PRISM-WM: A Structured World Model for Hybrid Robotic Dynamics
A new architecture for model-based planning in robotic systems featuring hybrid dynamics has been unveiled by researchers, named the Prismatic World Model (PRISM-WM). Traditional latent world models rely on monolithic neural networks that enforce a global continuity, which can lead to the oversmoothing of different dynamic modes (such as sticking versus sliding or flight versus stance), resulting in cumulative errors during long-horizon predictions. PRISM-WM counters this issue by breaking down intricate hybrid dynamics into composable primitives through a context-aware Mixture-of-Experts (MoE) framework. This framework includes a gating mechanism that identifies the current physical mode, while specialized experts forecast the corresponding transition dynamics, ultimately enhancing planning reliability at physical boundaries.
Key facts
- PRISM-WM is designed for model-based planning in hybrid systems.
- Hybrid dynamics involve continuous motion punctuated by discrete events.
- Conventional models over-smooth distinct dynamic modes.
- PRISM-WM uses a Mixture-of-Experts framework.
- A gating mechanism identifies the current physical mode.
- Specialized experts predict transition dynamics.
- The architecture decomposes dynamics into composable primitives.
- The paper is available on arXiv with ID 2512.08411.
Entities
Institutions
- arXiv