AIM reportedly matches end-effector recall with 85× fewer training episodes in a robot task

The researchers say world models trained on expert demonstrations can predict plausible futures while ignoring the commanded action. Their Action Influence Maximization (AIM) method actively collects data to better capture what actions do. In a PushT manipulation test, they report that a model trained on 3,000 expert episodes plus 500 AIM-collected episodes matched the end-effector recall of one trained on 300,000 expert episodes—roughly 85× fewer training episodes.