RAIN: Robotic Task Generalization with Region-Aware Interaction Networks

RAIN is a new method for robotic task generalization that formulates tasks as sequences of goal-conditioned subtasks, each specified by an action type and a target-region mask. It uses a Target-adaptive Cross-View Encoder and a PlanDiT module to predict target-directed trajectories and actions, along with a Transition Head to enable automatic subtask transitions. The Reference-based Retargeting Strategy allows smooth adaptation to new target regions without additional data collection. Experiments on LIBERO-Analogy and real-world tasks show improved generalization compared to existing VLA baselines.