Software Defined Vehicles

Interview with Dr Liu Qiang

“The Chinese SDV model tends to favour in-house AI development”

Benjamin Müller
Benjamin Müller International Editor for ADT, aIT, AP & All-Electr.
4 min
Speaker onstage at a congress with a large blue presentation screen and audience silhouettes in the foreground.
Dr Liu Qiang is Vice President of Li Auto and General Manager of Li Auto Germany R&D Center. The PhD graduate in Power Machinery and Engineering from Harbin Institute of Technology previously led power electronics development and industrialisation programmes at UAES before joining Li Auto in 2021.

As SDVs and AI become central to automotive competition, manufacturers face growing pressure to industrialise both at scale. Dr Liu Qiang, Vice President of Li Auto and General Manager of Li Auto Germany R&D Center, explains how AI, SDVs and embodied intelligence are reshaping vehicle architecture.

Software-defined vehicles, foundation models and embodied AI are moving from research topics to strategic priorities across the automotive industry.

At the 30th AUTOMOBIL-ELEKTRONIK Kongress, Dr Liu Qiang, Vice President of Li Auto and General Manager of Li Auto Germany R&D Center, outlined how the Chinese EV manufacturer is approaching this transformation. In his presentation, “From Smart Vehicles to Embodied AI: Building the ‘Silicon-Based Life’ Architecture of the Future”, he described how AI is increasingly shaping vehicle architecture, software development and product definition.

In the following interview, he discusses the biggest challenges in scaling SDVs, the role of foundation models and why embodied AI could become the next major step in automotive intelligence.

ADT: Looking ahead three to five years, what will be the biggest bottleneck in turning SDV and AI strategies into scalable, industrialised vehicle platforms?

The biggest bottleneck is the mismatch between compute development and algorithm iteration. Automotive hardware cycles typically take two to three years, while AI models evolve within weeks or months. This creates a fundamental challenge: neither the full potential of algorithms nor the available computing power is fully utilised. Under traditional, decoupled development approaches, rigid hardware architectures constrain rapidly evolving algorithms, while expensive compute resources remain underutilised without targeted optimisation. This lose-lose situation significantly reduces overall system efficiency.

From Li Auto’s perspective, hardware-software co-design has evolved from a nice-to-have into a necessity. To industrialise SDVs and AI, traditional organisational silos must be broken down. Only through deep alignment and joint optimisation at the foundational level can performance bottlenecks be eliminated, hardware utilisation maximised and the full potential of AI unlocked.

Which decision being made today will most strongly determine where value is created in the future automotive ecosystem?

The decisions that will shape future value creation are concentrated across four technology layers. At the application layer, software focuses on intelligent cabin experiences and autonomous driving. This is where users directly interact with the vehicle and where differentiated customer experiences are created. Our approach centres on Cabin Agents and AD Agents. At the model layer, foundation models act as the core intelligence of the vehicle and define the upper limit of its capabilities. They create a data-driven moat for perception and decision-making. Our approach combines a self-developed general foundation model with a self-developed VLA model for autonomous driving. 

At the system layer, compute and operating systems form the central hub connecting software and hardware. This layer determines overall system efficiency and the depth of hardware-software co-design. Our approach is based on our self-developed SoC series and Halo OS. At the physical layer, hardware provides the foundation for executing advanced AI capabilities. Our approach includes a full drive-by-wire chassis, an 800 V fully active suspension system and full-scenario sensor arrays.

Where do current approaches to SDVs and next-generation E/E architectures still fall short in real-world programmes?

Real-world SDV and E/E architecture deployments face challenges across three dimensions: hardware, software and safety. From a hardware perspective, the industry faces a miniaturisation bottleneck. Expanding electrical functions and the introduction of smart PDUs are increasing PCB integration density. This creates significant packaging conflicts between traditional hardware, limited installation space, harsh operating environments and increasingly complex wiring topologies. Our view is that the industry must move towards higher levels of chip integration and device miniaturisation. New packaging approaches, including 3D packaging and flexible substrates, will also be required to make better use of irregular installation spaces. From a software perspective, the challenge lies in balancing SOA decoupling with communication overhead. Service-oriented architectures face a fundamental dilemma in automotive applications. 

If services are too coarse, software coupling remains high and limits independent OTA updates. If services are too granular, the growing number of service instances overwhelms compute-constrained MCUs through communication overhead, service discovery latency and increased memory requirements. Our approach is to deploy lightweight service frameworks that significantly reduce resource consumption. We also see a transition from static communication matrices towards dynamic service composition. 

In parallel, true app-to-app end-to-end hard real-time connections must be established, enabling a gradual migration towards a fully service-oriented architecture. From a safety perspective, AI introduces a new challenge because its inherent black-box characteristics conflict with the deterministic requirements of standards such as ISO 26262. Our view is that the industry must move beyond ISO 26262 and integrate SOTIF and emerging AI safety standards. Validation should shift from traditional code coverage towards scenario coverage using equivalence testing based on scenario libraries. In addition, a closed-loop data compliance framework and explainable AI technologies are required. We also advocate a dual-channel safety architecture that physically separates AI decision-making from safety arbitration functions.

How is AI shaping vehicle architecture in China today?

AI is reshaping vehicle architecture in China from two perspectives: technical architecture and product definition. From a technical standpoint, the key shift is the use of large AI models as the vehicle’s digital brain. Rather than treating AI as an additional module, manufacturers are integrating AI capabilities directly into the vehicle architecture. Our implementation is based on a unified intelligent architecture built around our self-developed VLA autonomous driving model and our self-developed foundation model as dual cores. Through the underlying system architecture, this enables coordinated management of compute, data and control. 

From a product-definition perspective, AI agents are becoming the central interaction layer of the vehicle. The vehicle is evolving from a system that passively executes commands into one that actively understands user intent. Our implementation reflects this shift. Our autonomous driving architecture operates in the physical world, while our cabin architecture manages digital services. Together, they create a full-scenario intelligent experience in which the vehicle can both handle driving and manage tasks.

What differentiates Chinese SDV development from Europe?

Based on the experience gained at Li Auto’s German R&D centre and through our internal development programmes, we see fundamental differences in strategic priorities between Chinese and European SDV development. The Chinese SDV model tends to favour in-house AI development, enabling rapid closed-loop iteration cycles. It places strong emphasis on user experience, agile development and hardware-software co-optimisation. In contrast, the European SDV model relies more heavily on collaborative AI development through established supplier ecosystems. It places greater emphasis on functional safety and engineering reliability, with a stronger preference for hardware-software decoupling and standardisation.

How does Li Auto approach the concept of embodied AI in vehicles?

Embodied AI means that a vehicle does more than simply execute commands. It perceives its environment, understands user intent and anticipates needs in a way that increasingly resembles human interaction. As the system evolves through daily use, the vehicle becomes more than a means of transport. It becomes an intelligent entity that recognises its users, understands their preferences and proactively provides support. For Li Auto, the vehicle is only the first large-scale application of embodied AI. Our long-term goal is to build a general intelligence system for the physical world. Our approach to embodiment is inspired by human physiology. The vehicle’s technical architecture is designed around five major in-house R&D domains that correspond to core human systems. This enables deep integration across hardware and software and creates the foundation for increasingly capable embodied intelligence.

top stories li auto interview electric vehicle technology ecosystem china architecture ai aek top story software defined vehicles

More software defined vehicles

Our 2026 events