Percentile requests as generation targets: method realizes 1,422 of 1,440 requests (98.75%) on highD

It proposes a history-conditioned risk-percentile request interface, giving the same numerical request a common relative meaning across traffic contexts. Existing generators can enforce behavior, adversity or feasibility conditions, but offer limited control over how extreme a result is relative to plausible futures in the same context; this work defines the request as a position in the natural conditional future-risk distribution. The text reports that at a PET of one second the empirical CDF rises from 0.319 for histories with 3-5 vehicles to 0.609 for histories with at least nine, showing a fixed physical threshold signals different adversity in different contexts.

A learned reference risk distribution maps the percentile request to a physical risk target, and a percentile-conditioned joint diffusion model with sampling-time risk guidance realizes it. The reference is trained with CRPS using temporal-relational encoders and PET-range queries; the generator receives the request, physical target and CDF descriptor, and occupancy-geometry updates push minimum PET toward the target during the final 15 DDIM steps. On the primary evaluation set, 1,422 of 1,440 requests (98.75%) fall within the 0.05 tolerance, with P-MAE 0.00673 and PET-target MAE 0.00991 seconds; continuous-target requests are met in 98.65% of cases and all 111 point-mass-target requests are matched exactly.

The ablation identifies sampling-time risk guidance as the main source of fine precision, with learned conditioning adding further gain. Guidance alone reaches 95.07%, conditioning alone 24.79%, and removing both leaves 16.18%; adding conditioning to guidance raises Fine by 3.68 percentage points and more than halves P-MAE from 0.01469 to 0.00673. Controlled experiments use the same trained weights, reference, histories and noise draws, and show that pooled physical outcomes cannot separate configurations while realized percentiles can.

The interface transfers to time-to-collision (TTC) requests in car following, and higher requests pose harder scenarios to an IDM and MOBIL planner. TTC guidance meets 98.61% of requests without retraining the main generator (17.22% without guidance); at the highest request the planner's hard braking reaches 12.85% and rear TET rises to 0.578 seconds. The TTC reference more than halves the CRPS of the unconditional distribution on car-following clips, from 1.21 to 0.52 seconds; in the planner test the median Spearman correlation between request and encountered percentile is 0.74.

The interface targets multi-vehicle highway scenes, and the experiments use highD natural trajectories throughout, with the reference and generator fitted on 9,913 complete training clips and the primary evaluation set drawn from 11,649 clips of 16 test recordings. It suits scenario libraries organized by context-relative adversity: the physical target is known before generation, so a test that also requires a physical severity, such as a PET below one second, can first select histories whose targets reach it. Researchers extending it to another surrogate or scenario would need to estimate that surrogate's history-conditioned natural distribution and implement a mechanism for realizing its inverse-CDF targets; the text instantiates PET and TTC and mentions merging at highway entries and exits and crossing conflicts at urban intersections as further settings.

A percentile is as accurate as the reference that defines it, which is why the text reports reference prediction quality separately from request realization; references are trained under leave-recording-out labels, and a leave-recording-out reference can estimate a given history differently from the full reference. Each model was trained once, so reported intervals describe variation across histories and recordings for one trained model. In the planner test the surrounding vehicles replay generated futures without reacting, holding the scenario fixed while the planner responds. Comparisons among reference encoders show an absolute-history encoder is slightly more accurate in CRPS, indicating room in the reference implementation. A reader seeing only a fast parse without figures or tables should return to the original for the tail calibration diagnostics and the same-history visualizations of external priors.