Data-Centric Autonomous Driving: A Data-Utilization Framework for Learning from Large-Scale Driving Logs

Closed-loop autonomous driving increasingly combines model-centric architectural development with data-centric learning. These two directions are complementary: model-centric methods improve representation, prediction, planning, and control capabilities, whereas data-centric methods improve how training samples, supervision signals, and closed-loop failures are selected and reused. A key remaining challenge is therefore how to use heterogeneous driving logs more effectively to improve closed-loop policy learning. Existing methods have made certain progress in aspects such as representation learning, generative planning, and closed-loop evaluation. This paper proposes a data-centric closed-loop autonomous driving learning framework. Combining risk-aware data valuation, supervision-calibrated distillation, and counterfactual attribution-guided corrective replay, the framework adjusts the training distribution, calibrates sample-level supervision, and converts training-side failure trajectories into intervention-validated corrective signals. Closed-loop experiments on Bench2Drive demonstrate improved driving performance, while NAVSIM provides supplementary non-reactive cross-platform evaluation under the tested benchmark conditions. The experimental results show that data-centric optimization strategies can help reduce performance degradation under the evaluated perturbations and provide a feasible approach to high-risk sample management, supervision-quality calibration, and failure-informed corrective replay.