The paper introduces ChronoGraph, a functional 4D scene graph that represents how actions on affordance parts change both semantics and geometry. The same graph structure handles observed transitions (understanding past interactions) and anticipated ones (planning future actions). To train and evaluate this capability, the authors built ChronoGraphBench, an automatic data engine that converts human-interaction videos and simulated robot trajectories into graph-annotated questions. They then developed ChronoGraphVLM, a trained in two stages: Graph-as-Chain-of-Thought supervised fine-tuning teaches it to reconstruct and predict graph traces, followed by joint 4D graph reinforcement learning that rewards graph correctness and answer accuracy.
Experiments across multiple model scales show that ChronoGraphVLM outperforms pretrained baselines on both interaction understanding and grounded planning tasks. It also transfers zero-shot to the VLM4D benchmark. Real-world mobile manipulation demonstrations confirm that the model's graph-based planning and affordance grounding can drive existing robot skills without additional fine-tuning. The work is interesting because it directly connects two traditionally separate aspects of embodied intelligence—understanding how past actions changed the scene and predicting how future actions will change it—within a single representation and learning framework.