World Action Models give robots the ability to simulate consequences before they move

Researchers from China and Singapore published the first systematic survey of World Action Models for robotics.

The paper reviews about 100 studies, grouping them into cascaded and joint architectures that link world-state prediction with robot control.

It also compares where these models get training data, from teleoperation and action-labeled sets to large-scale video sources.

The authors argue WAMs could help robots generalize better to new objects and environments by learning both future change and action generation.