Build the CONEs You Need

Scalable work zone generation for closed-loop driving simulation

Teaser figure
assets/teaser.png
Overview. Real work zones from ROADWork are mined into archetypes, re-generated as MUTCD-compliant layouts, deployed in CARLA, and used to evaluate and finetune closed-loop planners.

Abstract

Construction work zones are among the most common and most hazardous departures from the road as mapped, yet they are nearly absent from the benchmarks used to evaluate autonomous driving planners. We present CONE, a framework for generating construction work-zone scenarios at scale for closed-loop simulation.

Starting from real-world ROADWork data, CONE clusters observed work zones into a small set of archetypes and produces new layouts with two complementary generators: a rule-based generator that instantiates MUTCD typical applications with archetype-conditioned priors, and a learned VAE-LDM generator trained on real scene layouts. Generated layouts are deployed in CARLA and evaluated in closed loop with state-of-the-art planners.

We show that a planner strong on standard benchmarks fails on the majority of generated work zones, that finetuning on CONE data recovers in-distribution performance, and, critically, that the gains transfer to held-out geometric and prop variants never seen in training. Transfer across archetypes remains limited, motivating diverse generated layouts rather than narrow finetuning.

From real work zones to closed-loop tests

CONE is a four-stage pipeline. Each stage can be run on its own, and evaluation results feed back into which layouts get generated next.

1

Mine real data

ROADWork scenes are parsed into object layouts and clustered into archetypes C0–C4 that cover the common work-zone configurations.

2

Generate layouts

A rule-based generator instantiates MUTCD typical applications; a VAE-LDM learned from real layouts adds diversity beyond the templates.

3

Deploy in CARLA

Layouts are placed onto CARLA routes with the right props, lane closures and traffic, producing drivable closed-loop scenarios.

4

Evaluate and finetune

Planners drive the scenarios end to end. Failures are diagnosed, and the same scenarios become finetuning data.

MUTCD template derivation from real ROADWork scenes
assets/template_derivation.png
Rule-based templates are derived directly from real scenes, so the generator stays faithful to how work zones are actually laid out.

Five archetypes cover the road

Clustering real work zones yields five archetypes. Each has its own MUTCD templates and prior over props, geometry and taper length.

Finetuning on generated work zones transfers

Closed-loop success rate of a TransFuser-v6 planner before and after finetuning on CONE-generated scenarios, ordered by distance from the training distribution.

In-distribution

17.971.4%

Same archetype and template family as the finetuning data.

Held-out geometric & prop variants

12.587.0%

Taper lengths, offsets and prop types never seen in training. The gain is not memorisation of specific layouts.

Cross-archetype

Limited

Narrow finetuning on one archetype does not carry to others — which is exactly why diverse generated layouts are needed.

Full evaluation table across all six splits
assets/results.png
Per-split results for pretrained TFv6, finetuned TFv6 and SimLingo. See the paper for the full protocol and route counts.

What the planner sees

Same route, same generated work zone. The pretrained planner drives into the closed lane; the finetuned planner reads the taper and merges early.

Pretrained TFv6
assets/pretrained_fail.mp4
Fails Pretrained TFv6
Finetuned TFv6
assets/finetuned_pass.mp4
Passes TFv6 finetuned on CONE

Citation

Authors and venue will be filled in after the review period. Until then, please cite the anonymous submission.

@inproceedings{cone2027,
  title     = {Build the CONEs You Need: Scalable Work Zone Generation
               for Closed-Loop Driving Simulation},
  author    = {Anonymous},
  booktitle = {Under review},
  year      = {2027}
}
Code, generated scenario sets and finetuned checkpoints will be released on acceptance. This page intentionally contains no repository link, affiliation or contact details to comply with double-blind review.