DrivingBench researchers conducted a remarkable experiment on a real Toyota Corolla with general-purpose AI models. In this study, different AI models attempted to control the vehicle on a course marked with small cones.

The testing process provided valuable data for understanding the relationship between AI models and autonomous driving technologies. Researchers prepared a special course to observe how the models processed camera and telemetry data.

AI Models Behind the Wheel of a Toyota Corolla

In tests conducted on a 2022 Toyota Corolla, vehicle control was provided via the CAN bus through a Comma Four device. Images and telemetry information from the vehicle cameras were used to generate steering, throttle, and brake commands for the AI ​​models.

The tested models included the GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol. Each model was given three attempts to analyze its mistakes and complete the course.

The course consisted of curves marked with colored cones and a parking area marked with blue cones. According to the test results, the only model that successfully completed the course from start to finish was the GPT-6 Astra. The GPT-6 Astra completed approximately 49% of the course in its first attempt and finished it in 5 minutes and 22 seconds in its second attempt. The other models did not achieve similar success and struggled to complete the course.

Claude Fable 5.1 was able to complete half of the course on its third attempt. Some models failed the test, even failing to navigate the first corner.

Difficulties Encountered During Driving

The biggest challenge faced by the AI ​​​​models was accurately interpreting the course and the vehicle’s position. Some models veered in the wrong direction, unable to understand which side of the cones they should follow.

Models that couldn’t accurately assess the vehicle’s physical dimensions from the camera approached obstacles too closely. Decision-making speed was another significant problem.

Some models left the vehicle motionless for extended periods while calculating the next move.

The GPT-6 Astra

processed images approximately every 5-6 seconds, generating 6 commands per minute. Thanks to this, the vehicle continued to progress on the course without interrupting its driving. Safety precautions were also kept at the highest level during the tests.

The vehicle speed was limited to between 1.8 km/h and 12.6 km/h. Throughout the entire test process, a safety driver was on standby with their foot on the brake pedal.

The researchers stated that this study was not conducted to prove that general-purpose AI models are ready for autonomous driving on open roads. The main aim of the experiment was to observe what these models could do on a real vehicle with camera and telemetry information.

Do you think AI models could reliably manage driverless vehicles in the future?

DrivingBench researchers conducted a driving test on a Toyota Corolla with AI models. Which model completed the course?