SSAD 2026 researchers question generalization and accountability of end-to-end autonomous driving

At the SSAD 2026 summer school/workshop, researcher abursuc presented and relayed a set of systematically pessimistic observations about technical approaches in autonomous driving and embodied AI, centering on two major shortcomings of end-to-end approaches: generalization and accountability.

Confirmed

• Public datasets are insufficient for training; scaling is limited when data lacks diversity, and camera-only data is not enough—accidents are hard to avoid
• Open-loop evaluation does not generalize to real-world scenarios
• Diffusion models bring at most a few percentage points of improvement on point-cloud synthetic data; novel view synthesis also fails to generalize on public datasets
• Vision-language-action (VLA) models do not generalize to new cameras; their chain-of-thought explanations do not reflect the true internal reasons behind actions and were called post-hoc rationalization
• End-to-end autonomous driving suffers from an accountability gap: when Dutch authorities approved Tesla FSD, it was unclear what information had been submitted
• Planning verification for end-to-end methods is believed to require hundreds of millions of kilometers of driving mileage
• End-to-end approaches still require human supervision

Why it matters

• These judgments collectively challenge the current narrative centered on end-to-end methods and data scaling in autonomous driving, exposing structural gaps in evaluation methods (open-loop), data sources (public datasets), and safety accountability mechanisms
• The observation that VLA chain-of-thought is disconnected from true internal motivations sounds an alarm for embodied AI practices that rely on models' self-reported explanations
• The speaker also acknowledged bright spots: interpretability methods are emerging, scaling laws are encouraging, and AD is accelerating across Europe and worldwide—so these criticisms serve as a correction to overheated expectations rather than a wholesale rejection

2026-09-18 ~ 2026-09-18 · 6 related posts

• SSAD2026 Talk Argues Camera-Only Autonomous Driving Faces Scaling Limits Without Diverse Data — abursuc · 2026-09-18
• [source] SSAD 2026 Talk: Open Datasets Fall Short, End-to-End AV Needs a Babysitter — abursuc · 2026-09-18
• SSAD 2026: Diffusion models and public datasets both fail to generalize — abursuc · 2026-09-18
• [source] SSAD 2026: VLAs fail on new cameras, CoT doesn't reflect real causes — abursuc · 2026-09-18
• [source] Researcher: VLAs don't generalize, and e2e self-driving planning needs 100M+ km to benchmark — abursuc · 2026-09-18
• Is computer vision solved? Researcher flags e2e accountability gap in FSD oversight — abursuc · 2026-09-18