DSWAM: A Dual-System World Action Foundation Model for Fine-Grained Robot Manipulation

arXiv 2026 36.1 method

TLDR

DSWAM combines a world action model executor with a vision-language subtask planner for fine-grained robot manipulation, enabling task decomposition.

Reasoning

The paper addresses a clear gap in WAM-VLA comparison and introduces a dual-system approach. Strengths include real-robot experiments and a novel decomposition method; weaknesses are limited detail on the executor's video co-training and lack of explicit future video generation.

Read-first score

Read-first score 36.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 30.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,evaluation,result

Topical relevance 29%

42.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 171.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Dual-system architecture combining a System 1 World Action Model executor with a System 2 vision-language subtask planner for task decomposition
Fair real-robot comparison framework between WAM and VLA policies under matched conditions (DeMaVLA setting)
Integration of TensorRT acceleration, asynchronous execution, and real-time chunking (RTC) for practical real-robot deployment

Methodology

DSWAM uses a System 1 WAM executor trained with action prediction and video co-training, which directly predicts action chunks at inference without explicit video generation. A System 2 vision-language planner optionally decomposes coarse instructions into executable subtasks using short-term visual history and a global task prompt. The system is evaluated on a real-robot deformable manipulation benchmark (DeMaVLA) with matched robot, data, and evaluation criteria to enable a fair comparison with VLA policies.

Key Results

The abstract does not report specific quantitative results; it describes the system design and the setup for a fair real-robot comparison between WAM and VLA.

Resources

Paper