What is it about?
Self-driving cars must keep steering safely when their camera view is disturbed by glare, fog, blur or sensor noise. Testing this on a real vehicle is slow and expensive, so engineers can only run a limited number of tests and need to choose which ones to run first. This work asks whether cheap offline tests, which apply such disturbances to recorded camera images, can tell us which disturbances are most likely to make the real car fail. Using data from 40 camera disturbances tested on a research vehicle in Munich, ranking them by their offline effect found failures in all of the first 10 vehicle runs, compared to about 4 when picking at random.
Featured Image
Photo by Charlie Deets on Unsplash
Why is it important?
Real-world tests of self-driving cars are slow, costly and need a safety driver, so test campaigns often stop before every scenario has been tried. Choosing a smart order means dangerous failures are found earlier, with fewer test drives. That allows us to fine-tune the driving model at a earlier stage. This study shows that cheap offline tests on recorded images can guide that order, even though they cannot predict every failure exactly.
Perspectives
This work grew out of my bachelor’s thesis, where we tested a self-driving model on a real research car on public roads in Munich. Seeing how much effort each single test drive takes made me ask how we could use those drives more wisely. I’m now extending the idea from camera images to LiDAR data from the same vehicle.
Matthias Weil
Read the Original
This page is a summary of: Which Camera Perturbation Should We Test First? Offline-Guided Prioritization of On-Vehicle ADS Tests, October 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3832783.3844568.
You can read the full text:
Contributors
The following have contributed to this page
