license: apache-2.0
task_categories:
- image-feature-extraction
tags:
- autonomous-driving
- formal-verification
- carla
- automatic-emergency-braking
size_categories:
- n<1K
Camera frames and trained networks for the braking verification study
The inputs to AD-Assurance-Lab/formal-verification--aeb--code, measured on 2026-09-10.
Why this exists. The study's results are in git. These files are not: they are large and git ignores them. Without them the published numbers can be reproduced in kind but never exactly, because the simulator does not render bit-identical frames from one run to the next. Different frames give different networks, which give different certificates.
What is here
captures/Town01/ |
66 frame sets, one per scenario and sun altitude, plus the campaign manifests |
models/Town01/ |
6 trained networks, three policies for each of two scenarios |
MANIFEST.json |
a hash of every file, and the capture harness recorded inside every frame set |
How the pieces fit
A frame set is a rendered approach at one sun altitude: images, the range to the conflict point at each pose, and a stamp saying what made it. The stamp records the map, the cloud cover, the settle time, the determinism rules digest, the texture streaming and quality settings, and the order the knots were captured in. That last field exists because capture order silently ruined a dark endpoint once and nothing recorded it.
The networks are what the certificates are computed on. Each certificate in the code
repository names its network by hash, and those hashes are in MANIFEST.json. If they do
not match, the certificate and the network are not a pair and nothing may be concluded from
them together.
Using them
Clone the code repository, put these directories at results/captures/ and
results/models/, then:
python -m study.status # where the study stands
python tools/family_fidelity.py # how much light each piece of the range spans
Both run without a simulator.
Training reproduces exactly from these frames. Four separate runs gave byte-identical checkpoints. So the networks here can be regenerated from the frames here, and the certificates from the networks. The chain is only broken if the frames are lost.
What the study found
Three policies differing only in which light levels their training frames came from. All three pass all three lighting conditions the standard tests, ten times out of ten, so the standard's own procedure cannot tell them apart.
Driven at 65 light levels between those conditions, three repetitions each: the policy trained on the whole range crashes 0 times, and the policy trained on the regulatory test points crashes 21 times on the lead-vehicle scenario and 9 on the crossing pedestrian. A certificate names those light levels without simulating them.
Method and results: AD-Assurance-Lab/formal-verification--aeb--code, docs/STUDY_REPORT.md.