AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control
Latent world models are typically trained to predict factual transitions, whereas model predictive control (MPC) must compare alternative actions from the same state. A model can therefore achieve low factual prediction error yet poorly distinguish candidate actions. We introduce AD-WM, an action-discriminative joint-embedding world model for counterfactual MPC. AD-WM combines
Reading in Plain English mode — toggle below to switch
Latent world models are typically trained to predict factual transitions, whereas model predictive control (MPC) must compare alternative actions from the same state. A model can therefore achieve low factual prediction error yet poorly distinguish candidate actions. We introduce AD-WM, an action-discriminative joint-embedding world model for counterfactual MPC. AD-WM combines
This paper introduces a novel world model approach that significantly improves the performance of model predictive control in complex environments. For ML engineers, understanding how to build models that can effectively reason about counterfactuals is crucial for developing robust and intelligent agents in real-world applications.