The paper introduces Policy-Aware Data Valuation (PAD), a framework that assigns utility scores to driving demonstrations specific to a target policy rather than relying on intrinsic sample properties or training-dependent proxies. PAD computes scalable influence-based values by combining scene-level calibration feedback with projected policy gradients and a regularized approximation of local training geometry. A controlled group-removal study validates the directional fidelity of these valuations, while cross-planner analysis demonstrates that identical candidate pools produce markedly different rankings across diffusion-, regression-, and flow-matching-based planner families. Evaluated on DiffusionDrive, policy-specific curation via PAD improves EPDMS by 3.5 points on NavTest and 5.0 points on NavHard over real-only training. The method consistently outperforms both real-only and full-pool training baselines across planner architectures while substantially reducing scene-level regressions, establishing policy-aware valuation as a mechanism for targeted policy improvement over indiscriminate data accumulation in autonomous driving systems.

Read original