These Tech Workers Made ChatGPT Drive a Toyota Corolla
The team used frontier LLMs with no prior training data navigate a simple parking lot course.
Three tech workers in San Francisco, known as DrivingBench, have demonstrated that large language models (LLMs) like ChatGPT can be hooked up to a Toyota Corolla and taught to navigate a parking lot course. The team, consisting of Aditya Ramabadran, Tobias Gessler, and Simon Mahns, posted their experiment online, including code, prompts, and videos.
They used the Comma system to connect chatbots - GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol - to the car's steering, accelerator, and brakes. The goal was to find out if untrained, off-the-shelf AI can drive a real car. Grok, Sol, and Fable only managed a few meters and failed to complete the course, while GPT-6 Astra eventually learned to navigate it.
The researchers stress that this is not a practical application for real-world driving, but rather a proof-of-concept to demonstrate the potential of LLMs. They warn that these models are unlikely to be specifically trained for this purpose, nor do they have real-world driving environments. The Comma system, an off-the-shelf solution, was used to connect the car to the laptops running the LLMs.
The three members of DrivingBench met through their day job at Axiom Math, an AI math startup. The idea for the project came after they saw demos of Astra and other models performing robotic tasks that LLMs couldn't do before.
Written by urgent.news from 404 Media's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
