Adding more demonstrations is a common way to scale end-to-end autonomous-driving training, but a larger dataset does not guarantee a better driving policy. The usefulness of an example can depend on the planner learning from it: a demonstration that improves one model may be redundant or even counterproductive for another. Policy-Aware Data Valuation (PAD) addresses this mismatch by estimating how useful each candidate demonstration is to a specified target policy.
Key points
- Value is policy-dependent. PAD ranks examples for a particular driving policy rather than treating data quality as a fixed property. The authors’ cross-planner analysis finds markedly different rankings for the same candidate pool across planner families.
- The estimate uses learning and scene signals. PAD combines scene-level calibration feedback with projected policy gradients and a regularized approximation of local training geometry. Together, these components produce influence-based values intended to estimate how examples affect the target policy, rather than relying only on scene attributes or indirect training proxies.
- Removal tests examine the direction of the estimates. A controlled group-removal study is used to test whether the valuations capture the direction of data influence. The abstract says the study supports their directional fidelity, but does not provide its detailed setup, error measures, or data scale.
- Reported results span multiple planners. On DiffusionDrive, PAD improves EPDMS over real-only training by 3.5 points on NavTest and 5.0 points on NavHard. The paper also reports that policy-specific curation beats both real-only and full-pool training across diffusion-, regression-, and flow-matching-based planners, while substantially reducing scene-level regressions.
Why it matters—and what remains unclear
PAD reframes data curation as a relationship between a demonstration and the policy that learns from it. Instead of indiscriminately adding examples, a training pipeline could estimate which candidates are most likely to improve its particular planner. If these estimates can be computed at practical cost, the approach could make dataset growth more targeted and potentially avoid examples that do not help a given model.
The available summary, however, leaves important questions open. It does not specify valuation cost, candidate-pool composition, the fraction of data retained, or the full experimental tables. The reported gains therefore establish a promising direction, not a complete account of performance under every data scale or deployment condition. Nor does a reduction in scene-level regressions by itself establish that all safety risks have been addressed. More detailed experiments and independent validation will be needed to assess the method’s broader usefulness.
Source: Hugging Face Daily Papers
Comments
Checking sign-in status...
Loading comments...