bustling five or six years ago, with many players in the field, but now the number has decreased significantly and will continue to decline.

Whether it is moving towards end - to - end or further towards world - model control, assisted driving largely depends on data - driven development. Data - driven development essentially competes in terms of data volume, computing power, and algorithms. At that time, people will find that a common intelligent platform is very important for this industry because it is not cost - effective for a single company to invest alone.

Today, because NVIDIA's computing - power chips are prohibited from being sold in China, various domestic computing - power chips have emerged.

Question: How long does it usually take for Huawei to match a vehicle model?

Jin Yuzhi: The fastest is about 6 - 9 months.

Question: Previously, Bosch proposed that users should pay for autonomous driving, and automakers should not offer it for free. If so, will there be a situation where Huawei's Qiankun Intelligent Driving system is priced at 70,000 yuan, while competitors price theirs at 50,000 yuan or 40,000 yuan?

Jin Yuzhi: First of all, I think there is no such thing as a free lunch in the world. The seemingly free services we enjoy actually have their business models charged in another way, such as through advertising or other forms. This is the so - called "the wool comes from the pig, not the sheep", which is essentially a transfer of the payment method.

Secondly, regarding the final pricing, currently some automakers advertising free assisted driving either offer it for free only for a limited number of years, or the price of assisted driving is already included in the vehicle price, or their assisted - driving systems are not good enough and they use users as guinea pigs and will charge later when the capabilities are improved. These are the main forms. Apart from these, there is no free lunch in the world.

Thirdly, the pricing you mentioned is definitely reasonable. Because selling the assisted - driving function to consumers may seem like a one - time deal, but during the entire usage cycle of the vehicle, assisted - driving suppliers need to continuously iterate, maintain, and perform OTA updates, which cannot be done without cost input.

Take Huawei as an example. From ADS 1 to ADS 4 and then to ADS 5, these are major version iterations, and there are also minor upgrades in between each year. In the past, users only went to 4S stores for vehicle maintenance after buying a car. Now, they buy a car that is constantly updated. They can receive updates at regular intervals to make the car safer and provide a smoother service experience. All of these involve input, and it is a long - term and continuous input.

For those who bought the initial version of ADS, Huawei has been iterating and upgrading it for them every year. Moreover, we had the concept of lifecycle management for software and hardware from the very beginning. So, the first set of hardware that users bought can still be upgraded. However, the hardware of some users' cars cannot be upgraded after two years of use. It may seem cheap or even free at first, but it becomes useless after a few years and has to be discarded.

This means that a good product must be designed from the start to be able to be continuously iterated and upgraded in the future. Although users pay more at the beginning, they can use it for a longer time and have a better experience. Therefore, the cost is not high when calculated over time.

Question: Some people say that Huawei is installing more and more lidars in its Qiankun Intelligent Driving system to support price premiums. What's your view on this?

Jin Yuzhi: Essentially, Huawei increases the sensor configuration for greater safety. Our pursuit is zero casualty accidents. We believe that any measure that can improve safety is worth investing in. For example, in addition to the forward - facing lidar, we added two side - facing and one rear - facing solid - state lidars to the Zunjie S800.

Take the rear - facing solid - state lidar as an example. We added it to make parking safer. There were originally cameras and ultrasonic radars at the rear. However, the accuracy of ultrasonic radars is not very high, and cameras can only judge whether there is an obstacle through training. Moreover, the photos taken by cameras are two - dimensional and lack depth information and specific details.

For example, if there is a water pipe hanging on the wall behind the car or a long pipe protruding from the wall, and the camera has not been trained to recognize it, it cannot make a judgment. As a result, the car will definitely hit it when reversing. We have encountered such situations when parking before.

By adding the rear - facing radar, the system's accuracy can reach the centimeter level, enabling it to judge whether there is an obstacle behind and avoid scratches.

We also have some reversing scenarios. For example, after buying a new car and going back to the countryside during the Spring Festival to show off the automatic - parking function to relatives, the car may drive into a farmland because the rear is a farmland.

This shows that the system also has problems in vertical recognition. With the solid - state lidar, the system can detect a pit in the vertical direction, so the car will not drive into the farmland when reversing.

These needs objectively exist, and the increase in configuration is driven by users' usage scenarios, not deliberately added. Essentially, Huawei starts from users' scenarios and needs. Our pursuit is to make assisted driving safer whether it is for parking, in urban areas, or on highways. Other speculations are unfounded.

Author's WeChat: luckg17305264638