Controlled comparison of how to add world models to VLA robot policiesExpert
What Matters for World Modeling in Vision-Language-Action Models?
A preprint from the Chinese Academy of Sciences, Alibaba and FiveAges tests the main ways of teaching robot AI models to predict what happens next, all on the same backbone, and finds that the add-on helps but no single design wins everywhere. Keeping the prediction module separate from the main model worked best for pretrained robot models and raised real-robot success from 45% to as much as 73.8%.