Event

DynaSlots: A Slot-Level 3D Dynamics Pretraining Framework

1 reports · 1 sources

Meng Li and co-authors released DynaSlots, a 3D visual pretraining framework that requires no action labels. It decomposes each frame's point cloud into a small number of temporally consistent slots and models dynamics at the slot level rather than the scene level: a temporal slot propagation module keeps slot assignments consistent across frames, an inverse dynamics model extracts each slot's latent action from feature differences, and a diffusion-model forward dynamics model together with a self-supervised interaction head predicts each slot's future state and learns pairwise adjacency and motion relationships. The authors report that experiments on Meta-World, Adroit, and the RealMan RM65-B robot show that DynaSlots outperforms prior scene-level baselines, achieving higher success rates in simulation and exhibiting strong robustness to scene clutter and unseen objects in real manipulation. These results are as reported by the authors.