PersonaDrive: Controllable Trajectory Prediction with Multi-Dimensional Driving Personas

Every time a self-driving car navigates a busy city street, it must make a series of split-second decisions about where to go and how fast to get there. For years, researchers have taught these vehicles to predict the future paths of other cars and pedestrians, but the vehicles themselves have largely been trained to drive in a single, safe, and predictable way. They are programmed to be the ideal, cautious driver who never speeds up unnecessarily and never takes a sharp turn. While this approach is safe, it misses a crucial part of human reality: people drive differently depending on who they are, where they are going, and how they feel. A parent rushing a sick child to a hospital needs a smooth, gentle ride even if they are in a hurry, while a firefighter racing to a blaze might need to accept a rougher, more aggressive path to save time. Until now, autonomous systems have struggled to understand these subtle, conflicting human needs, often defaulting to a one-size-fits-all style that feels robotic and disconnected from the passenger's actual situation.

A team of researchers from South Korea has tackled this problem by creating a new way to teach self-driving cars to understand the personality behind the wheel. They realized that human driving behavior is not just about being fast or slow; it is shaped by two distinct forces. The first is urgency, which dictates how quickly a driver wants to reach their destination. The second is comfort, which determines how smoothly the driver wants to travel, prioritizing a gentle ride over speed. By separating these two factors, the researchers created a new framework that allows a vehicle to be instructed not just to "drive fast" or "drive slow," but to adopt a specific driving persona that balances these competing desires.

To test this idea, the team first had to build a new kind of training library. They took thousands of real-world driving scenes and used advanced language models to generate nine different versions of the same journey. Each version corresponded to a unique combination of urgency and comfort, ranging from a frantic, bumpy ride for someone in a desperate crisis to a slow, ultra-smooth journey for a passenger who is feeling unwell. For each of these nine scenarios, they wrote natural language descriptions that a passenger might actually say to a driver, such as "Please, step on it! My kid's locked in the car," or "Could you please drive gently? I have a hot coffee with me." This collection, which they call the PCT dataset, pairs these text descriptions with the precise physical paths a car should take to match that specific mood.

With this new data in hand, the researchers developed a system they named PersonaDrive. This system acts as a translator, taking the passenger's text request and converting it into specific instructions for the car's steering and acceleration. Instead of treating every instruction as a simple switch, the system learns to adjust the car's path along two separate axes. If a passenger asks for speed, the car learns to cover more distance in less time. If a passenger asks for comfort, the car learns to smooth out its turns and avoid sudden jerks. Crucially, the system can handle requests that mix these needs, such as a passenger who is in a rush but needs a smooth ride because they are carrying fragile medical equipment. The researchers found that by teaching the car to respect both axes simultaneously, it could generate trajectories that were far more accurate to the passenger's intent than previous methods, which often got stuck in a single mode of driving.

The results of their experiments showed that this approach works significantly better than older models. When tested on a standard driving simulation, the new system consistently predicted the correct path for all nine different driving personas, whereas older systems often failed to distinguish between a driver who wanted to be fast and one who wanted to be smooth. The researchers also discovered that simply giving the car a text description was more effective than giving it a simple code or label. The language itself helped the system understand the nuances of the request, allowing it to create a continuous range of driving styles rather than just a few rigid categories. This suggests that the future of autonomous driving may not be about programming a car to be the perfect, emotionless machine, but about teaching it to understand the human stories behind the journey, adjusting its behavior to fit the unique needs of every passenger.