ProWAM: Progressive Visual Planning for Long-Horizon Robotic Control
September 30, 2026
ProWAM improves long-horizon robotic prediction by jointly generating actions and an ordered sequence of sparse visual sub-goals. This approach provides explicit visual guidance to anchor action generation, allowing the model to scale by learning sub-goal prediction from large-scale, action-free video datasets.
HOW THIS AFFECTS YOU
●
builderYou can use sparse sub-goals to stabilize long-horizon robotic tasks.
●
researcherThis method enables training on action-free video to improve control capabilities.