HaM-World: Structured World Model for Stable Planning
A new study has unveiled HaM-World (HMW), a structured model designed to improve stability in planning based on models. The researchers argue that the instability seen in simulated rollouts results from a lack of structure in the latents that planners use, particularly in memory that relies on history to ensure Markov completeness. HMW separates the latent state into two parts: a canonical subspace (q, p) and a context subspace c, using Mamba's state-space memory for historical input. The (q, p) subspace evolves through a Hamiltonian vector field driven by energy, while the context includes various semantic elements. This method gives planners a cohesive latent state for predicting dynamics and calculating rewards. The study can be found on arXiv with ID 2605.05951.
Key facts
- HaM-World (HMW) is a structured world model for planning.
- It addresses instability in imagined rollouts due to missing history-conditioned memory and geometric organization.
- The latent state is decomposed into (q, p) subspace and context subspace c.
- Mamba selective state-space memory provides history-conditioned input.
- (q, p) evolves via Hamiltonian vector field plus residual/control dynamics.
- c captures semantic, dissipative, and non-conservative factors.
- The model provides a single latent state for dynamics prediction and reward.
- The paper is on arXiv with ID 2605.05951.
Entities
Institutions
- arXiv