Foreword
Edited by Zhang Yongwei, Chairman of the China EV100 Research Institute, "AI Revolution in the Automotive Industry: From Product Transformation to Underlying Capability Building" is grounded in the era of deep integration between automobiles and AI (Artificial Intelligence). Through a systematic framework of five chapters—"Overview, Products, Enterprises, Elements, and Outlook"—it deeply analyzes the comprehensive impact of AI technology on the automotive industry. This book is suitable for professionals engaged in intelligent vehicles and their upstream and downstream industries, as well as technology and industry decision-makers, and serves as a reference for students and readers in related fields.
Section 4: Autonomous Driving Data Closed-Loop
1. Strategic Positioning of Data in the Development of Autonomous Driving
Autonomous driving represents the largest application scenario for AI technology. It has now entered an era of "equal access," with a market scale exceeding 300 billion RMB. It possesses a strong driving effect on cutting-edge AI technological innovation and application, and is expected to leverage a 0.1% GDP growth in the future. As autonomous driving shifts from "algorithm-driven" to "data and model collaborative-driven," high-quality data has become the fuel for the evolution of algorithms and models, as well as a core strategic asset for enterprises. A large proportion of automotive data is unstructured, and traditional processing methods cannot meet the requirements, necessitating the construction of software platforms and toolchains adapted to data-driven development. The data closed-loop has also been upgraded from an early "supporting capability" to "underlying infrastructure." Only by taking the lead in connecting and standardizing the closed-loop chain of "collection—screening—annotation—training—deployment—operation—feedback optimization" to form a "data flywheel effect" can we seize the initiative in autonomous driving competition and gain the upper hand in the broader spatio-temporal intelligence competition in the next stage.
The data closed-loop propels autonomous driving from mere technological innovation toward the construction of a new systematic competitiveness. At the technical level, the data closed-loop connects real-world roads, environments, driving behaviors, and in-vehicle sensor information, transforming them into a structured and manageable spatio-temporal sample library. It closely integrates with cloud-based training platforms to provide a continuous supply of scenario nutrients for algorithms. This closed-loop chain from "collection—screening—annotation—training—deployment—operation—feedback optimization" not only enhances data effectiveness but also promotes the formation of a "data flywheel effect," winning a head start for enterprises in the autonomous driving competition.
At the business level, the data closed-loop is not merely a supporting facility for existing technical architectures, but a critical prerequisite determining whether the next-generation autonomous driving technology forms can be sustainably implemented. New network architectures such as BEV (Bird's Eye View) and end-to-end models require the construction of scenario understanding under unified spatio-temporal coordinates and physical semantics, which demands high consistency and standardization from the data closed-loop.
At the service level, autonomous driving technology is no longer a "one-time consumable," but a service capable of iterative innovation and continuous renewal. The core of autonomous driving competition lies not only in who can launch L2+/L3 level autonomous driving functions first, but also in who can master more high-quality, sustainably reusable data assets, along with the corresponding data production and governance capabilities.
2. Market Scale and Marginal Cost of Autonomous Driving Data Closed-Loop
The value chain of the autonomous driving data closed-loop is relatively long. It is not a single business but forms a complete "data value chain" around multiple links in the entire data lifecycle. In the data collection link, a standard data collection vehicle includes a 128-line LiDAR, 11 cameras, high-precision inertial navigation, and other equipment, with hardware costs of approximately 300,000 to 700,000 RMB. In the data annotation link, the unit price for 4D annotation is about 0.5 to 0.6 RMB per element. In the data transaction link, normal cases cost about 1 RMB per frame, while corner cases can reach 5 RMB per frame. High-quality edge scenario libraries (such as 60 seconds of data before takeover) carry a scarcity premium. Once the data platform is built, road-collected data can feed back into the simulation library. Using simulation synthesis, infinite scenarios can be generated at near-zero cost, causing marginal costs to drop sharply.
The autonomous driving data closed-loop features "high fixed costs and low marginal costs." As scenarios such as urban NOA (Navigate on Autopilot), long-haul trunk lines, and industrial parks/mining areas gradually achieve mass production, the demand for training data will grow exponentially on a PB scale, demonstrating significant economies of scale in the data closed-loop. The same road and the same type of scenario can be reused multiple times and across domains among different customers, causing the marginal cost of "adding a piece of data or a scenario" to decrease in a stepped manner. Its economics are highly related to data characteristics: the initial construction costs for standardized collection, clip structures, and annotation systems are relatively high; however, after completion, as standard samples, toolchains, and quality control systems gradually stabilize, the marginal data costs for new scenarios and new vehicle models will drop rapidly. This characteristic of "high fixed costs and low marginal costs" makes the data closed-loop naturally suitable for standardized products and subscription models.
High fixed costs are mainly reflected in the "initial construction" phase, namely the early setup stage, which primarily includes platform construction, standard system design, infrastructure construction, and compliance system building. After the above fixed investments are completed, the data closed-loop enters a stable operation phase. New vehicle models can reuse existing collection modalities, requiring at most minor adjustments and calibration without redesign; new scenarios can be expanded based on the existing scenario system, only needing a small number of new labels for long-tail or special working conditions; the efficiency of AI pre-annotation and human-machine collaboration in the annotation link is steadily improving, gradually reducing the manpower required to complete the same tasks; governance and quality control rules are embedded into the toolchain, achieving "design once, reuse multiple times," thereby reducing repetitive labor; data reuse can be conducted across projects and customers, where one road or one type of scenario can serve multiple automakers and algorithm teams. Therefore, the unit cost of newly added data will drop rapidly, and the speed of decline is usually faster than that in traditional labor-intensive industries.
3. Evolution Roadmap of Autonomous Driving Data Closed-Loop
Given the growth of autonomous driving penetration rates, the increasingly mature business models, and the continuously improving policy environment, the data closed-loop will continue to expand across the entire chain of collection, annotation, governance, evaluation, and subscription, evolving from an "outsourced tool" to "infrastructure."
Stage 1: Outsourced Tools. In the early days of the industry, automakers and algorithm companies outsourced collection, annotation, and some data governance processes to service providers, focusing mainly on delivery capabilities and unit prices. The characteristics of this stage are: project-based, labor-intensive, lacking standards, and low reuse rate.
Stage 2: Platform-based Services. As projects increase and experience accumulates, some enterprises begin to precipitate toolchains, standards, and middle-office systems, gradually productizing data and services to form reusable clip structures and scenario libraries. In this stage, obvious head effects and Matthew effects begin to appear in the market: participants with closed-loop and platform capabilities can continuously acquire large clients and long-term contracts with lower marginal costs, better quality, and superior delivery experiences.
Stage 3: Infrastructure. When data closed-loop capabilities are integrated into the overall R&D and operation systems of automakers, and docked with regional data policies, industry standards, and data element markets, it is no longer just "the capability of a certain company," but gradually evolves into the "infrastructure" of the industry, becoming the public foundation for autonomous driving and even broader spatio-temporal intelligence. Through subscription and transaction mechanisms, data can be transformed into priced, mortgageable, and configurable assets, which is expected to promote the compliant circulation of automotive data elements in the future.
4. Analysis of Autonomous Driving Data Closed-Loop Toolchain
A single autonomous driving vehicle generates TB-level unstructured data (such as images, point clouds, and videos) daily, with varying data quality. Facing hundreds of millions of frames of data, relying on manual annotation is inefficient, and traditional graphics engines struggle to simulate complex lighting, weather conditions, and sensor noise. Only by reconstructing the full-chain data toolchain into a "production system" that can be automatically scheduled, observed, and measure ROI, can the data closed-loop truly shift from "tangible" to "useful" and from "operable" to "scalable." This requires solving the unified definition, management, and measurement of data within the system, connecting the output of each link with its preceding and subsequent links to form an automated and traceable chain, and linking data investment with model performance and business metrics, thereby building an optimizable decision-making closed-loop.
Unified collection and annotation protocols are the starting point for the reconstruction of the entire toolchain. The first step is to clarify sensor configurations, including the number of cameras, resolution, frame rate, exposure strategy, LiDAR line count, deployment positions, etc., and to unify time synchronization and coordinate system definitions, such as the transformation relationships among vehicle body coordinates, world coordinates, map coordinates, and BEV coordinates; simultaneously, "access standardization" is applied to data from different vehicle models, different regions, and different suppliers. This ensures that all subsequent annotation, governance, training, and evaluation can be conducted on a common "physical language" without the need to customize a separate process for each collection scheme.
The second step is to unify the annotation protocol, including unifying the definitions of target categories, attribute categories, behavior categories, and relationship categories, establishing a unified semantic ontology for lane lines, drivable areas, road surface structures, dynamic targets, etc., and formulating consistent annotation principles and conflict resolution rules applicable to multi-tasks such as 2D/3D/BEV/MapTR/OCC. With a unified annotation protocol, the same road and the same scenario can be interpreted as "the same set of semantic facts" under different tasks and different network structures, thereby avoiding the situation of "2D and 3D annotations conflicting with each other," and laying the foundation for subsequent "multi-task training from the same source." In this way, data assets possess reusability and tradability.
Building a multi-task unified-source annotation platform can transform "dispersed operations" into "unified production." Above the protocol, the true landing carrier is a tool platform supporting multi-task unified-source annotation, which can complete the annotation and quality inspection of multiple tasks for the same clip and the same trajectory under the same coordinates and time sequence. It ensures that tasks such as 2D detection/segmentation/instance segmentation, 3D point cloud detection/tracking, BEV semantic grid/occupancy network, 4D lane lines/topological structure, and temporal behavior annotation/interaction intention annotation are processed on the same underlying data foundation.
On this basis, through a unified task management and work order circulation mechanism, a closed-loop process from assignment, annotation, review, spot check to rework is realized, supporting the automatic selection of annotation modes such as fully manual, AI pre-annotation + human-machine collaboration, and template reuse based on scenario difficulty and task type. Furthermore, unified operational specifications and online training systems are used to reduce personnel switching costs.
Finally, the results of all tasks are stored uniformly in the same data warehouse with "clip + scenario + version" as the basic management unit, ensuring that a single target object maintains ID-level consistency across multiple tasks, facilitating subsequent multi-task joint training and alignment.
Establish an active learning and sample filtering engine to focus on high-value data. In the data closed-loop, the truly expensive costs are not storage and computing, but annotation quality and governance. It is necessary to establish an active learning and sample filtering engine oriented towards long-tail scenarios, focusing model capabilities on the most intractable long-tail and extreme scenarios in real road conditions. AI large models can be utilized to filter high-value scenario data, and a scenario and strategy-based sample management mechanism can be established to tilt limited budgets towards scenarios that truly affect safety and experience, avoiding "a large number of repeated simple scenarios" squeezing resources. Meanwhile, dynamic management of the sample library drives the adjustment of collection and annotation strategies, forming a continuously iterating optimization closed-loop.
Build a quality control and audit middle-office to create an auditable data quality system. In the early stage dominated by the project system, data quality control highly relies on "senior quality inspectors + subjective experience," making it difficult to achieve scaling and standardization. As autonomous driving enters the era of data closed-loop and assetization, with data scale rising exponentially, it is necessary to build a quality control and audit middle-office to ensure data credibility and regulability. First, refined management is realized through a multi-level quality control mechanism. The spot check ratio is automatically set based on task type, annotator capability, and scenario complexity. Weighted re-inspection is implemented for key samples such as model feedback samples, high-frequency misjudgments, or long-tail corner cases, and cross-review and double-blind mechanisms are introduced for high-value data, significantly improving annotation consistency.
Second, all quality inspection behaviors are structurally precipitated, automatically generating three types of key profiles: annotator capability profile, single dataset quality profile, and overall project quality profile, providing a data basis for resource scheduling and risk warning.
Third, establish a quality indicator system and SLA (Service Level Agreement) management mechanism matching business forms. Core indicators such as accuracy, consistency, missed detection rate, and error type distribution are clearly defined and embedded into project contracts or subscription service level agreements, forming quantifiable, monitorable, and accountable service commitments. Once a quality anomaly triggers a threshold, the system will automatically warn and drive rapid tracing and repair through the work order mechanism.
Finally, at the audit and compliance level, the middle-office generates complete and tamper-proof operation logs for every piece of data and each of its modifications, achieving full-chain tracing for data sources, processing flows, quality inspection records, and usage trajectories. This meets the internal audit needs of automakers and provides transparent, searchable, and verifiable evidence chains for third-party evaluation institutions and government regulatory departments.
5. Embodied Intelligence and Low-Altitude Economy Expected to Become New Market Increments for Data Closed-Loop
The so-called "pan-autonomous driving" industry refers to intelligent agent applications relying on core technologies such as multi-sensor fusion, real-time decision-making control, and autonomous navigation. Besides intelligent vehicles, it also includes unmanned logistics equipment, embodied robots, and eVTOL (electric Vertical Take-Off and Landing) aircraft. These fields all belong to "agents moving in the real physical world and interacting with the environment," requiring the perception, modeling, and prediction of the environment in continuous space and time, and then outputting control instructions combined with task objectives, and continuously optimizing through result feedback. They are highly homologous in technical fields such as perception, planning and control, and simulation (scenario reconstruction), and can be reused after technical adaptation.
These fields are also shifting from "code-driven" to "data-driven." Some enterprises have publicly stated their "willingness to purchase low-cost, high-quality data to compress self-collection costs." However, their data dilemmas are more prominent: first, less data volume. The scale of unmanned fleets is far smaller than that of passenger vehicles, and embodied robots are in the early stage of development with even more limited scale. Second, more data demands. The operating environment needs to cover complex scenarios such as indoor-outdoor switching, human-machine interaction, and narrow passages. Third, greater difficulty in data acquisition and processing. The robot's action space is continuous, and annotation requires 6D pose + semantic information, far exceeding 2D box/3D box annotation. Currently, the robotics field heavily relies on foreign datasets. Data shortage has become a long-term bottleneck for the development of pan-autonomous driving. Therefore, building a data closed-loop suitable for pan-autonomous driving is a prerequisite for the industrialization and scaling of the industry.
As robots enter factories, supermarkets, and communities, and drones penetrate logistics, inspection, and low-altitude travel fields, these emerging industries also need to build a standardized data closed-loop from data collection and annotation to model training and feedback optimization around typical task scenarios. The clip structures, BEV modeling frameworks, long-tail scenario filtering mechanisms, and data assetization and subscription models accumulated during the autonomous driving stage are expected to be applied to the fields of embodied intelligence and low-altitude economy after transfer or modification, forming a cross-domain flywheel effect of "one set of technology, multi-industry reuse." On the one hand, this can amortize underlying investments and reduce the trial-and-error costs for new industries; on the other hand, through cross-scenario transfer learning, models can continuously evolve in richer spatio-temporal environments, which in turn improves the perception and decision-making levels of autonomous driving itself, thereby constituting a broader spatio-temporal intelligence data closed-loop ecosystem.