Event

ORD-WM proposes object-centric rigid-body dynamics 3D world model

1 reports · 1 sources

Yueyang Weng et al. propose ORD-WM, a 3D world model for robotic manipulation tasks. It uses a vector-neuron variational autoencoder supervised by signed distance fields to encode each object's point cloud into a geometry-aware latent representation, then employs an action-conditioned dynamics model to predict per-object SE(3) motion increments, and analytically applies rigid-body transformations to the current latent state rather than generating an unconstrained next latent state, thereby separating persistent geometry from pose changes. The paper reports that experiments show the model achieves stable, geometry-preserving multi-step predictions in simulated and real manipulation scenarios, and supports real-time replanning for closed-loop model predictive control.