Autonomous driving · Policy-aware learning

Policy-Aware Data Valuation
for Autonomous Driving

Better driving begins with the right data.

Chengkai Xu1Yicheng Guo1Jiaqi Liu2Yiming Cui1Shiyu Fang1Peng Hang1Hao Zhang3Jian Sun1
1 College of Transportation, Tongji University2 Department of Computer Science, UNC Chapel Hill3 Department of Control Science and Engineering, Tongji University
A value for every demonstration.
A dataset for your policy.
Explore PADLearn what matters.
Drive better.

A different way to scale

More data?
The right data.

A demonstration’s value depends on the policy learning from it.

PAD connects real driving outcomes to the learning signal of each demonstration. The result: a synthetic dataset selected for your planner, with every retained example ranked by policy-specific value.

Discover the mechanism
88.3NavTest EPDMS
with DiffusionDrive
25%of the synthetic pool
retained for training
3policy families
one valuation framework

The mechanism

Driving outcomes.
Turned into data value.

Measure what your policy learns from.
Keep what helps it drive better.

Let the policy’s outcomes lead.

Evaluate the real-only base policy on 4,000 held-out real scenes. Its driving outcomes define a policy-specific direction for valuing demonstrations.

4,000 real scenes · Policy-specific feedback

The results speak

Less synthetic data.
More capable driving.

Across diffusion, regression and flow.
Data selection that follows the learner.

PAD · NAVTEST EPDMS
88.3
+3.5 points over real-only

A quarter of the synthetic data.
A stronger DiffusionDrive.

DiffusionDrive on NavTest

EPDMS / 100
Real only84.8
Full synthetic pool85.9
PAD 25% selected88.3
050100

Same real training set. A shared pool of 208,413 synthetic scenes.

25% retainedof the synthetic candidate pool
Best PAD result on NavHard32.9+5.0 vs. real-only

Paper-reported results · NavTest: one-stage · NavHard: two-stage · EPDMS on a 0–100 scale

From selection to behavior

The difference
is in the driving.

Policy-aware data choices translate
into stronger driving trajectories.

−0.900Spearman correlation

A ranking with real consequences.

The five-group removal experiment connects PAD’s values to actual retraining effects. Removing lower-valued groups improves driving utility; removing the highest-valued group causes the largest reduction.

Policy-aware data. Purposeful learning.

Learn what matters.
Drive better.

Let’s rethink what makes driving data valuable.

Contact Me

Scroll to explore · Escape to close