Spatially Aware World Action Model via Geometric Latent Diffusion

arXiv 2026 42.9 method, application

TLDR

SA-WAM repurposes a pretrained video diffusion model to jointly predict actions, RGB, and depth, enabling 3D-aware world modeling for robot policy learning.

Reasoning

The paper's strength is integrating geometric depth into a frozen video diffusion tokenizer, achieving strong results on benchmarks and real-world robot evaluations. Weaknesses include limited visible detail on limitations and reliance on existing benchmarks rather than new datasets.

Read-first score

Read-first score 42.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.

Methodology quality 15%

100

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

Recency 5%

100

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

Topical relevance 25%

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 15%

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

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

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

Citation velocity 10%

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: 676.

Keyword Scores

world model

generative world model

video world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces SA-WAM, a Spatially Aware World Action Model that repurposes a pretrained video diffusion model for joint action, RGB, and depth prediction within a single diffusion backbone.
Uses a nonlinear encoding to map unbounded depth signals into the bounded input domain expected by the frozen VAE tokenizer, enabling reuse of the tokenizer without 3D-specific fine-tuning.
Enables 3D-aware world modeling and action prediction while preserving pretrained visual and physical priors.

Methodology

SA-WAM repurposes a pretrained video diffusion model to jointly predict future actions, RGB observations, and depth in a single diffusion backbone. A nonlinear encoding maps unbounded depth into the bounded input domain required by the frozen VAE tokenizer, avoiding 3D-specific fine-tuning. The approach is evaluated on the RoboCasa and LIBERO-Plus benchmarks and in real-world UR5 robotic arm settings, with comparisons against strong baselines and analysis of prediction quality versus rollout success.

Key Results

SA-WAM achieves state-of-the-art results on the RoboCasa and LIBERO-Plus benchmarks while improving future-state predictions. It also outperforms strong baselines in real-world UR5 robotic arm evaluations, with especially strong gains in randomized environments, and shows a correlation between world model prediction quality and rollout success.

Resources

Paper