Robots can plan more effectively if they anticipate how a scene will change. WorldFlow, reported on October 29, predicts both future video frames and the movement of selected points in three-dimensional space^[1]. This is a predictive world model^[2], not a system that directly commands motors.

The researchers trained the joint architecture on 1.44 million object-level samples. Their WorldFlowBench evaluation set contains 1,678 samples from 11 sources. On SpatialCog, average displacement error^[3] was 0.060 metres, a 54.2% reduction compared with the strongest external baseline in the authors' tests.

A real deployment would still require latency measurement, camera calibration checks, occlusion robustness and analysis of accumulated error. Forecasting geometry accurately does not itself demonstrate safe or reliable manipulation.