Reproduce “MWM: Mobile World Models for Action-Conditioned Consistent Prediction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Han Yan, Zishang Xiang, Zeyu Zhang, Hao Tang

MWM (Mobile World Models) improves visual planning for robot navigation by training a world model to generate action-conditioned future observations with enhanced consistency and faster inference. The system achieves a 4x speedup in rollout time while improving multi-step rollout quality and leads to better goal-image navigation performance in real-world scenarios.

Key insights

  • Decomposing world model training into distinct structural learning and consistency refinement stages effectively balances visual fidelity with multi-step autoregressive accuracy.
  • Explicitly training for action-conditioned consistency using self-forcing and perceptual losses significantly reduces error accumulation in multi-step rollouts.
  • Introducing an 'inference-consistent state' during distillation helps preserve sequence consistency when accelerating diffusion models, addressing the mismatch between training and efficient inference.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.