ChronoGraph is a functional 4D scene graph that links actions on object parts to the semantic and geometric changes they cause, using one form for both observed and predicted transitions. A data engine turns human-interaction videos and simulated robot trajectories into graph-annotated questions (ChronoGraphBench); VLMs are trained to write graph traces as chain of thought and then refined with reinforcement learning on graph properties and answers. The trained models beat their pretrained baselines, transfer zero-shot to VLM4D, and drive mobile manipulation through existing skills on a real robot.
Relevance
ChronoGraph records what changed and why, alongside what is where, as dynamic scenes need. Relevant to 3D scene graphs in multi-robot perception and to long-term autonomy.
Links
- arXiv
- Subfield: Semantic mapping, Perception