When Chinese ADAS Firms Sell Solutions to Germany, Algorithms Aren't Their Biggest Worry

WeRide and Bosch's L2++ solution going overseas has already begun road testing in Germany, France, and Japan. This news reminds me of a common question from product managers: When a solution needs to adapt to three completely different markets, what exactly is its "Minimum Viable Product"?

The answer might not be algorithms, but the ability to "obtain local road test licenses + understand local driving habits." WeRide chose a path that seems conservative but is actually clever—using Bosch's automotive-grade domain controller platform as the "safety shell" and their own end-to-end algorithm as the "intelligence core." The product logic of this combination is clear: Bosch handles trust endorsement and compliance thresholds, while WeRide handles the upper limit of experience.

Deconstructing the Product Combination: Why Bosch?

This isn't WeRide's first attempt at going overseas. But the difficulty of exporting L2++ (often called advanced driver assistance domestically) isn't about the technology itself, but that "the road conditions in every market are a new product." I made a comparison table to see the "special requirements" for intelligent driving systems in these three markets:

Market	Core Road Characteristics	Key Constraints for Intelligent Driving Systems
Germany	Unlimited speed highways, strict lane markings, many roundabouts without traffic lights	Aggressiveness of lane change strategies, roundabout passage logic
France	Dense roundabouts, narrow roads, inconsistent road markings	Low-speed gaming capability, tolerance for marking recognition errors
Japan	Left-hand traffic, extremely narrow streets, dense pedestrians/non-motorized vehicles	Left-side adaptation, narrow road passage, priority for yielding to pedestrians

These differences cannot be automatically generalized solely by "end-to-end" algorithms. The advantage of end-to-end is learning complex scenarios, but the disadvantage is it requires massive amounts of local, high signal-to-noise ratio training data. Currently, WeRide is only doing "road testing and adaptation verification," still a long way from true mass production and user experience optimization.

Key Numbers: According to public data, domestic L2++ solution test mileage is usually counted in millions of kilometers, while the effective scenario data collected by one overseas test car in a year might be only one-tenth of that in China. This is the first product threshold for going overseas—data loop efficiency.

Business Model Value of "End-to-End + Bosch" from a Product Perspective

WeRide didn't choose to sell algorithm software alone (like Mobileye) nor do full-stack integration themselves (like Waymo), but deeply partnered with Bosch. This decision makes clear sense commercially:

• Lower Installation Barriers: Bosch's domain controller platform has already passed automotive certifications like ASPICE CL3 and ISO 26262 ASIL-D, so OEMs don't need additional platform adaptation. WeRide only needs to run algorithms on Bosch's reference hardware.
• Shorter Validation Cycles: Bosch has ready-made test tracks, partner OEM resources, and regulatory compliance teams globally. Using Bosch's shell allows WeRide to skip the "0 to 1" compliance process.
• Risk Sharing: If a market fails (e.g., sudden regulatory changes), the algorithm can switch to another market, while Bosch's hardware platform can continue supplying.

However, risks exist too: WeRide's dependency on Bosch is extremely high. Once Bosch prioritizes other algorithm partners in other regions, WeRide's bargaining power will decline. From a PM perspective: This is a typical "platform lock-in" risk.

Underlying Contradiction in User Experience: Globalization vs. Localization

WeRide's solution is "one-stage end-to-end"—using deep learning networks for everything from perception to decision to control. The benefit is unified experience, but the drawback is: It is difficult for one model to handle both aggressive lane changes on German highways and low-speed yielding on Japanese narrow roads simultaneously.

A PM would ask: Do users really need this global capability? German consumers might care more about the strength of the "machine feel" during lane changes, while Japanese users care more about obstacle recognition during parking. WeRide currently chooses the product rhythm of "unify the algorithm first, then accumulate adaptation data through testing." This rhythm is feasible in the early validation stage, but once entering mass production delivery, they will inevitably face the choice of "making separate versions for each market."

From a business value assessment, I am more concerned about whether WeRide plans for localized OTA frequency. If a German user's car updates its intelligent driving system at the same pace as Chinese test cars, the product will struggle to truly land. OEMs need to see a clear "data return—model iteration—user upgrade" loop cycle.

Summary: Three Key Judgments on Product Logic

1. Algorithms are not the barrier; localized operations are. What WeRide is doing now is "driving piles"—testing in various countries, essentially building a data collection network. This network requires continuous investment and cannot be monetized quickly.

2. Bosch is an "entry ticket" but not a "moat." PMs should be wary: The partner's globalization capabilities might limit the freedom of their own algorithm iterations. For example, the computing power ceiling of Bosch's domain controller platform could become a bottleneck.

3. Market priorities need re-ranking. European markets (Germany, France) are main markets for high-end cars, but Japan has limited high order volumes. WeRide should concentrate resources on conquering one market first, rather than running three lines in parallel.

Leaving an Open Question

If WeRide later enters the US market (which has stricter regulatory restrictions and more complex multi-sensor fusion needs), will they need to abandon the end-to-end solution and switch to more traditional modular solutions to fit local OEM validation systems? In other words, to build a global product, should you "use one algorithm to conquer the world" or "build an independent product for each market"?

Original Link: https://www.ithome.com/0/976/190.htm