A Novel Multi-layer Task-centric and Data Quality Framework for Autonomous Driving

Abstract: The next-generation autonomous vehicles (AVs), embedded with frequent real-time decision-making, will rely heavily on a large volume of multisource and multimodal data. In real-world settings, the data quality (DQ) of different sources and modalities usually varies due to unexpected environmental factors or sensor issues. However, both researchers and practitioners in the AV field overwhelmingly concentrate on models/algorithms while undervaluing the DQ. To fulfill the needs of the next-generation AVs with guarantees of functionality, efficiency, and trustworthiness, this paper proposes a novel task-centric and data quality vase framework which consists of five layers: data layer, DQ layer, task layer, application layer, and goal layer. The proposed framework aims to map DQ with task requirements and performance goals. To illustrate, a case study investigating redundancy on the nuScenes dataset proves that partially removing redundancy on multisource image data could improve YOLOv8 object detection task performance. Analysis on multimodal data of image and LiDAR further presents existing redundancy DQ issues. This paper opens up a range of critical but unexplored challenges at the intersection of DQ, task orchestration, and performance-oriented system development in AVs. It is expected to guide the AV community toward building more adaptive, explainable, and resilient AVs that respond intelligently to dynamic environments and heterogeneous data streams. Code, data, and implementation details are publicly available at: https://anonymous.4open.science/r/dq4av-framework/README.md.

• The paper presents a modular five-layer framework that aligns multisource data quality dimensions with specific AV tasks to optimize detection performance and efficiency.
• It validates the framework with redundancy analysis on nuScenes data, showing that targeted redundancy pruning can maintain or improve mAP50 and recall in object detection.
• The study demonstrates practical benefits including reduced computational load, adaptive robustness under adverse conditions, and enhanced system explainability in AV design.

A Multi-layer Task-centric and Data Quality Framework for Autonomous Driving

This paper introduces a modular, five-layer framework designed to systematically align data quality (DQ) evaluation with task, application, and goal requirements in autonomous driving (AD) systems. The authors emphasize the necessity of moving beyond model-centric paradigms, advocating instead for a structured, data-quality-aware approach to enhance reliability, efficiency, and interpretability in real-world AV deployments.

Framework Architecture and Rationale

The proposed Vase Framework comprises the following layers:

1. Data Layer: Encompasses multisource (e.g., multiple cameras, LiDAR, radar, infrastructure) and multimodal (e.g., images, point clouds, text, GPS) sensor data. This layer acknowledges the heterogeneous, overlapping, and potentially redundant characteristic of raw AV data.
2. Data Quality (DQ) Layer: Introduces critical, context-specific DQ dimensions, including completeness, consistency (cross- and temporal), correctness, noise levels, redundancy, relevance, and timeliness. The DQ assessment is tailored to both the nature of the data and the downstream application.
3. Task Layer: Recognizes that different AV tasks (e.g., object detection, tracking, trajectory prediction, planning) impose unique DQ requirements, and mandates that DQ evaluation and amelioration must be task-centric.
4. Application Layer: Aggregates tasks into functional AV pipelines (e.g., perception, prediction, planning, control), facilitating a mapping from task-specific DQ needs to end-to-end application-level requirements.
5. Goal Layer: Articulates high-level objectives such as safety, latency, accuracy, and efficiency, and provides system-level feedback for iterative improvement of the preceding layers.

The framework structure enables clear traceability from operational goals down to data acquisition, supporting iterative refinement and modular system upgrades.

Task-centric Redundancy Analysis: Empirical Case Study

To substantiate the framework, the authors conduct a series of experiments centered on the redundancy DQ dimension within the object detection task using the nuScenes dataset. The methodological approach consists of:

• Sensor Overlap Analysis: By identifying overlapping fields of view between pairs of nuScenes cameras, the authors define redundancy at the instance level based on spatial overlap and a Bounding Box Completeness Score (BCS).
• Controlled Redundancy Pruning: Training YOLOv8 detectors on datasets with varying levels of redundancy demonstrates that judicious redundancy removal has negligible or even positive effects on mean average precision (mAP50) and recall.
• Multimodal Redundancy Evaluation: By comparing detections from LiDAR-only and LiDAR-camera fusion, and systematically pruning LiDAR data based on spatial proximity, the study quantifies the cross-modal redundancy and its impact on model performance.

Key Quantitative Findings

• On multisource image data, removing redundant object instances through BCS-guided selection maintained or improved detection performance relative to using the full, more redundant training set.
• In multimodal experiments, high redundancy was localized to objects near the AV (closer distances), and pruning near-field LiDAR detections did not degrade performance up to an empirically determined threshold.
• These results support the bold claim that, in AV systems, data quantity is not always beneficial—appropriate, task-informed data curation can yield smaller, more effective training sets and lower computational overhead without accuracy loss.

Implications for AV System Design

Practical Implications

• Operational Efficiency: The framework enables data selection strategies that reduce computational footprint, storage, and inference costs, critical for edge-deployed AV systems with hardware constraints.
• Adaptive Robustness: By incorporating dynamic DQ monitoring and feedback loops across application layers, the system can adjust to adverse conditions (e.g., sensor occlusion, weather degradation) and maintain task performance.
• Explainability and Debuggability: Tightly coupling DQ evaluation with task outcomes improves traceability, enabling better diagnosis of performance failures and more transparent AV decision-making pipelines.

Theoretical Significance

• The framework formalizes the many-to-many interdependencies between DQ metrics, task specifications, and system goals in AV. This provides a foundation for the development of task-conditioned DQ taxonomies, mathematical trade-off models, and adaptive metric selection methods.

Open Research Directions

The authors identify several research questions that remain to be addressed for comprehensive DQ-aware AV system development:

• Task-specific DQ Taxonomies: How to standardize and systematically map DQ dimensions to specific AV tasks and modalities.
• Adaptive and Dynamic Metric Selection: Mechanisms for real-time adjustment of DQ metrics according to environmental context and mission-critical goals.
• Latency–Efficiency–Quality Trade-offs: Optimization algorithms to balance data volume, processing latency, and DQ for safety-critical operations.
• Integration with XAI and Human-in-the-Loop Systems: Leveraging DQ metrics to enhance the interpretability of AV decisions and facilitate collaborative, semi-autonomous control workflows.

Prospects for Future AI Systems

This paper's framework anticipates a paradigm shift in AV development toward data-centric engineering, where improvement cycles prioritize DQ remediation and task alignment over mere model complexity escalation. There is strong potential for extending this methodology to:

• DQ-aware simulation and digital twinning for robust scenario testing and certification.
• Reinforcement learning agents that optimize DQ-task-goal alignment in real time.
• Practical edge deployments in distributed, federated, and privacy-preserving AV systems.

Conclusion

The proposed Vase Framework provides a formal, multi-layered structure to design, implement, and iteratively refine autonomous driving systems where data quality is a first-class concern. The empirical analysis on redundancy in multisource and multimodal data substantiates that task-centric data curation can reduce system resource demands without sacrificing—and occasionally improving—model accuracy. This work lays the foundation for building more adaptable, explainable, and resilient autonomous systems and highlights the necessity of unifying data, model, task, and goal considerations under a principled, feedback-driven design approach.