Intelligent Driving Simulation and World Model Research Report, 2025
1. The world model brings innovation to intelligent driving simulation
In the advancement towards L3 and higher-level autonomous driving, the development of end-to-end technology has raised higher requirements for the scale of high-quality data, coverage of diverse scenarios, assurance of physical realism, synchronized generation of multimodal data, rationality of behavioral logic, and improvement of iteration efficiency.
Among the three core elements of high-quality intelligent driving (data, models, and computing power), the quality and quantity of scenario data are becoming key differentiators in the intelligent driving experience. Meanwhile, training high-level advanced driver-assistance system (ADAS) algorithm models requires tens of millions of video clips and the generation of long-sequence multimodal driving scenarios. However, the long-tail scenarios captured in real-world road data are relatively limited and cannot meet the demand for high-quality data to feed end-to-end algorithm training.
Automated simulation testing is becoming a powerful tool for OEMs and suppliers to shorten development cycles, reduce costs, improve efficiency, address insufficient coverage of long-tail scenarios, and overcome challenges in reproducing high-risk operating conditions. At the same time, world models, which can understand the physical characteristics and spatial attributes of the real-world environment, are being adopted by an increasing number of OEMs and leading Tier1 suppliers.
Currently, for intelligent driving training, scenario data mainly comes from the following sources:
One is simulation technology based on the replay of real road test data, with the advantage of high scenario authenticity, primarily used to reproduce road test problem scenarios and verify the effectiveness of algorithm fixes;
The second is artificially defined parametric scenarios (such as OpenScenario format), characterized by standardized testing, exploration of boundary conditions, and strong scenario controllability;
The third involves converting real road test data (logsim) into generalizable virtual simulation scenarios (Worldsim), with the core function being data-driven scenario generation and generalization, building a high-confidence simulation scenario library, supporting scenario derivation and automated testing, thereby improving scenario coverage efficiency.
The fourth is the World Model, which uses AI to construct an internal representation of the physical world, enabling intelligent models for environmental state prediction and counterfactual reasoning. Its data sources include multimodal data (images, text, physical rules) and reinforcement learning-generated data. Its advantages include causal reasoning capabilities and support for unknown scenario prediction. However, world models require significant computational resources, their interpretability needs improvement, and they also carry the risk of data bias.
World models also demonstrate advantages in multiple aspects, such as environmental perception and understanding, prediction of future scenario evolution, decision and planning optimization, data generation and training enhancement, simulation and test validation, and improvement of system generalization capabilities. The following table provides a glimpse of the innovation world models bring to intelligent driving training through case studies of typical OEMs and Tier1 suppliers applying world models.
2. Full-chain safety validation is driving simulation toward cross-domain integration
Current autonomous driving safety validation has shifted from single-function testing to full-chain closed-loop verification. Simulation technology is breaking through traditional boundaries and moving toward deep cross-domain collaboration, with core drivers including accelerated technological convergence and toolchain integration.
Specifically:
Accelerated Technological Convergence: AI-driven scenario generation is crucial for building high-quality training datasets. For example, DriveDreamer4D and OASIS SIM’s generative AI technologies have improved long-tail scenario generation efficiency by 10 times (e.g., 51Sim generates 32,000 extreme scenarios per day). Meanwhile, multi-domain model collaboration is becoming more prominent, such as vehicle dynamics (PanoCar), sensors (physics-level radar modeling), traffic flow (SUMO/VISSIM), and cloud-based world models (e.g., Li Auto’s MindGPT) working together to build a digital twin closed loop.
Toolchain Integration: Leading solution providers (e.g., Horizon’s AIDI platform, Synkrotron.ai’s OASIS) have achieved full-stack toolchain integration from "perception - planning - control - vehicle-road-cloud," supporting seamless transitions from MIL to VIL. For instance, Horizon’s UniAD framework uses an end-to-end model to compress perception-planning latency to around 50ms and validates multi-vehicle gaming strategies in simulation.
Due to the development of cockpit-driving integration and cross-domain integration applications, simulation is also moving toward cross-domain integration. The industry has introduced simulation testing solutions for various automotive domains, integrating software and hardware tools/platforms to actively promote joint cross-domain testing. Overall, OEMs and suppliers are currently advancing cross-domain simulation, mainly focusing on: Intelligent cockpit + intelligent driving integration, Intelligent chassis + intelligent driving cross-domain integration, Three-electric systems (battery, motor, electronic control) + thermal management integration, IoV + intelligent driving integration, global digital twins.
Examples include:
Tsing Standard’s active suspension HIL and Zhejiang PanoSim’s PanoCar conducting suspension-planning co-simulation to address cross-domain control latency and improve stability in extreme conditions (reducing roll by ≥15%).
AUMO (under Alinx Electronic) collaborated with BYD to develop a cockpit-driving domain integration testing solution, using the W50 platform to validate in-cabin vision systems (DMS/OMS) alongside autonomous driving algorithms, enabling data exchange between cockpit and driving domains to accelerate "cockpit-driving integration."
In Q1 2025, Mercedes-Benz partnered with VECTOR to conduct centralized electronic architecture virtualization validation, using SIL Kit middleware for distributed simulation testing of domain controllers (e.g., autonomous driving, body domains) to optimize cross-domain communication and functional integration efficiency.
In October 2024, Beijing Oriental Jicheng and Great Wall Motors collaborated on cross-domain joint testing for intelligent cockpit, intelligent driving, and vehicle connectivity, covering signal-level simulation to full-vehicle testing in a one-stop service.
Kunyi Electronics’ cockpit HIL testing, based on Kunyi’s high-real-time RTPC system and combined with its high-level autonomous driving data closed-loop testing solution, provides simulations for 360° surround view, driver monitoring, and streaming rearview mirrors, meeting high-performance testing needs for vehicle-cloud integration, cockpit-parking integration, and cockpit-driving integration.
3. Industry progress in enhancing simulation credibility
One of the biggest pain points in simulation testing is credibility. The industry needs to consider how to ensure high fidelity in scenario simulation, high accuracy in sensor models, high confidence in dynamics models, as well as challenges in real-time performance, data bandwidth, and stability of data interfaces.
In terms of improving simulation credibility, the following approaches are being adopted.
1. Application of AI Technology
AI technology is gradually being applied to simulation testing in engineering practice, significantly accelerating the automation efficiency of testing and validation, thereby improving automotive development efficiency. For example:
In February 2025, IAE partnered with VDBP to launch the industry's first AI scenario generation tool integrated with the DeepSeek R1 large model. It pioneered an end-to-end solution for "generating high-quality OpenDRIVE and OpenSCENARIO standard scenarios with text commands," supporting intelligent generation from simple ADAS tests to complex traffic rules and extreme working conditions. It covers full-scenario needs such as ADAS, urban NOA, and V2X, improving scenario construction efficiency by 300% and enabling seamless integration with mainstream simulation software like CARLA, VTD, and Prescan.
In December 2024, AVL released the AI simulation assistant ChatSDT to simplify and enhance user interaction with AVL simulation components. MathWorks also introduced the MATLAB Large Language Support Package, aiming to deeply integrate large language models (such as ChatGPT, Qwen, and DeepSeek) with MATLAB/Simulink to improve engineering development efficiency.
2. Open-Source Datasets
Additionally, organizations like the China Association of Automobile Manufacturers (CAAM) are actively promoting open-source data initiatives. Nearly 20 datasets have been released, including Coral Data, vehicle-road-cloud integrated simulation scenario open-source data, OEM-open-sourced end-to-end autonomous driving public datasets, and publicly available training datasets related to intelligent driving world models. The goal of open-sourcing is to facilitate efficient reuse of these high-quality scenario datasets and avoid redundant development within the industry.
In April 2025, the ASAM OpenMATERIAL 3D 1.0.0 standard was officially released. This standard specifies a standardized format for physical material properties and 3D object descriptions, precisely defining parameters such as refractive index, surface roughness, and permeability. By providing accurate and standardized 3D assets and material properties, the standard enhances the realism of perception sensor simulations, making the outputs of LiDAR, radar, and cameras more lifelike.
3. Simulation Tool Upgrades
Simulation testing companies have also updated and upgraded the functions of simulation software tools/platforms, such as PreScan software version 2503, HEXAGON VTD/MSC/ADAMS/KISSoft simulation software, CarMaker14.0, AURELION 24.3, MATLAB/Simulink R2025a, Ansys 2025R1, Oasis SIM 3.0, aiSim intelligent driving simulation software UE5.5 upgrade, Qianxing system V3.0 with 20+ new features, PanoCarV1.7 PanoSim V33 version, etc. (see the report for details).
Intelligent Driving Simulation and World Model Research Report, 2025
1. The world model brings innovation to intelligent driving simulation
In the advancement towards L3 and higher-level autonomous driving, the development of end-to-end technology has raised higher re...
Autonomous Driving Map (HD/LD/SD MAP, Online Reconstruction, Real-time Generative Map) Industry Report 2025
Research on Autonomous Driving Maps: Evolve from Recording the Past to Previewing the Future with "Real-time Generative Maps"
"Mapless NOA" has become the mainstream solution for autonomous driving s...
End-to-End Autonomous Driving Research Report, 2025
End-to-End Autonomous Driving Research: E2E Evolution towards the VLA Paradigm via Synergy of Reinforcement Learning and World Models??The essence of end-to-end autonomous driving lies in mimicking dr...
Research Report on OEMs and Tier1s’ Intelligent Cockpit Platforms (Hardware & Software) and Supply Chain Construction Strategies, 2025
Research on intelligent cockpit platforms: in the first year of mass production of L3 AI cockpits, the supply chain accelerates deployment of new products
An intelligent cockpit platform primarily r...
Automotive EMS and ECU Industry Report, 2025
Research on automotive EMS: Analysis on the incremental logic of more than 40 types of automotive ECUs and EMS market segments
In this report, we divide automotive ECUs into five major categories (in...
Automotive Intelligent Cockpit SoC Research Report, 2025
Cockpit SoC research: The localization rate exceeds 10%, and AI-oriented cockpit SoC will become the mainstream in the next 2-3 years
In the Chinese automotive intelligent cockpit SoC market, althoug...
Auto Shanghai 2025 Summary Report
The post-show summary report of 2025 Shanghai Auto Show, which mainly includes three parts: the exhibition introduction, OEM, and suppliers. Among them, OEM includes the introduction of models a...
Automotive Operating System and AIOS Integration Research Report, 2025
Research on automotive AI operating system (AIOS): from AI application and AI-driven to AI-native
Automotive Operating System and AIOS Integration Research Report, 2025, released by ResearchInChina, ...
Software-Defined Vehicles in 2025: OEM Software Development and Supply Chain Deployment Strategy Research Report
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, ...
Research Report on Automotive Memory Chip Industry and Its Impact on Foundation Models, 2025
Research on automotive memory chips: driven by foundation models, performance requirements and costs of automotive memory chips are greatly improved.
From 2D+CNN small models to BEV+Transformer found...
48V Low-voltage Power Distribution Network (PDN) Architecture and Supply Chain Panorama Research Report, 2025
For a long time, the 48V low-voltage PDN architecture has been dominated by 48V mild hybrids. The electrical topology of 48V mild hybrids is relatively outdated, and Chinese OEMs have not given it suf...
Research Report on Overseas Cockpit Configuration and Supply Chain of Key Models, 2025
Overseas Cockpit Research: Tariffs stir up the global automotive market, and intelligent cockpits promote automobile exports
ResearchInChina has released the Research Report on Overseas Cockpit Co...
Automotive Display, Center Console and Cluster Industry Report, 2025
In addition to cockpit interaction, automotive display is another important carrier of the intelligent cockpit. In recent years, the intelligence level of cockpits has continued to improve, and automo...
Vehicle Functional Safety and Safety Of The Intended Functionality (SOTIF) Research Report, 2025
Functional safety research: under the "equal rights for intelligent driving", safety of the intended functionality (SOTIF) design is crucial
As Chinese new energy vehicle manufacturers propose "Equal...
Chinese OEMs’ AI-Defined Vehicle Strategy Research Report, 2025
AI-Defined Vehicle Report: How AI Reshapes Vehicle Intelligence?
Chinese OEMs’ AI-Defined Vehicle Strategy Research Report, 2025, released by ResearchInChina, studies, analyzes, and summarizes the c...
Automotive Digital Key (UWB, NearLink, and BLE 6.0) Industry Trend Report, 2025
Digital key research: which will dominate digital keys, growing UWB, emerging NearLink or promising Bluetooth 6.0?ResearchInChina has analyzed and predicted the digital key market, communication techn...
Integrated Battery (CTP, CTB, CTC, and CTV) and Battery Innovation Technology Report, 2025
Power battery research: 17 vehicle models use integrated batteries, and 34 battery innovation technologies are released
ResearchInChina released Integrated Battery (CTP, CTB, CTC, and CTV)and Battery...
AI/AR Glasses Industry Research Report, 2025
ResearchInChina released the " AI/AR Glasses Industry Research Report, 2025", which deeply explores the field of AI smart glasses, sorts out product R&D and ecological layout of leading domestic a...