Model predictive control turns control into repeated optimization: at every tick, solve for the best short trajectory under a dynamics model and cost function, apply only the first action, then re-solve from the new state. The receding horizon means model errors never compound for long, because the plan is refreshed with fresh measurements many times per second.
MPC is the crown jewel of the classical robotics stack, powering model-based legged locomotion and precise industrial motion. Its dependence on an explicit dynamics model and cost function is exactly what learning-based methods relax: where the model is unknowable (contact with unknown objects, unstructured scenes), learned policies take over, and hybrid systems increasingly use learning to propose and MPC to guarantee.
