AI

How corner is a corner case? Percentile control for highway scenario generation

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 19:18:24 UTC.

Briefings

A new method uses history-conditioned risk percentiles to generate autonomous vehicle scenario simulations, allowing precise control over scenario extremity. The study applies a percentile-conditioned joint diffusion model to generate multi-agent futures, achieving 98.75% success rate in meeting a 0.05 percentile tolerance on highD metrics. Key metrics include a mean percentile error of 0.00673 and PET-target error of 0.00991 seconds, with the interface linking risk specification to physical realization.

Key points

• The study introduces a method where the adversity of a generated scenario is defined by its percentile in the conditional distribution of future risk given observed history.
• Experiments on the primary evaluation set achieved 1,422 of 1,440 requests within a 0.05 percentile tolerance, with a mean percentile error of 0.00673 and PET-target error of 0.00991 seconds.

Context

Autonomous driving scenario generators currently offer limited control over scenario extremity relative to plausible futures. This research addresses that gap by defining scenario adversity as its percentile in the conditional distribution of future risk, enabling calibrated answers about how 'corner' a scenario is and how its cornerness can be fine-tuned.

Limitations

The primary evaluation set used minimum post-encroachment time (PET) as the risk surrogate on highD. Results are limited to this specific evaluation set and do not represent generalizable performance across all traffic contexts or vehicle types.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 11/11/2026, 21:33:24 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeidxphgatocbjvtw2ssvcvhvg7omt4z6dt2afxwhqm4lyuf7egmdr4