Series
Adas E2e 2026
The "Adas E2e 2026" series has 6 parts — read them in order from part 1.
ADAS & Autonomous Driving
From UniAD to VLA: Mapping End-to-End Driving in 2026
A four-axis vocabulary — input representation, planning output, supervision signal, evaluation protocol — for reading any end-to-end driving system.

Run a Driving VLA Yourself: OpenDriveVLA on nuScenes
Hands-on tutorial covering the architecture, installation, and inference of OpenDriveVLA — the AAAI 2026 driving VLA from TU Munich — on the nuScenes dataset.

Benchmarks Are Arguments: NAVSIM v2, Bench2Drive, WOD-E2E
Three types of benchmark scores that cannot substitute each other. Read this before trusting any autonomous driving number in 2026.

Alpamayo-R1: Reasoning You Can Actually Inspect
NVIDIA Alpamayo-R1 proves a self-driving car can explain why it braked — with structured causal chains you can audit, not decorative text.

World Models Stopped Being About Pretty Video
The 2026 shift: world models are judged not by FVD but by whether generated scenes measure policy quality. GAIA-4, Orbis 2, WorldLens.
What Robotaxi Fleets Actually Publish About Safety
Series finale: Waymo's 220M miles, an IIHS report showing 68% fewer crashes, Apollo Go's 22M trips — and why methodology beats the headline number.