GPT-6 Astra AI successfully pilots a real Toyota Corolla in parking lot test

Researchers from the DrivingBench project have demonstrated that general-purpose large language models (LLMs) can navigate a physical vehicle. In an experiment conducted in a parking lot, tech workers connected various AI models, including Claude, Grok, and OpenAI’s GPT-6 Astra, to the steering, accelerator, and brakes of a Toyota Corolla.

While other models like Grok and Claude failed to complete the course, GPT-6 Astra successfully navigated a 134.7-meter cone course in five minutes and 22 seconds. The experiment highlighted the emergent capabilities of off-the-shelf AI, though the driving speed averaged only 0.94 miles per hour.

The researchers noted that the models initially refused to drive due to safety concerns upon realizing they were controlling a real vehicle. To bypass this, the team had to convince the AI that it was operating within a simulation by renaming their server to ‘DrivingBench Sandbox.’ Researchers described the results as a demonstration of how AI utility is beginning to shift from digital tasks into the physical world.