AFFECT from KAIST makes world models dynamics-faithful via action effect matching

John Won · hf · 2026-11-20

KAIST researchers propose AFFECT, an auxiliary objective that trains world models to match visual responses to temporally correlated action noise, addressing their failure to capture how small action changes affect future observations. Across OGBench simulation and real-world DROID and ALOHA-2 setups, AFFECT improves rollout fidelity and planning success without requiring repeated executions from the same initial state.

Original post →

More from Embodied

Embodied channel →