Researcher: VLAs don't generalize, and e2e self-driving planning needs 100M+ km to benchmark
abursuc · x · 2026-09-18
At the SSAD 2026 summer school, a researcher laid out sharp criticisms of end-to-end autonomous driving:
- VLAs do not generalize: they fail on new cameras, and their chain-of-thought explanations do not reflect the internal cause of actions — they're post-hoc rationalizations
- Accountability gap: it's unclear what information was given to Dutch authorities for FSD clearance; meanwhile planning can now be benchmarked, but statistical significance requires 100M+ km of driving
The upshot: there's a major gap between public accountability and the statistical evidence end-to-end systems can actually provide.
More from Embodied
- KUKA CEO Talks Physical AI at Augsburg HQ After 600,000 Robots Deployed — lukas_m_ziegler · 2026-09-18
- Mostaque predicts open-source robotic intelligence will handle 95% of daily tasks by end of next year — AjdDavison · 2026-09-18
- Notmisha: imitation learning works when trained on the job — pretraining is what falls short — notmisha · 2026-09-18
- Microduck robot duck head design iterations revealed by its creators — Sentdex · 2026-09-18
- Tesla FSD data shows 40% fewer collisions than manual driving in Australia and NZ — elonmusk · 2026-09-18
- KRAFTON AI launches Autonomy hackathon with 10M won top prize and interview fast-track — Kangwook_Lee · 2026-09-18