World Models for Embodied Intelligence: From Plausible to Controllable to Actionable

Submitted to IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2026

Abstract

This survey asks what makes a world model useful for embodied decisions. It organizes the literature into three levels: models that preserve plausible structure, models that predict the effects of interventions, and models whose predictions measurably improve an agent’s actions or learning. It then examines methods and evaluation across manipulation, navigation, locomotion, driving, and other embodied settings.