When you watch a ball move, you can imagine where it will be a moment later. A robot has a harder version of that problem: it needs to anticipate both the image and the geometry of motion. WorldFlow, introduced on October 29, predicts future video frames^[1] and the trajectories of selected points in three-dimensional space^[2].

The researchers trained the model on 1.44 million object-level samples from simulated and real scenes. WorldFlowBench, their evaluation collection, contains 1,678 samples from 11 sources. On SpatialCog, average displacement error was 0.060 metres, a 54.2% reduction relative to the strongest external method in the authors' comparison.

That does not mean a robot will always catch a ball or move an object safely. Predicting motion and selecting safe physical actions are different tasks. Real use would require checks on computation speed, camera accuracy and behavior in unexpected environments.