Researchers test large language models driving a Toyota Corolla

Researchers have conducted a new experiment, titled DrivingBench, to test the ability of large language models (LLMs) to control a physical vehicle. Using a rented Toyota Corolla, the team integrated advanced AI models—including OpenAI’s GPT-6 Astra, Claude, and Grok—to translate natural language instructions into driving commands.

In a test conducted in a Bay Area parking lot, GPT-6 Astra successfully navigated a 134.7-meter course of mini-cones and parked in a designated zone, completing the run in 5 minutes and 22 seconds. This marks what researchers describe as the first time a general-purpose language model has driven a real car through a course from start to finish.

The experiment aimed to observe how software designed for text and code handles physical realities like momentum and geometry. While the models demonstrated varying levels of success in navigating traffic and making decisions, the researchers noted that these general-purpose models are not intended to replace purpose-built autonomous systems like Waymo or Tesla.