Breaking News | High-speed L3 autonomous driving is coming. Huawei plans to promote its implementation this year.

On April 22nd, Huawei held the Qiankun Intelligent Technology Conference and centrally released a series of intelligent vehicle solutions. The biggest surprise among them was Huawei's Qiankun Intelligent Driving Solution ADS 4.

ADS 4 is a series of solutions with a total of 4 versions. The highest - level one is the ADS Ultra flagship version, which can achieve high - speed L3. Its release means that the industry has finally caught a glimpse of the prototype of autonomous driving. In the future, the vehicle chassis, algorithm capabilities, and even the supporting vehicle insurance forms may all undergo significant changes.

The scene of Huawei's Qiankun Intelligent Technology Press Conference

The breakthrough from L2 to L3 depends on a series of upgrades made by Huawei in terms of hardware, architecture, chassis, etc.

On the day of the press conference, Huawei launched three lidars: an all - direction mass - produced high - precision solid - state lidar, an in - cabin laser vision sensor, and a distributed millimeter - wave radar. The high - precision solid - state lidar has a smaller volume with a minimum accuracy of 3 cm, which can improve the vehicle's perception ability at the rear. The in - cabin laser vision sensor can enhance the detection ability of small targets at long distances. Even in low - light conditions at night, for small obstacles 30 cm high, AEB can brake the vehicle at a speed of 100 km/h. The distributed millimeter - wave radar can improve the detection confidence and resolution of obstacles and has a stronger ability to handle rainy, foggy, and dusty weather. The data collected by these sensors will be used to train Huawei's WEWA architecture.

The scene of Huawei's Qiankun Intelligent Technology Press Conference

The WEWA architecture runs on the MDC 1000 computing platform based on Huawei's Ascend 910B chip. The computing power of up to 1000 TOPS of the Ascend 910B chip provides strong support for the real - time processing of massive data. At the same time, the collaboration between the WEWA architecture and Huawei's XMC digital chassis engine makes the ADS 4 Ultra version respond faster, reduces the end - to - end latency by 50%, improves the traffic efficiency by 20%, and reduces the emergency braking rate by 30%.

The WEWA architecture consists of a world engine in the cloud and a world behavior model on the vehicle side.

The world engine in the cloud uses diffusion generative model technology to generate various high - density virtual difficult scenarios to train the world behavior model on the vehicle side, so that the ADS 4 Ultra version is expected to break through the bottleneck of difficult scenario data. The world behavior model on the vehicle side is the industry's first intelligent driving native base model, which can understand the driving environment based on different modalities of data and then generate the optimal driving decision.

As we all know, in real - world driving scenarios, humans often encounter various corner cases such as "sudden appearance of pedestrians", animals crossing the road, and traffic light failures. These special driving scenarios, which occur infrequently and come in a wide variety, cannot be 100% covered by traditional autonomous driving codes.

Therefore, Huawei's ADS 3.0 adopts an end - to - end architecture. Through training in various real - world scenarios, the large model is continuously optimized, making its ability to handle various scenarios stronger and stronger. This is what Yu Chengdong means by "the more you drive, the better it gets".

However, there is a problem with this architecture: for some rare scenarios, due to the small number of data samples, the large model cannot be fully trained. Therefore, if a rare scenario is encountered during driving, problems such as difficult decision - making or slow reaction may occur.

So, Huawei's ADS 4 introduced the WEWA architecture. The world engine in this architecture can generate various complex virtual driving scenarios. Before the corner cases in the real world appear, the world behavior model on the vehicle side has already been fully trained in these virtual scenarios. Since the coping methodologies in many scenarios have already been optimized in repeated training, when the vehicle - side model encounters some rare scenarios, it can make more reasonable driving decisions more quickly.

The scene of Huawei's Qiankun Intelligent Technology Press Conference

With a smarter brain, more agile limbs are also needed for execution. Huawei's XMC digital chassis engine is the more agile limbs. The difference between it and the traditional vehicle control system is that it is a full - domain fusion architecture. Originally, the intelligent driving domain and the vehicle control domain "operated independently". Now, the XMC digital chassis deeply integrates the two through a centralized decision - making and planning module. As a result, the control and processing ability of the entire architecture is 10 times that of the traditional vehicle control system, and the number of controllable devices that can be coordinated has increased by 5 times.

These are the technical logics for Huawei's ADS 4 Ultra version to achieve high - speed L3 commercialization. The emergence of the first high - speed L3 has undoubtedly promoted the progress of the entire industry. However, some people question whether the high - speed L3, which is trained and verified on a large scale in virtual scenarios, is really that reliable. Relevant personnel said that there are sometimes only subtle differences between virtual scenarios and real scenarios, but sometimes the differences are very large.

Moreover, it is hard to say that the virtual scenarios generated by the large model can fully cover the corner cases in the real world. For example, the large model can generate 1000 corner cases to train the vehicle - side AI, but in the real world, there may still be the 1001st and even more corner cases that have not been discovered yet. Therefore, in his opinion, training and verification in virtual scenarios should indeed be carried out, but its greater significance lies in the improvement of efficiency. At this stage, the technical value of high - speed L3 is greater than its commercial value.

On the same day as the release of ADS 4, Huawei also released the HarmonyOS cockpit HarmonySpace 5. It can bring users four intelligent space experiences: in the travel space, the success rate of voice interaction is increased to 85%, and the Qiankun Yunque vertical - domain large model can provide professional diagnosis of vehicle problems; in the music space, the HUAWEI SOUND Premium series supports a 43 - speaker configuration, creating a 7.5.10 super - surround sound field; in the video space, the 16.1 - inch in - vehicle smart screen, combined with the "Mountain and Sea Picture Quality" engine, can enhance the visual experience of 720P videos to 1440P; in the theme space, intelligent linkage of multiple elements such as wallpapers, screensavers, and ambient lights can be achieved.