DynScene: Scalable Generation of Dynamic Robotic Manipulation Scenes for Embodied AI

Figure 1. Types of text-guided data generation for embodied AI. Static scene generation often prevents task completion due to object positioning, regardless of the task requirements (left). Action generation is limited by a lack of interaction diversity, failing to account for the effects of surrounding environments (middle). To the best of our knowledge, we are the first to propose a method for dynamic scene generation that enables more realistic and diverse task interactions by the joint generation of static scenes and robotic actions (right).