Researchers led by Ziyu Zhu, with Shenlong Wang among the authors, have built MoRE, a 3D world model that predicts how objects move when a robot acts on them, according to a paper on arXiv. Within one model it hands rigid, deformable and grasping interactions to separate motion experts.
Trained on a mix of simulated and real data, MoRE works with model predictive control for different manipulation tasks. The authors report better predictions for all three kinds of physics, which led to better real-world control. The paper has been accepted at CoRL 2026.



