Planning with object-centric world models for robotic control

Slot-MPC combines object-centric representations (learned via slot attention) with Model Predictive Control for robot manipulation. The system learns a differentiable dynamics model that reasons about individual objects, then uses gradient-based optimization to plan actions at test time. On simulated robotic tasks, it achieves better performance and planning efficiency than non-object-centric baselines, and outperforms gradient-free sampling-based MPC in offline settings with sparse state-action coverage. Code and results are released.