ACID: Action Consistency via Inverse Dynamics for Planning with World Models

ACID addresses a weakness in decision-time planning with action-conditioned world models: terminal-state error can look good even when intermediate predicted transitions are not realizable. It adds a cycle action-consistency residual, requiring an inverse dynamics model to recover the action that produced each predicted transition. A scale-invariant adaptive weight incorporates this residual into the planning objective. The authors report consistent improvements across four action-conditioned world models and six tasks covering rigid and deformable manipulation, articulated control, and visual navigation, while matching baseline accuracy with substantially lower planning compute.