PersonaDrive: controllable trajectory prediction with multi-dimensional driving personas
PersonaDrive provides a framework that learns driving personas from natural language and generates persona-specific trajectories for autonomous driving. It introduces Persona-Conditioned Anchor Transform and Multi-Modal Fusion to reshape anchors along urgency and comfort axes, and uses hierarchical losses to enforce ordered behavior. The project targets researchers and engineers building predictive planning systems who need fine-grained control over driving style. Compared to existing predictors, it offers multi-dimensional persona control rather than a single urgency level, enabling richer, more customizable behavior.
View on GitHub →VisualAIKHU/PersonaDrive