DrivingBench researchers conducted a remarkable experiment with general-purpose artificial intelligence models on a real Toyota Corolla. In this study, different artificial intelligence models tried to control the vehicle on a track determined by small cones.
The testing process provided important data to understand the relationship between artificial intelligence models and autonomous driving technologies. The researchers created a special track to observe how the models process camera and telemetry data.
Artificial Intelligence Models at the Wheel of Toyota Corolla
In the tests carried out on a 2022 model Toyota Corolla, vehicle control was provided over the CAN bus via the comma four device. Images and telemetry data from vehicle cameras were used for artificial intelligence models to generate steering, gas and brake commands.
Models tested included GPT-6 Astra, Claude Fable 5.1, Grok 4.6 and GPT-5.6 Sol. Each model was given a total of three chances to analyze their mistakes and complete the course.
The course consisted of bends delimited by colored cones and a parking area marked with blue cones. According to the test results, GPT-6 Astra was the only model that could successfully complete the course from start to finish.
While GPT-6 Astra completed approximately 49 percent of the course in its first attempt, it completed the course in 5 minutes and 22 seconds in its second attempt. Other models could not show similar success and had difficulty completing the course.
Claude Fable 5.1 was able to advance halfway through the course on his third attempt. Some models failed the test by not even being able to pass the first corner.
Difficulties Encountered While Driving
The biggest difficulty experienced by artificial intelligence models was to correctly interpret the track and the location of the vehicle. Some models failed to understand which side of the cones they were supposed to go on and went the wrong way.
Models that could not accurately evaluate the physical dimensions of the vehicle from the camera got too close to obstacles. Decision-making speed also stood out as another important problem.
Some models left the vehicle stationary for long periods of time while calculating the next move. GPT-6 Astra evaluated images approximately every 5-6 seconds and produced 6 commands per minute.
In this way, the vehicle continued to move on the track without interrupting driving. Security measures were kept at a high level during the tests.
Vehicle speed was limited between 1.8 km/h and 12.6 km/h. A safety driver stood with his foot on the brake pedal throughout the entire testing process.
The researchers stated that this study was not conducted to prove that general-purpose artificial intelligence models are ready for autonomous driving on open roads. The main purpose of the experiment was to observe what these models could do on a real vehicle with camera and telemetry data.
Do you think artificial intelligence models can safely manage driverless vehicles in the future?
