Tsinghua & NVIDIA propose PrismWAM, splitting semantics, physics and actions in world models
jiqizhixin · x · 2026-11-04
Tsinghua University and NVIDIA researchers propose PrismWAM, a world model that separates a scene's appearance from its physical state, splitting semantic planning, physical rollout and action generation.
- Division of labor: RGB features describe task context and goals; depth-based representations model geometric changes; chain attention passes information from semantics to predicted physical states, then to actions, rather than letting the action module read RGB directly.
- Results, with caveats: the exploratory 15B PrismWAM reaches 75.50% success on RoboTwin's randomized condition (83.67% average across clean and randomized) — but those headline numbers belong to the larger variant, not the 6B main model.
- Precise pretraining claims: the method avoids embodied action pretraining.
The accompanying figure shows the architecture: RGB VAE / Depth VAE towers and the chain-attention masks used at train/inference time.
More from Embodied
- Baidu unveils Apollo World, a physical AI world model for autonomous driving simulation — Baidu_Inc · 2026-11-04
- Gaussian splatting game built with Playcanvas: real2sim as a training ground for robotics — willeastcott · 2026-11-04
- Multi-fingered robotic hand demos coordinated bends, real dexterity still unproven — burhop · 2026-11-04
- This robot runs an entire 3D printing farm unattended, print after print — lukas_m_ziegler · 2026-11-04
- Liber-0 hits 36% zero-shot success across 60 robot manipulation tasks without task-specific demos — heyshrutimishra · 2026-11-04
- 30 days, six teams: lessons from delivering a full humanoid robot system on a tight timeline — yongqianme · 2026-11-04