Long-horizon robot manipulation is usually sensed as high-dimensional images, but the facts that matter for planning are sparse: which actions are possible now and what they change. Classical planners use symbolic operators with preconditions and effects, yet building those representations from raw visual experience has been an open problem.
STRIPS-WM tackles this by learning a grounded STRIPS-style model directly from image transitions. It first induces a finite abstract transition graph from observed (image, action, image) tuples, then learns latent binary predicates and one propositional operator per action label. The resulting operators have sparse preconditions and add/delete effects, and the learned predicates are distilled into a visual encoder so that classical planning can be applied to novel start and goal images.
In visual rearrangement experiments, STRIPS-WM improved image-to-plan success over visual rollout, latent graph-search, and latent-symbolic baselines. The paper was accepted at the Conference on Robot Learning (CoRL) 2026.