Interview by Dr Jürgen Dickmann, Senior Advisor, DVN
Note: On 17–18 Nov., our DVN conference will take place in Stuttgart and for the first time, host special sessions on Dual-Use and on “The Road to Type Approval: Mastering End-to-End AI Systems”.
In the runup to the DVN conference, we are presenting Key ADAS/AV players , and sharing their results, perspectives in this field. The first in this series was our report on a preview drive with Wayve’s latest vehicle; now comes the Mercedes interview, and further instalments will follow every fortnight until the conference begins
DVN-Dickmann: It was recently reported about Mercedes-Benz that you have paused Level 3 because of the currently low level of demand. Does that mean Mercedes is fundamentally giving up on Level 3?
Massing: No, we are not giving up on Level 3. Our main goal at Mercedes-Benz is to provide our customers with products and features that add value to them. Even though the feedback on our Level 3 system was very positive, we also understood the customer’s desire that the system had to offer more. A new, improved, faster and even more powerful version of MB. DRIVE PILOT can be expected in the next few years. We are currently working on designing the system boundaries in such a way that Level 3 can be used up to 130 km/h on motorways more frequently and in different weather conditions.
DVN-Dickmann: Almost at the same time, it was reported that you want to enter the Level 4 fleet business together with NVIDIA. How does this pause in Level 3 fit with the push towards Level 4?
Massing: We have a clear roadmap for the development of automated driving systems. Level 3 will come back at Mercedes-Benz soon. At the same time, we are exploring different Level 4 development paths, such as Robotaxis. Although this commitment is focused on a commercial use case, the data and experience collected will also be beneficial for our main goal: Level 4 in a Mercedes-Benz.
It becomes truly interesting at Level 4, because that is where customers can move much further away from the driving task: at Level 4, you can sleep. But that also raises the requirements for how we functionally safeguard safe operation and the fallback level.
DVN-Dickmann: At Level 4, I think one has to distinguish clearly between fleet operation or robotaxi services and the private market. In the specialist literature, the “winner takes it all” thesis is often discussed: the provider of the software stack effectively supplies the operating system and thereby captures the largest share of the value creation. How does Mercedes view that question?
Massing: Mercedes-Benz is the architect in the development of driving assistance systems – in the end our customers will always get a Mercedes-Benz system. When it comes to Level 4 and robotaxi, I do not believe in a simple “winner takes it all” logic. Mercedes-Benz has recently announced, that we want to accelerate the development of a future robotaxi ecosystem and collaborate with industry-leading partners to accomplish this. It is therefore necessary, to distribute value creation across three strong domains: first, the scalable vehicle platform from assistance systems to higher levels of automation; second, the development of the software models; and third, the fleet operators. The platform side in particular benefits from using high volumes, standardisation and regional scaling to get cost and robustness under control.
On the model side, we see a great deal of capital flowing into companies such as Momenta, NVIDIA and others. As an OEM, this helps us by integrating pre-trained models on a flexible platform and then adding only the Mercedes-specific delta training, for example for our desired driving behaviour or cooperative steering. And on the operator side, there are players that already know how to run fleets, handle certification and operate control centres. If each party does its part well, one player does not have to win everything. Then several actors can succeed, and in the end the customer benefits above all.
DVN-Dickmann: At the moment, I am seeing something among OEMs that I would call “multi-path investments”. Mercedes has been developing ADAS up to and including Level 3 for a long time and identified suitable partners early on: Momenta for China and NVIDIA for the rest of the world. At the same time, you are investing in companies such as Wayve and other software-stack providers. Why are you deliberately pursuing several paths in parallel?
Massing: The AI world is very agile and still at its beginning.
There is still a great deal to happen. We are seeing right now how quickly the technology is shifting: from classic machine learning to reinforcement learning and on to combinations with vision and language models. Which architecture will ultimately prevail in which domains has not yet been decided. At the same time, major players and investors are investing heavily in different AI companies precisely because the market is still moving so much.
We are deliberately pursuing several paths, and planting seeds across different regions. That gives us access to innovation, helps us learn early which approaches scale, and creates strategic options for us. In such a dynamic technology field, that is, in my view, exactly the right course.
DVN-Dickmann: That leads directly to the next question: how do you want to organise the division of labour with current and potential partners in future when it comes to type approval and the safety case? Are responsibilities shifting here, for example with software-stack providers taking on more of the safety case around AI models and their interaction in the vehicle, while you as the OEM take on the classical part of overall vehicle type approval?
Massing: To put it very clearly: complete vehicle approval for a basic series-production system must ultimately be carried out by Mercedes. That is the basic rule. Partners and software-stack providers supply important building blocks, evidence and tests.
In the robotaxi area this works differently. The operator, in addition to a vehicle that is compliant for approval, needs its own approval for the application and the operation, because it also carries responsibility for liability and insurance. So we provide a vehicle that is technically and regulatorily roadworthy; for the specific driverless robotaxi deployment, however, the operator must also carry additional approval and operational responsibility.
DVN-Dickmann: Just so this is clearly framed: would that mean that providers such as NVIDIA, Momenta or Wayve are responsible for the software stack and the associated evidence, while you remain responsible for the classical part of overall vehicle type approval?
Massing: One has to distinguish between two things: we provide a vehicle-side system that is compliant with type-approval requirements, and depending on the operating model there may then be a further layer of approval for the operator or the specific application.
In the robotaxi use case that responsibility shifts to the operator, because a different operating model applies there.
DVN-Dickmann: But the reverse is also true: Mercedes itself has to present the argumentation to the KBA, for example on the data basis, data handling, redundancy concepts and the interaction between AI and classical signal processing. So this responsibility cannot simply be delegated to NVIDIA, Momenta or Wayve, correct?
Massing: Our partners have to make their contribution, just as in a classic supplier relationship. If, for example, a supplier provides a steering system, we need robust evidence that the component is roadworthy. Without that evidence, Mercedes cannot approve the overall vehicle.
The same applies to software, sensors and AI functions. We need data, test evidence, safety argumentation and technical input from our partners.
For series-production systems sold to end customers, that responsibility remains with Mercedes. The partner provides the building blocks, and we bring them together for overall approval.
DVN-Dickmann: Let us come to the final question.
One issue that is currently occupying the industry and leading to greater fragmentation of markets, supply chains and organisations is “local for local”, especially in China.
You have already responded to that with partners such as Momenta for China and NVIDIA for other markets. How far are you prepared to go with market-specific functional behaviour? Would it be acceptable to Mercedes if functions in China were deliberately configured or trained differently from those in the rest of the world?
Massing: We already take that into account today. Our current generations already have market-specific characteristics because regulatory requirements and customer expectations are not the same everywhere.
DVN-Dickmann: Could that, in the future, go so far as to result in market-specific sensor set-ups?
Massing: While the functions can be adapted to market-specific requirements, an OEM always has to keep an eye on costs as well. In addition to that we are very confident that our sensor setup we introduced with MB.DRIVE is the perfect foundation for our current and future ADAS functions. Never change a running system, right?
Dickmann-DVN: Georges, thank you very much for this fascinating interview and for the insight into the strategy of Mercedes-Benz AG.I hope we will be able to welcome your colleagues to our 9th DVN Conference on 17/18 November and continue the discussion.
