The paper proposes ACID, a decision-time planning framework for embodied control that addresses the issue of unrealizable intermediate transitions in standard planning costs. It introduces cycle action consistency by using an inverse dynamics model to verify that inferred actions match conditioned ones, folding this residual into the planning cost via a scale-invariant adaptive weight.
- The method is evaluated across four action-conditioned world models and six tasks including rigid and deformable manipulation, articulated control, and visual navigation.
- ACID consistently improves planning performance compared to baselines.
- It matches baseline accuracy while requiring substantially less planning compute.
This approach ensures that predicted trajectories remain realistic during environment rollouts, offering a more efficient and reliable planning paradigm for embodied agents.