HERMES++: Toward a Unified Driving World Model for 3D Scene Understanding and Generation
HERMES++ proposes a unified driving world model that combines 3D scene understanding with future geometry prediction. It uses a Bird's Eye View (BEV) representation to integrate multi-view data and LLMs for semantic reasoning. A key innovation is the Current-to-Future Link, which predicts scene evolution over three seconds by conditioning geometric changes on semantic context and ego-motion, leading to structurally consistent future predictions.