Researchers at TU Delft in the Netherlands have developed a new system that uses large language models to convert natural language instructions into motion planning parameters for autonomous vehicles. The system does not directly control driving decisions; instead, it describes behavior changes and requests confirmation from passengers to maintain “human control”. This ensures that vehicle speed and steering smoothness meet users’ vague preferences, such as “being in a hurry”, while ensuring safety. The research team tested the system using OpenAI’s GPT-4o-mini model and previously developed model-based path integral controllers, and found that the vehicle can adjust its driving style according to natural language instructions. The related preprint has been published on arXiv and will be presented at the IEEE Intelligent Transportation Systems Conference in September.
Researchers at TU Delft in the Netherlands have developed a system that uses large language models to convert natural language instructions into parameter adjustments for autonomous vehicle motion planning algorithms. The system does not directly control driving decisions; instead, it describes behavioral changes and requests confirmation from passengers to maintain “human control”. This ensures that vehicle speed and steering smoothness meet users’ vague preferences, such as those related to urgency. The research team tested the system using OpenAI’s GPT-4o-mini model and previously developed model-based path-integral controllers, and found that the vehicle can adjust its driving style based on natural language instructions. The related preprint has been published on arXiv and will be presented at the IEEE Intelligent Transportation Systems Conference in September.