Still an Open Problem -- Cracking the Code of AV Safety and Validation
Automated vehicle validation is still the big unsolved puzzle. We don't yet have a single method, metric, or standard that can prove an automated car is safe enough for the open road. Just piling on test miles won't cut it. Rare events are too rare, the real world is too messy, and software – and the world keeps changing after vehicles are deployed.
The keynote will unpack why this problem remains open — from statistical limits and simulation credibility to shifting operational design domains, machine learning brittleness, and the headaches of over-the-air updates and cyber threats.
Instead of pretending there's a silver bullet, the talk lays out a layered approach: combine scenario-based testing, credible simulation, safety cases, and continuous monitoring once cars are in the wild.
Most importantly, it will challenge the community with the open tasks — practical things we can actually work on together: building ODD schemas, creating rare-event testing benchmarks, stress-testing perception, agreeing on sim-to-real playbooks, and designing transparent reporting after deployment.
