Event

STRIPS-WM: Learning STRIPS-Style World Models from Image-Only Transitions

1 reports · 1 sources

A CoRL 2026 accepted paper proposes the STRIPS-WM framework, introduced by Abhiroop Ajith and Constantinos Chamzas, which enables robots to learn image-grounded STRIPS-style world models solely from image transitions (current image, executed high-level action, resulting image). In this setting, at test time the robot is given only a start image and a goal image and must output a sequence of high-level actions that reaches the goal. Experiments on visual rearrangement tasks show that STRIPS-WM outperforms the tested visual rollout, latent graph-search, and latent-symbolic baselines in image-to-planning success rate.