Summary
OA-WAM decomposes each frame into N+1 slot states (one robot slot + N object slots), where each slot has a persistent address vector and a time-varying content vector. This addressability lets the action decoder reliably reference specific objects even under scene shifts, solving the identity-entanglement problem in holistic WAMs. A world head predicts next-frame slot states while a flow-matching action head decodes a 16-step action chunk in the same forward pass, with slot attention enforced through address-only key routing. OA-WAM matches strong VLA and WAM baselines on LIBERO (97.8%) and SimplerEnv (79.3%) and achieves SOTA on the most challenging LIBERO-Plus geometric axes, with causal slot-intervention cosine 0.87 vs. ≤0.09 for holistic baselines.
Key Contributions
- Object-addressable slot representation separating persistent identity (address) from time-varying state (content)
- Cross-slot attention routed through address-only keys, enforcing structural binding without extra tokens
- Unified world + action decoding in a single forward pass via flow-matching action head
- LIBERO 97.8%, SimplerEnv 79.3%; high causal binding score validating object tracking
Significance
OA-WAM is the first WAM with explicit object-level addressability, showing that structured scene representations improve both manipulation accuracy and interpretability over holistic video-token approaches.