Robotic world models are meant to give a robot a predictive sense of how its actions will move objects, but many current approaches let object shape and motion drift over multiple prediction steps. The authors of this preprint argue that image-based and dense 3D models leave rigid-object geometry to be learned implicitly, which makes long-horizon rollouts unreliable.

Their proposed model, ORD-WM, treats each object as a point cloud and encodes it with a vector-neuron variational autoencoder supervised by signed distance fields. A dynamics model then predicts an SE(3) pose increment for each object, and the model applies that rigid transformation directly to the current latent state. This separates persistent geometry from pose changes, so the model does not have to regenerate an unconstrained next state.

In experiments, the approach produced stable, geometry-preserving multi-step predictions and enabled real-time replanning for closed-loop model predictive control in both simulated and real manipulation tasks. The results come from a single arXiv preprint and have not yet undergone peer review.