PUBLISHER: ResearchInChina | PRODUCT CODE: 1721398
PUBLISHER: ResearchInChina | PRODUCT CODE: 1721398
SDV Research: OEM software development and supply chain deployment strategies from 48 dimensions
The overall framework of software-defined vehicles: (1) Application software layer: cockpit software, intelligent driving software, vehicle control software, and AI empowerment, etc.; (2) Functional software layer: cloud services, security services, etc.; (3) System software layer: vehicle OS, middleware and SOA, etc.; (4) R&D tools: process and systematization tools, data closed loop, development toolchains, etc.
In this report, we specifically expound the OEM vehicle software development and toolchain framework from 13 subsystems and 48 dimensions in order to analyze the R&D focus, development strategies and supply models of OEMs, as shown in the following table:
Autonomous driving software algorithm: Compare automotive VLMs and VLA models with cloud world models, and fully mobilize supply chain resources to assist development
For example, Li Auto uses multimodal foundation models and generative AI technology to enable autonomous driving. The key technical architecture includes: an end-to-end intelligent driving model, VLM, VLA, cloud-based reconstruction and an generative world model.
Li Auto fully integrates academic and supply chain resources to accelerate the development of its autonomous driving system:
Li Auto has proposed an advanced hybrid BEV algorithm framework with Tsinghua University and MIT, and cooperated with Tsinghua MARS Lab to develop the DriveVLM model;
The next-generation architecture MindVLA released in 2025 and the GaussianAD from the University of California, Berkeley optimize scenario modeling and trajectory prediction;
Li Auto cooperates with GigaStudio to explore the construction of 4D world models, introduce the DriveDreamer4D framework, and use world models to enhance the reconstruction effect of 4D driving scenarios;
Keleyuan operates Li Auto's 4D BEV data collection project and is responsible for the annotation, analysis and processing of automotive sensor (such as LiDAR) data to generate high-precision road information data. Its services cover road data collection in multiple cities across the country, supporting the optimization of dynamic perception and prediction capabilities of Li Auto's intelligent driving platform;
Singstor provides a massive data collection and storage solution for autonomous driving with the core being customized automotive high-performance optical fiber high-speed storage devices, supporting real-time data recording and synchronization across regions;
4D automatic annotation system: Li Auto has built a full-process system covering dynamic obstacle detection and tracking, laser/visual SLAM reconstruction, static element annotation and OCC (common obstacle) annotation.
The physical world model framework built by Xiaomi Automobile consists of a data observation layer (Ot), an implicit feature layer (Zt), and a display symbol layer (St):
The data observation layer (Ot) is the input layer of the neural network. The sensor input includes images, LiDAR point cloud, and navigation information required by NOA.
In the implicit feature layer (Zt), the information of the input layer in the previous step is expressed as a private feature through the BEV coding network. Through different decoders, dynamic elements, static elements, and the future trajectory of the vehicle can be obtained respectively.
In the display symbol layer (St), in order to facilitate human understanding and connect with manual rule codes, the model will decode the displayed symbolic expressions, such as static lane lines, zebra crossings, etc., dynamic pedestrians, vehicles, etc., which are also the expression forms of manual or automatic value-added annotation in supervised learning.
This architecture breaks through the fragmentation of traditional modular systems and integrates perception, planning, and control into a single end-to-end model, significantly improving system coherence.
After the model architecture is set up, it is necessary to open up data-driven channels so that the model can automatically output the real required conclusions driven by data to form a data-driven flywheel. The technical iteration of Xiaomi SU7 leverages the closed loop of "production vehicle data - simulation training - OTA updates" to deliver vehicles running on roads and quickly accumulate a complete database of Chinese road scenarios.
In terms of toolchains, Xiaomi has conducted in-depth cooperation with NVIDIA to optimize cloud training and automotive inference:
Cloud: By reconstructing the inference pipeline based on Triton, the efficiency of automatic annotation is improved by 100%; by optimizing the training process through DALI and CV-CUDA, the GPU utilization rate is increased by 30%.
Vehicle: The Thor platform accelerates model reasoning, doubling performance compared to the initial version; it offloads tasks such as image processing and point cloud compression to heterogeneous computing units (VIC, ISP) to alleviate pressure on the main computing power.
Vehicle control software algorithm: Make full use of AI empowerment
The empowerment of AI technology in the development of automotive motion control software is mainly reflected in algorithm optimization, development process innovation, and improvement of system integration capabilities, spurring the industry towards a more intelligent, efficient and safe direction.
Geely Xingrui AI Cloud Power
Geely has developed the Xingrui AI Cloud Power. The Xingrui AI Cloud Power 1.0 mainly realizes AI empowerment in smart energy management, smart motion control, and smart cloud diagnosis.
AI smart motion control: Based on the real-time perception of the drive motor, vehicle posture, camera vision, weather and other information, AI intelligent recognition of driving style, road terrain, and environmental conditions is achieved, and the drive system is dynamically adjusted to perform precise real-time control "according to local conditions", reducing slippage by 50% and improving tracking stability by 15%;
AI smart energy management: Integrating the Xingrui AI cloud power model, AI realizes intelligent decision-making for global optimization, evolving the previous rule control into one that can perceive external temperature, humidity, altitude, slope and other road conditions in real time through AI, and can intelligently adjust the energy management strategy at any time, thereby reducing the fuel consumption of electric hybrid sedans to 2L and electric hybrid SUVs to 3L.
AI smart cloud diagnosis: Based on digital twin technology, power systems, "electric drive, battery and electric control" components and thermal management systems are fully monitored, and automatic repair and proactive maintenance are achieved.
Geely plans to further launch Xingrui AI Cloud Power 2.0 in 2025, which extends battery life by 15% and reduce energy consumption by 5% through model compression technology.
Xpeng AI Chassis Active Perception
Xpeng has added AI chassis active perception and automatic chassis adjustment to its two flagship models - X9 and G9:
The vehicle's sensors preview the road surface and the chassis suspension is adjusted accordingly. The standard intelligent dual-chamber air suspension of Xpeng X9 includes intelligent dual-chamber air springs and intelligent variable damping shock absorbers. It can efficiently filter road vibrations, effectively isolate minor bumps, and offer a gentle and smooth driving experience, taking into account both comfort and controllability.
It utilizes automotive sensors and cloud AI technology. When it determines that the vehicle passed a bumpy point, it immediately uploads relevant information to the cloud. A new bumpy layer is formed on the cloud, allowing the vehicle to detect bumps and potholes ahead earlier, giving the driver more time to operate. At the same time, the stiffness of the suspension will be adjusted before the bump to improve comfort.