Reproduce “Sword: Style-Robust World Models as Simulators via Dynamic Latent Bootstrapping for VLA Policy Post-Training” 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

Jiaxuan Gao, Yongjian Guo, Zhong Guan, Wen Huang, Wanlun Ma, Xi Xiao, Junwu Xiong, Sheng Wen

Sword trains a world model to simulate robot observations under visual styles and its own predicted context. It uses appearance augmentation and Dynamic Latent Bootstrapping, DLB, to reduce mismatch between training and inference. On LIBERO-Mixed, Sword achieves FVD 111.19 versus WoVR’s 198.84; lower is better. With GRPO, success reaches 61.72% versus 53.12% at step 172 in LIBERO-Spatial, not real-robot deployment.

Key insights

  • Disentangling visual styles from underlying physical dynamics via structured augmentation significantly improves a world model's generalization to unseen visual conditions.
  • A dynamic, self-updating cache for latent predictions enables a smoother and more effective alignment of training and inference distributions, stabilizing long-horizon rollouts.
  • A world model that consistently preserves structural and semantic fidelity under visual variations and over extended temporal sequences provides a more robust and effective simulator for VLA policy optimization.

Summary by alphaXiv.

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