Motivation

🙁 MPC alone doesn't work well for long-horizon tasks. 

It requires reference trajectories that are expensive to compute online

Our contirbutions

✅ A diffusion model that generates subgoals dynamically in a coarse-to-fine manner, trained by random play data

✅ A reachability-based strategy that determines appropriate temporal resolution, focusing on the challenging part of the task

✅  A system that integrates subgoal generation and MPC for robot manipulation.