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

Summary

A preprint introduces AD-WM, an action-discriminative joint-embedding world model for counterfactual model predictive control. Standard latent world models predict factual transitions, which can leave them poor at telling candidate actions apart. AD-WM adds residual latent dynamics and an action-recovery regularizer based on inverse dynamics. On OGBench-Cube it raises hard-start success from 3.7% to 52.0% over a matched baseline. It also improves mean success in four of five simulation environments. With a frozen V-JEPA 2 encoder and DROID post-training, basic pick-and-place success on a Franka setup rises from 42.2% to 71.1% zero-shot.