STAGE: STyle-controllable Action GEneration for Personalized Autonomous Driving

Researchers introduced STAGE, a style-controllable action generation method for personalized autonomous driving.

STAGE combines a continuous style value with imitation learning and style/latent action encoding to produce more human-like driving behavior.

Preference learning is used to infer a driver’s style, while rule-based pair comparisons reduce the need for manual labeling.

The method was reported to align well with human expectations in common road scenarios and was evaluated on controllability, continuity, style alignment, and safety.