Researchers led by Chengkai Xu have built Policy-Aware Data Valuation (PAD), which scores each training demonstration by how much it would help a particular self-driving planner, according to a paper on arXiv. The authors say more data does not always mean better driving, because a demonstration that helps one planner can add nothing to another or even harm it.
PAD estimates each sample's influence from scene-level feedback and projected policy gradients. Data chosen this way did better than real data alone or the full pool for planners based on diffusion, regression and flow matching. On DiffusionDrive it raised the EPDMS score by 3.5 points on the NavTest benchmark and by 5.0 on NavHard.



