Notes Describe Work in Progress on World Action Models and Object Permanence

Short entries posted around the same time point to research on world models and robot manipulation. The entry titled “DeltaWAM: Delta World Action Models for Bimanual Manipulation” says the work is in progress and that the authors are preparing the code.

Another entry, “World Action Agent: Harnessing VLMs for Robot Manipulation via World Action Rehearsal,” states that general-purpose vision-language models (VLMs) bring broad knowledge and spatial reasoning to robot manipulation. It notes that existing systems either use VLMs directly or, as the note indicates, point to another direction; the title links that direction to world action rehearsal.

A third entry, “Training Object Permanence in World Models,” says object permanence is the foundation of human cognition. In general, object permanence is the idea that objects continue to exist even when they are not visible. The entry describes a very complete data infrastructure composed of a very diverse set of object permanence cognitive tasks, and says that each task has a Blender-based data generator allowing one to scale each task to at least 10,000.

Taken together, the entries describe an emerging focus on scalable data and world-model methods for manipulation, while indicating that at least one project is still being prepared.