AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control

Source facts

AD-WM is a joint-embedding world model designed for counterfactual model predictive control. It combines residual latent dynamics with action-recovery regularization so the model preserves action-dependent differences needed to compare candidate actions. The abstract reports improvements over a matched baseline on OGBench-Cube and four of five simulation environments, plus zero-shot transfer improvements on a Franka pick-and-place setup with a frozen V-JEPA 2 encoder.

Analysis

The paper challenges evaluating planning world models only by factual prediction error. For control, a model can predict the observed transition well while still ranking alternative actions poorly; action-discriminative objectives target the decision boundary that matters to closed-loop behavior.

Uncertainty/Doubt

• This is an arXiv preprint. The reported simulation, transfer, and baseline comparisons were not independently reproduced here; the project page and code availability were not audited beyond the arXiv record.