First Overseas Launch of One-Stage End-to-End Solution

Author: Zhong Sheng

Editor: Mark

Producer: Red Interstellar

Header Image: Autonomous Driving Image

Recently, WeRide announced a one-stage end-to-end assisted driving solution developed in cooperation with Bosch, which is conducting road testing and adaptation verification in countries such as Germany, France, and Japan. This can be said to be the first overseas launch of a domestic one-stage end-to-end system.

WeRide has been very dominant domestically with its one-stage end-to-end assisted driving solution, not only achieving "six consecutive championships" in the Autonomous Driving Competition, but also continuously winning multiple project bids with leading companies, cumulatively securing over 30 project nominations from GAC Group and Chery Group.

AION N60 is equipped with this one-stage end-to-end solution

It is reported that in this overseas road test by WeRide, it demonstrated outstanding performance in dealing with overseas scenarios such as German unlimited speed limit highways, French dense roundabouts, and Japanese right-hand drive left-hand traffic rules.

From the perspective of autonomous driving going global, this is not just an overseas verification, but China's autonomous driving global expansion will enter a new mode stage.

In the past, autonomous driving going global not only had fewer features, generally limited to Highway NOA, and required deploying significant manpower overseas for localization engineering. But this time WeRide is bringing the one-stage end-to-end solution overseas, not only pushing City NOA features overseas, but also intends to solve the issue of how to move from single-market adaptation to global adaptation.

Therefore, this is also a global exam for China's Physical AI capabilities.

Autonomous driving is a typical Physical AI scenario, requiring deep interaction with the physical world and handling complex multi-tasks.

And L4 autonomous driving is the most stringent, direct verification field for Physical AI capabilities. It is not simply identifying objects and then doing path planning, but requires understanding the world and predicting trends to make safe and reliable decisions.

This requires the ability to reconstruct the physical world, able to generate real scenes at the pixel level within seconds; learning the physical laws behind complex traffic; deducing the causal logic behind driving behavior; pre-playing spatiotemporal evolution, deducing multiple possibilities within the next few seconds in a continuously changing road environment.

As the company with the widest global business coverage in the autonomous driving field, WeRide transforms the massive data, technical systems, and engineering experience accumulated over years of L4 autonomous driving R&D and operations into a one-stage end-to-end solution for mass-produced L2++ models. This is also why the one-stage end-to-end solution from WeRide&Bosch can win the "six consecutive championships" in the Autonomous Driving Competition, essentially bringing the capabilities of the more difficult L4 autonomous driving down to L2++.

WeRide&Bosch win China Autonomous Driving Competition · Six Consecutive Championships

In addition, a major threshold for autonomous driving going global is data, overseas real road scenario data.

Physical AI models and internet large language models have fundamental differences. Every decision plan interacts with the real world, so the error tolerance is extremely low. Therefore, Physical AI models need cognitive data from the real world to help models learn the laws of the real world.

Overseas data is a challenge for companies that have always done domestic L2++ autonomous driving, but for WeRide which has already operated Robotaxis globally, it is an existing advantage.

WeRide's Robotaxi operations currently cover Guangzhou, Beijing, Singapore, Abu Dhabi, Dubai, Riyadh, Zurich, and other domestic and overseas core markets, conducting autonomous driving R&D, testing, and operations in over 40 cities in 12 countries globally. According to previous disclosures, WeRide has deployed over 3000 autonomous driving vehicles globally. The global massive fleet provides sufficient scenario depth and breadth for training the L2++ one-stage end-to-end AI large model.

Chery Exeed EX7 comes standard with this one-stage end-to-end solution

One can say WeRide is one of the few companies globally that has accumulated real-world cognitive data. This is also a barrier for autonomous driving going global. To provide rich nutrition for the evolution of the one-stage end-to-end AI large model, the countries/regions covered must be wide enough, and the accumulated road scenario data must be rich enough.

More importantly, massive global data provides new gameplay for autonomous driving going global.

Previously, a huge difficulty faced by China's autonomous driving going global was the scenario gap.

Overseas road traffic scenarios differ significantly from domestic ones, and differences between different countries/regions are also very large. For example, German highway speed limit rules, French dense roundabouts, and many countries/regions use right-hand drive vehicles.

In the past, the differentiation of these overseas road scenarios required investing significant manpower in "localization" adaptation and deployment, doing a large amount of technical engineering hard labor and tiring work. This led to high difficulty and high cost for China's autonomous driving going global, and the speed of laying out overseas markets was very slow, because each country/region needed "localization" technical engineering to be done over again.

Chery Exeed ES also equipped with this one-stage end-to-end solution

Therefore, for Chinese autonomous driving companies, the problem at hand is: Can L2++ autonomous driving going global break through the scenario gap with lower cost and more efficient methods?

WeRide's answer is: Yes, and must do so.

By training the one-stage end-to-end AI large model through massive global real operation data, modeling using the WeRide GENESIS world model that conforms to the objective laws of the physical world, further generating scenarios, training algorithms, verifying safety. This is a new gameplay for China's autonomous driving going global, and also the underlying logic of AI. For humans, road scenarios and rules differ by country, but for AI, it belongs to the same world.

So autonomous driving going global ultimately comes down to one thing: Is AI capability strong enough? And WeRide is verifying this. Does AI learn Chinese road experience, or physical world laws?

Autonomous driving going global has welcomed the best timing. The global automotive market is looking up to the Chinese market in autonomous driving.

Firstly, the scale of Chinese car makers going global continues to surge, with overseas sales accounting for nearly one-third. Autonomous driving is becoming a key competitiveness for overseas models, and Chinese car makers are also preparing to upgrade to City NOA advanced autonomous driving features.

Secondly, not only will Chinese car makers' overseas models carry City NOA advanced autonomous driving features, but overseas car makers in their own home markets are also starting to layout City NOA. For example, Porsche and Audi chose to cooperate with Mobileye respectively, using Mobileye's solution, covering point-to-point assisted driving for highways and urban areas.

One-stage end-to-end solution test vehicle in Germany

This is different from the situation in the past few years. Previously overseas car makers were more conservative in Europe and the US markets, autonomous driving features were generally limited to Highway NOA, but now they also need to align with the Chinese market.

Therefore, currently is the best timing for China's autonomous driving going global. Firstly, the demand in overseas markets is being released. Secondly, autonomous driving suppliers in Europe and the US markets, such as Mobileye, International Tier 1, are behind domestic in autonomous driving solution technology and mass production efficiency.

Large market scale, rigid customer demand, weak competitors, this is a profitable track with a long slope and thick snow.

However, this time WeRide's overseas layout is not simply copying the global expansion model of International Tier 1s. In the past, International Tier 1s relied on a model of stacking manpower, stacking adaptation, stacking rule algorithms, while WeRide relies on AI capabilities, data-driven, using AI models as the leverage. Therefore, this is AI capability going global.

What WeRide represents is not just a globalization attempt of an autonomous driving enterprise, but also an important mark that China's Physical AI capabilities are starting to accept systematic verification from the global road system.

-END-