LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

arXiv 2026 56.7 method

TLDR

LaWAM uses latent visual subgoals from a world model for efficient, dynamics-aware robot control, achieving high success rates with low latency.

Reasoning

The paper introduces a novel latent world model that avoids expensive video generation, achieving strong results on benchmarks and real-world tasks. However, it lacks explicit comparison to model-based RL methods and the latent representation's interpretability is unclear.

Read-first score

Read-first score 56.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 32.

Recency 6%

100

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

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.93996694

Methodology quality 18%

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

Topical relevance 29%

45.7

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=code

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 380.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Replacing computationally expensive pixel-level video generation with compact latent visual subgoals for predictive dynamics
Introducing a latent-action-conditioned Latent World Model (LaWM) trained in the latent space of a pretrained vision foundation model
Conditioning action generation on predicted latent visual subgoals to enable dynamics-aware robot control with low latency

Methodology

LaWAM builds a latent world model (LaWM) by training a latent action model in the latent space of a pretrained vision foundation model and repurposing its forward decoder to predict future observation features. The policy then conditions action generation on these predicted latent visual subgoals, avoiding pixel-level video reconstruction. The approach is evaluated on LIBERO, RoboTwin, and real-world manipulation tasks, with baselines including pixel-space World-Action Models.

Key Results

LaWAM achieves state-of-the-art or competitive success rates: 98.6% on LIBERO, 91.22% on RoboTwin, and strong performance on real-world tasks. It runs in 187 ms per action-chunk prediction, achieving up to 24× lower wall-clock latency than pixel-space WAMs.

Limitations

Reliance on a pretrained vision foundation model may limit generalization to visual domains not well represented in its training data
Latent subgoals may discard fine-grained pixel-level information that could be critical for precise manipulation tasks
Evaluation is limited to specific benchmarks (LIBERO, RoboTwin) and a set of real-world tasks; broader generalizability is not demonstrated

Resources

Paper