Summary
HiMem-WAM is a Hierarchical Memory-Gated World Action Model that integrates motion-centric latent actions, high-level skill latents, and boundary-triggered memory updates. A joint hierarchical latent framework learns low-level motion and high-level skill representations simultaneously, while a boundary-aware memory gate writes compact task states at predicted skill transitions—enabling causal inference without test-time video generation or optical flow estimation.
Key Contributions
- Hierarchical latent framework jointly learning low-level motion and high-level skill latents
- Boundary-aware memory gate triggered at predicted skill transitions (not every step)
- Enables robust long-horizon manipulation without test-time video generation
- Evaluated on LIBERO, LIBERO-PLUS, RMBench, and real-world tasks
- Hierarchical latents improve robustness under deployment perturbations
Significance
Hierarchical temporal abstraction in world action models provides better structure for long-horizon tasks; the skill-boundary-triggered memory gate is an elegant way to avoid redundant writes while maintaining causal coherence.