SG-WAM: Self-Guided World Modeling in Geometry-Aware Policy Space
TLDR
SG-WAM learns geometry-aware action-conditioned dynamics in policy-derived representation space, achieving 98.5% on LIBERO benchmarks.
Reasoning
The paper introduces a self-guided framework that jointly optimizes latent future prediction, geometric grounding, and action generation, which is a strong contribution. However, the abstract is truncated and lacks detailed comparisons or limitations, and the model size (0.9B) may raise scalability questions.
Read-first score
Read-first score 26.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.
Field roles
Rank sensitivity
Stability: volatile; rank range: 68.
Keyword Scores
Deep Analysis
Innovations
- Self-guided world modeling with learnable dynamics tokens and EMA-generated targets from the policy backbone
- Joint optimization of latent future prediction, geometric grounding, and flow-matching action generation
- Geometry-aware policy representation space that aligns action relevance and scene geometry
Methodology
SG-WAM uses a 0.9B model with learnable dynamics tokens and a Self-Guided World Predictor that forecasts future latent states conditioned on actions. Targets are produced by an exponential moving average copy of the policy backbone, providing stable supervision. Geometric supervision structures the policy image-token representations, and the framework is trained end-to-end with flow-matching action generation.
Key Results
SG-WAM achieves 98.5% average success on LIBERO and 73% on LIBERO-Plus, outperforming strong baselines in both in-distribution and out-of-distribution real-world evaluations.