GPT-6 Astra Navigated Cone Course via Text Commands
OpenAI's latest model demonstrated remote vehicle control in a closed-course test, highlighting a new approach to robotics.
Updated on Sept. 23, 2026 in Robotics
Live Poll
Do you trust artificial intelligence models to operate physical machinery and vehicles safely?
OpenAI's GPT-6 Astra has successfully navigated a 130-meter traffic cone course by sending individual steering and throttle commands through a chat session. The research-stage experiment evaluated the model's ability to maintain a vehicle within four meters of a centerline.
Why it matters
This test demonstrates a shift toward using large language models as high-level controllers for physical hardware, exploring how text-based reasoning can translate into real-world kinetic action. It highlights an emerging research focus on whether models can move beyond text generation to execute complex, multi-step physical tasks.
GPT-6 Astra completed the 130-meter course in 5 minutes and 22 seconds while adhering to a speed cap of 3.5 meters per second. The vehicle utilized comma.ai hardware to execute commands for steering, acceleration, and braking while maintaining a maximum deviation of four meters from the course centerline.
The players
OpenAI
An AI research organization focused on the development of large language models and autonomous agents.
Toyota
A global automotive manufacturer whose Corolla vehicle served as the platform for the driving test.
comma.ai
A startup that produces hardware and software interfaces for vehicle control systems.
The details
The test required each AI model to pilot a Toyota Corolla by outputting individual driving commands through a live chat interface, which were then processed by comma.ai hardware—an aftermarket device that provides an interface for vehicle steering and braking control. While some models refused the task citing safety concerns, Astra was the only participant to achieve a 100 percent completion rate on its second attempt, navigating the course without human intervention.
Timeline
September 23, 2026: The performance results were published.
The Tech Race
The experiment follows the trajectory set by the DARPA Grand Challenge autonomous vehicle competitions, which long defined the benchmark for machine navigation. It signals a move away from hard-coded sensor-fusion stacks toward the use of multimodal models as general-purpose logic controllers.
This research remains confined to a controlled parking lot environment and is not intended for public road use. It offers a glimpse into potential future workflows where language models serve as intermediaries between user intent and automated machinery.
The takeaway
The experiment proves that large language models can perform basic, low-speed physical navigation, though consistency varies wildly between models. Watch for future benchmarks evaluating whether these models can scale to higher speeds or more complex urban obstacles.
Further reading
For more on the latest developments in autonomous control, explore our Robotics archives.
Live Poll
Do you trust artificial intelligence models to operate physical machinery and vehicles safely?
