Summary
DSWAM addresses a key limitation of existing World Action Models (WAMs): they excel at physical execution but lack explicit language-level planning for decomposing coarse instructions. DSWAM maintains a System 1 WAM executor as the default control path and optionally activates a System 2 VLM subtask planner only when task decomposition is needed — running in WAM-only mode for atomic instructions and invoking System 2 for coarse multi-step commands.
Key Contributions
- Dual-system architecture: System 1 (WAM executor) runs by default for all tasks; System 2 (VLM planner) is activated selectively only when coarse instruction decomposition is beneficial, preserving efficiency for simple tasks
- System 2 predicts an ordered sequence of executable subtasks from short-term visual history and the global task prompt, enabling fine-grained decomposition of household-level goals
- Foundation model design: trains a single unified WAM with both execution and optional planning paths rather than two separate models
Significance
Shows that the VLA advantage in instruction decomposition and the WAM advantage in physically grounded execution are not mutually exclusive — a dual-system design can capture both, outperforming pure WAMs and VLAs on complex multi-step manipulation benchmarks.