Slot-MPC: Goal-Conditioned Model Predictive Control with Object-Centric Representations

Researchers introduced Slot-MPC, a goal-conditioned control framework that combines slot-based object representations with an action-conditioned world model.

The method uses differentiable model predictive control at inference time to plan actions efficiently, rather than relying on sampling-heavy search.

In simulated robotic manipulation tasks, Slot-MPC beat non-object-centric baselines and gradient-free MPC, especially when offline data coverage was limited.

The results suggest that object-level scene representations can improve generalization to unseen situations while reducing compute for planning.