AFFECT trains world models to respond to small action changes
John Won, Yonghoon Dong, and Jinwoo Shin submitted a 27-page preprint introducing AFFECT, an auxiliary training objective that teaches world models how visual observations respond to action perturbations. The method injects temporally correlated noise into actions from a task-agnostic human teleoperation policy, avoiding repeated execution from the same initial state. The authors report improved world-model fidelity on OGBench, DROID, and ALOHA-2, with DROID policy-success correlation rising from 0.47 to 0.79 and planning task success improving by up to 8.3%.
Making world models sensitive to small action changes could improve their reliability for robotic policy evaluation and model-based planning, where similar actions can produce qualitatively different outcomes.
Computer Science > Robotics
arXiv:2611.25980 (cs)
[Submitted on 19 Nov 2026]
Title:Learning What Actions Change: Action Effect Matching for Dynamics-Faithful World Models
Authors:John Won, Yonghoon Dong, Jinwoo Shin
View a PDF of the paper titled Learning What Actions Change: Action Effect Matching for Dynamics-Faithful World Models, by John Won and 2 other authors