.jpg)
Shanghai, Munich, and Vienna. Across these global stages, the central theme has been autonomous driving and AI.
At The Autonomous, held in Vienna’s Hofburg Palace, Bosch Mobility CTO Mathias Pillin declared: “Bosch is now an AI company.”
The significance of his keynote lay in the fact that Bosch is not just a visionary but a leading technology player actively enabling autonomous driving on the foundation of real-world safety. Bosch has already moved Level 2++ systems into mass production in both China and Europe, and is now working to extend those capabilities beyond Level 3. Pillin also revealed that Bosch plans to publish a paper early next year on a mathematical approach to tackling the core challenges of autonomous driving, saying he is eager to see the reactions from both academia and industry. At the same time, he emphasized that AI alone will not deliver full autonomy, insisting that broad, industry-wide collaboration is the true breakthrough path forward. Bosch’s journey of End-to-End AI systems in China and Europe, along with its middleware-based SDV strategy, resonated like a future vision built upon tradition.
By Sang Min Han, han@autoelectronics.co.kr
한글로보기
Just before stepping onto the keynote stage, Pillin briefly spoke about Bosch’s technologies, pointing to models unveiled at Munich’s IAA and Auto Shanghai—from BMW and Mercedes-Benz to Volkswagen and Chinese OEMs. Panoramic Vision was one highlight, but the real spotlight was on Bosch’s ADAS and AI capabilities. At The Autonomous, Pillin centered his keynote on how Bosch’s story in autonomous driving is increasingly an AI story.
“Autonomous driving still requires us to address and overcome many challenges. But rather than simply pointing out problems, I want to share Bosch’s perspective on how we can solve them,” Pillin began.
Bosch’s first AI-enabled product was shipped 15–20 years ago, in the form of a smart camera. At the time, the terms “machine learning” and “computer vision” dominated. A turning point came in 2015, when then-CEO Volkmar Denner founded the Bosch Center for Artificial Intelligence (BCAI). Since then, Bosch rapidly expanded its AI capabilities, forming partnerships with universities worldwide and sponsoring professors. Today, Bosch works with the latest technologies such as Transformers and large language models (LLMs).
Another key milestone came in 2020, when Bosch abandoned traditional machine learning and computer vision in smart cameras, shifting decisively to deep learning. Previously stuck as a “second place” competitor, Bosch leapt to the front of the market with its 3rd and 4th generation cameras, which delivered higher performance with improved cost efficiency.
“We now call ourselves an AI company,” Pillin said. “By 2021 alone, Bosch had published over 800 papers at leading AI conferences such as CVPR, NeurIPS, and ICLR. Our teams want to show the world that Bosch’s DNA is deeply embedded with AI, and we are not only publishing but also winning recognition at these venues.”
So where does Bosch’s AI technology stand today? The answer lies in its activities across China and Europe.
Three years ago, as AI began to enter the market in earnest, Bosch faced significant pressure but secured a customer project with the daunting requirement of bringing a new system into production in just 18 months. The company quickly assembled partners and integrated its capabilities into the vehicle.
“As a result, the first SOP (Start of Production) in China was achieved in early 2024. It was a so-called ‘2-Stage End-to-End (E2E)’ AI system. Today, this system has been expanded across multiple platforms in China, and this September we launched the Stage 1 E2E model. I am proud to say this success placed Bosch among the top three in China,” Pillin said.
In May, Bosch brought the Stage 1 prototype to Germany, trained it for only a few hundred hours, and demonstrated it at the IAA Motor Show. Almost every major OEM representative attended, and the performance—its driving behavior and complex situation handling—was described as nothing short of remarkable.
The driving style, rooted in its Chinese origins, remained somewhat more aggressive than German drivers would prefer, occasionally surprising with bold maneuvers. But the performance itself was undeniably impressive. The truly significant point, however, was that what once required years of effort and billions of dollars to transfer an AI system across regions, now required only a few hundred hours of training.
Bosch has since transferred this experience directly to Europe, accelerating ADAS software development through its partnership with Volkswagen Group’s CARIAD and the Automated Driving Alliance.
“The technology package we apply with Volkswagen is identical. It’s already in vehicles, and we are expanding fleets to gather data. Working with one of the world’s largest automakers to collect and validate data is hugely meaningful. We are confident we can be SOP-ready by mid-next year,” Pillin explained.
Bosch’s ability to run two completely independent SoC-based developments simultaneously in China and Europe came down to its software-defined approach. Initially, customers demanded different chip vendors—NVIDIA on one side, QUALCOMM on the other—causing confusion. But Bosch turned this into an asset, learning to leverage AI accelerators and high-performance memory across both. Today, Bosch can deploy its AI software on NVIDIA, QUALCOMM, Horizon Robotics, Ambarella, and more.
“The decisive factor enabling this speed was our robust software-defined middleware interface. This middleware is a product available on the market, and it gave us the foundation to scale our capabilities globally,” Pillin said.
What does Bosch’s ADAS/autonomous driving technology actually look like?
For the past 15–20 years, driver assistance architectures were largely rule-based: perception modules behind sensors, data fusion, and planning layers. These systems often delivered limited performance. Today, the key lies in AI and end-to-end learning.
“End-to-end learning is one of the hottest debates in both industry and academia. Some argue for a fully monolithic, single-stage approach. But in the automotive context, I firmly believe structured architectures are essential,” Pillin explained.
Bosch’s system still consists of the classic blocks—perception, fusion, and planning—but the way they are trained makes all the difference. They can be trained separately, cyclically in loops, or fully end-to-end. Interfaces between modules adapt dynamically in this process. In other words, Bosch proposes an end-to-end learning system with structured, hybrid architecture.
“The structure is decisive. When something goes wrong on real roads, you must be able to trace, validate, and secure homologation. That requires structure. This technology is already on the road, and we are learning from it,” Pillin said.
The next step is to integrate generative AI into the end-to-end stack. Specifically, Bosch is adding Vision-Language Models (VLMs) into the middle layers. VLMs are valuable because they generate semantic information.
For example, when encountering a truck towing a trailer on the highway, traditional perception systems often struggle to detect this as a special situation. A VLM, however, can infer semantic meaning and interpret context.
The combination of E2E modules and VLM effectively adds a layer of “world knowledge.”
.jpg)
But VLMs will not solve everything. The interaction between E2E architectures and VLMs remains unresolved, and Bosch is aware of that. Still, Bosch plans to deploy this technology on real roads next year.
“People call this ‘hallucination.’ Generative AI models face three intrinsic limits: scaling, emergence, and alignment,” Pillin cautioned.
Scaling: Supervised learning demands virtually infinite data for proper training. Reinforcement learning requires infinite test cases to guarantee correct behavior. Either way, infinite effort is needed, even with generative AI. Moreover, these models are built on discrete tokens, while the real world is continuous. That’s a fundamental mismatch.
Emergence: AI models sometimes develop surprising internal behaviors—positive ones, like focusing attention on useful vehicle features, but also negative ones that could cause erratic or unsafe decisions.
Alignment: Training with prompts does not guarantee control. Different prompts can lead to bizarre or unintended behaviors, reflecting a mathematical limitation of these models.
“Yes, Level 3 or even Level 4 is possible. But Level 3 ODDs must be extremely limited, and immense effort is required to ensure reliable operation within them. Level 4 comes with back operators. AI alone will not make full autonomy possible,” Pillin stated.
So how can the industry move from Level 2++ to Level 3 and Level 4? What solutions are being discussed?
Pillin cited Yann LeCun, Meta’s Chief AI Scientist and Turing Award laureate, pointing to world model integration as a potential path forward. It’s mathematically sophisticated, though still unproven in practice.
Bosch’s own breakthrough lies in reducing complexity: cutting the dimensionality of test problems and embedding physics into models, thereby challenging the notion that “autonomous driving involves infinitely many variables that cannot all be validated.”
“I do not believe in an ‘open world problem’ for autonomous driving,” Pillin argued. “We applied a trick developed 20 years ago in theoretical robotics to driving spaces. This mathematical approach fundamentally reduces complexity in testing autonomous vehicles. We plan to publish the results early next year, and I am very curious to see how academia and the automotive industry will react.”
Today, every company entering the Level 2+ and Level 2++ journey is pouring enormous resources into infrastructure and data. According to Pillin, the industry must cooperate on shared data, data processing, tooling, certification, and explainable AI. Initiatives like The Autonomous represent a foundation for this collaboration.
At the same time, differentiation remains. For Bosch, the advantage is delivering Level 2++ and Level 3 systems quickly, expanding ODDs, and bringing vehicles to market in record timeframes—sometimes as short as 18 months.
“I don’t believe we’ll need to wait until 2050 for these technologies to hit the road. If we solve the current challenges correctly, it will happen much sooner. But it’s critical not to mislead ourselves into thinking that simply adding VLMs solves everything. That’s not the case,” Pillin concluded.
AEM(오토모티브일렉트로닉스매거진)
<저작권자 © AEM. 무단전재 및 재배포 금지>
