DrivingBench Sandbox rename got AI to drive a real Corolla

DrivingBench was not set up to build a practical self-driving system. Its three founders — Aditya Ramabadran, Tobias Gessler and Simon Marns, who work at the AI startup Axiom Math — built it after seeing recent demos of AI drawing pictures or moving objects with robots, and wondering whether general-purpose AI could already drive. The test car was a 2022 Toyota Corolla fitted with the in-car device 'comma four' and the open-source driver-assistance software 'openpilot', with a human driver seated and ready to brake.

The instructions given to the AI ran under 600 words: stay inside a reverse U-shaped parking-lot course marked by cones, then park in a space marked by a small blue cone. Increasing distance was the main goal, with a fast finish requested where possible. Two problems surfaced ahead of driving skill itself. Models that recognized they were operating a real car refused on safety grounds, with GPT-6 Astra refusing especially often, forcing hours of adjusting prompts and system names; and because the car keeps moving on the previous command while the AI considers its next move, slower responses raise the risk of leaving the course. DrivingBench says curve failures came down mainly to 'perception' — reading the correct driving position from the arrangement of cones.

Marns described the result as an interesting experiment showing a new capability emerging in AI, and argued such tests have value because AI long used mainly inside computer screens is likely to reach the physical world more often through robots. Whether that reading holds may depend on whether the team's planned DrivingBench v2, with more trials per model, different inference settings and harder courses, reproduces the result.