Researchers Develop AI System That Learns Individual Driving Preferences

Autonomous driving systems today largely replicate patterns from training data, but they struggle to capture the rich variety of how humans actually drive or to understand what individual drivers prefer. Researchers have now proposed a solution: a machine learning framework that interprets user intent and generates driving behaviors customized to match personal preferences.

The Challenge with Current Systems#

Existing imitation learning-based autonomous planners tend to flatten human driving behavior into a single mode, missing the natural diversity in how people handle the road. They also fail to grasp contextual factors—why a driver might choose a cautious route one moment and an aggressive maneuver the next. For real-world deployment of autonomous vehicles serving different users with different needs, these limitations are critical.

How the New Approach Works#

The team developed what they call an M-Diffusion Planner, a multi-strategy framework that pairs a diffusion-based motion planning model with language model capabilities. The system works in two stages:

• Stage one uses imitation learning to train multiple "policy heads" (different driving strategies) on high-quality, safe trajectories
• Stage two applies a technique called constrained Group Relative Policy Optimization (GRPO) to fine-tune each strategy head so it aligns with user preferences

The LLM integration allows the system to understand semantic information about the driving context and user intent, moving beyond raw sensor data.

Testing and Performance#

The framework was evaluated on nuPlan, a standard autonomous driving benchmark, under both simulation (open-loop) and real-world-like conditions (closed-loop). Results show the system achieves competitive performance while staying within real-time planning constraints and successfully adapting to user preferences—suggesting it could support diverse, personalized driving services in the field.