Scene2Demo, described in a new arXiv paper, is a system that automatically generates embodied data offline. It takes a single RGB image of a scene and a user query as inputs, then builds an interactive simulated scene and produces executable tasks.
Because the generation happens offline, the data is created ahead of time rather than during live task execution. The abstract does not elaborate on specific applications or evaluation results, so the practical scope of the framework remains unclear from the announcement alone.
The paper's full title mentions a self-evolving mechanism and an object-action graph, but the abstract does not describe how these work. The announcement focuses on the framework's inputs and outputs, leaving the internal details to the full manuscript.