DeXplicit: Making Physical State Explicit in Dexterous Manipulation World Models
Before acting on an object, a robot needs to anticipate how its hand motion will change the physical world and what the resulting scene will look like. Existing approaches typically entangle action-conditioned dynamics, camera motion, and visual synthesis within a video generation model, or represent dynamics as point trajectories without explicit object structure. We introduce DeXplicit, a dexterous manipulation world model that predicts how the world evolves explicitly in metric 3D before rendering its visual consequences. Given an initial scene and a future hand trajectory, a lightweight 3D dynamics network predicts object-centric motion and deformation, providing structured and geometrically grounded future world states. These predictions guide video generation through appearance transport and geometric projection, allowing the renderer to focus on uncertain and newly visible content rather than implicitly learning object dynamics from pixels. To retain these controls in an efficient autoregressive generator, we develop a two-stage causal adaptation procedure: alignment of the transferred geometry, hand, and appearance conditions to a frozen few-step backbone, followed by on-policy distillation with a conditional video critic. The resulting model generates four-frame latent blocks with two denoising steps while conditioning on its generated history. Across HOT3D, OakInk2, and PhysTwin, DeXplicit reduces moved-point error by 27-40% relative to PointWorld and improves object IoU+ by 47% on HOT3D under a shared renderer. Causal-DeXplicit retains an object IoU+ of 0.490 while accelerating video synthesis by 57 times relative to our 14B bidirectional configuration.