Self-Driving Cars Could Someday Take Requests
Researchers at Delft University of Technology have created a system that allows passengers to adjust an autonomous vehicle's driving style using conversational language. By employing OpenAI's GPT-4o-mini model, the system converts natural-language requests into parameter tweaks for a vehicle's motion planner. The approach keeps humans in the loop by asking for confirmation before applying changes to driving behavior.
Key Takeaways
- Traditional motion planners in self-driving cars are pre-tuned by engineers, leaving passengers with virtually no ability to customize how a vehicle drives during a trip.
To solve this limitation, researchers led by Diego Martinez-Baselga at Delft University of Technology integrated OpenAI's GPT-4o-mini model with a motion-planning controller.
- Instead of placing the language model in direct control of steering or acceleration, the system interprets statements such as "I am running late, go fast" to alter the weighting parameters of a safety-bounded motion planner.
Because large language models lack deterministic safety guarantees and suffer from slower response times, using them for real-time vehicular control presents significant risks.
- By restricting the artificial intelligence model to translating fuzzy human preferences into scenario-based parameter updates, the vehicle remains safely inside established control bounds.
Before any adjustments take effect, the vehicle describes the proposed behavior changes in plain language and asks the passenger to confirm, ensuring clear communication between the user and the automated control system.
- Delft University of Technology researchers developed a system that lets passengers use natural language to customize autonomous driving styles.
The system utilizes OpenAI's GPT-4o-mini model to translate spoken requests into mathematical parameter adjustments for safety-aware motion planners.
- Passengers retain safety control because the language model modifies underlying algorithm parameters rather than directly driving the car.

Traditional motion planners in self-driving cars are pre-tuned by engineers, leaving passengers with virtually no ability to customize how a vehicle drives during a trip. To solve this limitation, researchers led by Diego Martinez-Baselga at Delft University of Technology integrated OpenAI's GPT-4o-mini model with a motion-planning controller. Instead of placing the language model in direct control of steering or acceleration, the system interprets statements such as "I am running late, go fast" to alter the weighting parameters of a safety-bounded motion planner.
Because large language models lack deterministic safety guarantees and suffer from slower response times, using them for real-time vehicular control presents significant risks. By restricting the artificial intelligence model to translating fuzzy human preferences into scenario-based parameter updates, the vehicle remains safely inside established control bounds. Before any adjustments take effect, the vehicle describes the proposed behavior changes in plain language and asks the passenger to confirm, ensuring clear communication between the user and the automated control system.
Delft University of Technology researchers developed a system that lets passengers use natural language to customize autonomous driving styles. The system utilizes OpenAI's GPT-4o-mini model to translate spoken requests into mathematical parameter adjustments for safety-aware motion planners. Passengers retain safety control because the language model modifies underlying algorithm parameters rather than directly driving the car.
For more details please read the original article at IEEE Spectrum AI.
Continue Learning
Comments
Comments appear only after moderation. Your email identifies your submission to the moderator and is never displayed here.
No approved comments yet.