Researchers led by John Won, with Jinwoo Shin among the authors, have built AFFECT, a training objective that teaches robot world models how small changes in actions alter what happens next, according to a paper on arXiv. The authors add correlated noise to a policy's actions and train the model so that its generated videos respond as the recorded ones do.
In simulation and in real-world DROID and ALOHA-2 setups, the world models predicted outcomes more faithfully. On DROID, the correlation between a policy's success in the world model and on the real robot rose from 0.47 to 0.79. Choosing actions with these models improved task success in planning by up to 8.3%.



