Blog
Releases, reports and research from DrivingBench.
v2 Release
A Tesla on a harder course, and a second kind of driver: VLAs and WAMs fine-tuned on 94 minutes of our driving data, alongside general-purpose VLMs.
View the results →
v2 Report
How the Tesla connects, the fine-tuning dataset, how we trained and ran π0.5 and FLUX 3 Action, and what the results show.
DrivingBench: Can Vision-Language Models Drive a Toyota Corolla? ↗
The v1 paper on arXiv: setup, protocol and results.
v1 Report
How the system works, the course, the harness and tools, what we saw in the traces, and what comes next.
v1 Release
Four frontier models drive a real Toyota Corolla through a cone course, with up to three attempts each in one conversation.