FlowMo-WM: A World Model with Object Momentum and Hidden Ambient Drift

arXiv 2026 63.3 method, application

TLDR

FlowMo-WM factorizes image-action history into short-term motion and long-term drift for accurate long-horizon prediction in aquatic vehicles.

Reasoning

The paper introduces a novel factorization of visual history to handle inertia and hidden drift, which is a clear strength. However, it is only evaluated in simulated environments, lacking real-world validation, and the abstract does not discuss limitations or comparisons to non-latent models.

Read-first score

Read-first score 63.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 56.

Recency 6%

100

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

Citation impact 18%

94.2

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.94158875

Topical relevance 29%

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

Methodology quality 18%

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

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

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 484.

Keyword Scores

world model

interactive world model

world dynamics prediction

generative world model

video world model

model-based reinforcement learning world model

world simulator

Deep Analysis

Innovations

Factorizing image-action history into a short-history latent state for object-centric motion and a longer-history context for slowly varying exogenous influences like hidden drift
Zero-context residual transition that separates action-conditioned base dynamics from context-dependent drift effects during latent rollout
End-to-end trainable visual world model that infers object-centric motion state and hidden drift from image-action histories without direct supervision of flow fields

Methodology

FlowMo-WM is an end-to-end trainable visual world model that factorizes image-action history into a short-history latent state (summarizing object-centric motion) and a longer-history context (summarizing slowly varying exogenous influences like hidden drift). A zero-context residual transition separates action-conditioned base dynamics from context-dependent drift effects during latent rollout. The model is trained on simulated aquatic surface-vehicle environments with diverse hidden flows, disturbances, and randomized vehicle dynamics, and evaluated against representative action-conditioned latent world models using long-horizon rollout accuracy.

Key Results

FlowMo-WM improves long-horizon rollout accuracy over representative action-conditioned latent world models in simulated aquatic surface-vehicle environments with hidden flows and disturbances. Prediction-time context ablations (zeroing or shuffling the inferred context) demonstrate that the ambient context is important for stable prediction under hidden drift, and frozen linear probes characterize the information encoded in the learned factors.

Limitations

Evaluated only in simulated aquatic surface-vehicle environments, not on real-world robots or other domains
Assumes hidden drift is slowly varying, which may not hold for rapidly changing exogenous influences
Requires a long history context, potentially increasing computational and memory costs
Limited to object-centric motion and hidden drift; applicability to other types of dynamics or disturbances is not demonstrated

Resources

Paper