Large models speed up the development of autonomous driving algorithms - IDC released the 2024 China Autonomous Driving Development Platform Vendor Evaluation Research Report

Today, decision-making and planning modeling and end-to-end models have become hot topics in the field of autonomous driving, and models based on neural networks have been fully infiltrated in autonomous driving algorithms. Under the trend of autonomous driving modeling, data-driven algorithm upgrades have begun to become an industry consensus, and the efficiency of enterprises in collecting and using data will become the key to their software upgrades and iterations. Therefore, an efficient and integrated autonomous driving algorithm development tool chain/platform has begun to become a key infrastructure to promote the differentiation of autonomous driving products. With the maturity of large model technology, the degree of automation and functional completeness of autonomous driving development platforms have been significantly improved again, and the development of data-driven autonomous driving algorithms has been accelerated.

IDC believes that the autonomous driving development platform is an algorithm development platform for car companies or autonomous driving software providers, which meets the needs of technology buyers to process massive data and update autonomous driving solutions. The platform needs to have:

● Comprehensive processing capabilities such as data collection, labeling, and mining;

● Model training and parameter adjustment;

● Functional modules such as simulation test, etc.

The autonomous driving development platform takes data closed loop as the core, providing enterprises with one-stop data processing, algorithm development and test verification services, helping enterprises to efficiently promote autonomous driving technology, such as the research and development and application of perception modules, decision-making and planning modules, etc.

IDC recently released the report "IDC MarketScape: China Autonomous Driving Development Platform 2024 Vendor Assessment" (Doc#CHC51571424, June 2024). The research has attracted the attention of many technology providers. IDC finally selected 7 representative vendors for in-depth research, including (in alphabetical order): Baidu, Horizon Robotics, Huawei, Volcano Engine, Qingzhou Zhihang, Ruqi Mobility, and Tencent. The vendor evaluation results are displayed in graphical form based on the IDC MarketScape model.

After a detailed market survey, IDC has the following insights into the product development trends of autonomous driving development platforms:

● Data collection and transmission gradually transition from large-scale collection by collection vehicles to long-tail scene collection of mass-produced vehicles: Data has become an important asset, and collecting various perception, decision-making, and environmental data has become an important prerequisite for forming a closed data loop. As the penetration and usage rate of intelligent driving functions in passenger cars gradually increase, the solution of using passenger cars to collect data has begun to be implemented. Although the current method of uploading data by collection vehicles is still mainly based on disk, it can be observed that the collection of mass-produced vehicles will become a trend.

● Explore the potential of Big Models in scene mining/retrieval: In the future, as the penetration and usage rates of intelligent driving functions continue to increase, and as more vehicle-side data is transmitted back in the form of vehicle-cloud intercommunication, the amount of data owned by enterprises will increase exponentially. However, for the development stage, only some long-tail scenes or specific scene data can be used as high-value data and enter subsequent training and other stages. This means that companies need to efficiently screen massive amounts of data. At present, based on the capabilities of Big Models , multi-modal retrieval functions such as text search, image search, and even video search have been launched one after another, helping companies to identify and retrieve high-value data. At present, some leading manufacturers have corresponding functions and access, but most have not yet formed standardized products.

● Automated labeling has become the main labeling solution, and 4D labeling will usher in rapid development: Driven by the development of visual algorithms and the demand for large amounts of data labeling, automatic labeling as a labeling solution of pre-labeling + manual review has become the main solution in the industry. At present, the accuracy of 2D image segmentation or recognition and 3D point cloud automatic labeling has generally reached more than 90%, helping companies to significantly reduce labor costs and improve efficiency. At the same time, the popularization of the Transformer+BEV algorithm has promoted the development of 4D labeling, and IDC has noticed that this function is accelerating its implementation.

● From roadside testing to simulation testing, parallel simulation capabilities and simulation scenarios are top priorities: the number of simulation test kilometers of each company has begun to become an important dimension of the strength of autonomous driving products, and the accumulated simulation test mileage of car companies has begun to move towards hundreds of millions of kilometers. This puts higher requirements on the operating efficiency/resource utilization and scenario coverage of the simulation test system. The key to improving efficiency and computing cluster utilization is to achieve multi-task and multi-scenario parallel testing through containerization/cloud native and other means. In addition, simulation testing is not just a simple accumulation of mileage data. It requires testing of various long-tail scenarios to improve the overall strength of autonomous driving software. To this end, technology providers have taken a multi-pronged approach to broaden the coverage of their scenario libraries. For example, not only extract key scenarios from real road conditions, but also generalize the scenarios to expand the diversity of scenarios. At the same time, manufacturers are actively exploring the generation of scenarios through large models, NeRF and other technologies to make technical reserves for the flexible construction of rare long-tail scenarios.

● Integrated platforms are gradually being connected: Although technology providers are currently able to build integrated platforms, they have not yet achieved full interoperability among modules (data flow). With the improvement of the automation capabilities of each module and the increase in the speed of software iteration, connecting the capabilities of each module will become an important development direction in the future. Only in this way can a true data closed loop be formed to help the rapid upgrade of autonomous driving algorithms.

Advice for technology buyers