Do Robotic World Models Really Follow Actions? Diagnosing and Aligning Action-Conditioned Generation for Policy Learning

Summary

Action-conditioned world models are increasingly used as learned simulators for policy evaluation and improvement in robotics, but their effectiveness rests on an unverified assumption: that generated futures faithfully reflect arbitrary valid actions. This paper introduces WorldEcho, a diagnostic framework that probes action following over a broader action distribution using visual integrity and SE(3) trajectory alignment. The diagnosis reveals that current world models execute expert actions reasonably well but struggle with diverse off-expert trajectories, either ignoring commanded actions or producing visually invalid rollouts. To address these failures, the authors propose WorldSync, which strengthens action following along three complementary dimensions: distribution coverage, representation grounding, and intervention-effect alignment. WorldSync expands the training distribution of action consequences, uses action forcing to ground intermediate video representations in action-induced robot dynamics, and aligns predicted changes under action interventions with corresponding changes in real futures. Experiments on the RoboTwin benchmark and real-world robot tasks show that WorldSync improves WorldEcho metrics and serves as a more reliable simulator for iterative policy improvement, enabling policies to reach higher success rates. arXiv: 2608.24885.