Engineers Use GPT-6 to Pilot a Toyota Corolla Through an In-N-Out Drive-Thru

Aditya Ramabadran, Simon Mahns and Tobias Gessler of the startup Axiom equipped a 2024 Toyota Corolla with windscreen cameras and a power-steering link, then gave OpenAI’s GPT-6 Astra access via a chat interface on a laptop. While a safety driver monitored the brakes, the model interpreted the visual feed and issued steering commands, slowly moving the car to the In-N-Out window. The trio noted that the model, originally built for text, code and image generation, managed the task without any pre-training for driving, suggesting emergent physical understanding.

They compared the result to other large models, finding Astra the only one to complete their own DrivingBench course, albeit slowly. The experiment underscores a growing focus on multimodal reasoning and hints at future applications in robotics and home automation. Researchers such as Andrew Dai of Elorian AI and Xingang Guo of Scale AI are also developing benchmarks like Humanity’s Sixth Sense to measure such physical scene comprehension.