AD-WM improves MPC success via action-discriminative world models
September 25, 2026
AD-WM uses residual latent dynamics and action-recovery regularization to improve counterfactual model predictive control. The method increased hard-start success from 3.7% to 52.0% on OGBench-Cube by forcing the model to better distinguish between alternative actions.
HOW THIS AFFECTS YOU
●
researcherYou can improve the planning capabilities of latent world models by incorporating action-recovery objectives.