Generating varied scenarios through simulation is crucial for training and evaluating safety-critical systems, such as autonomous vehicles. Text-to-Drive (T2D) synthesizes diverse driving behaviors via LLMs, offering a scalable method for simulating various driving interactions. The paper showcases how T2D generates more diverse trajectories than baselines and provides a user-friendly interface for human preference incorporation.

• Knowledge-driven approach using LLMs
• Synthetic diverse driving behaviors
• Natural language interface for interactive simulations
• Enhancing safety-critical system training