AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control
Training robot brains to imagine what different actions will do, not just what happens next
Robots that plan their own actions need world models that can reliably tell the difference between what would happen if they did action A versus action B—but standard models optimize for predicting the actual future accurately, which doesn't guarantee this. Researchers built AD-WM, a world model that explicitly trains to preserve information about which action led to which outcome, and showed it boosted task success from 3.7% to 52% on hard-start robot manipulation and improved real-robot pick-and-place from 42.2% to 71.1% without retraining.
Robots that plan using world models are entering real warehouses and labs. If their internal models can't reliably distinguish good actions from bad ones, the robots fail or waste time. This approach makes planning more reliable without slowing down inference, directly improving how well robots solve physical tasks on their first try—critical when resets are expensive or impossible.