MVISTA-4D: View-Consistent 4D World Model with Test-Time Action Inference for Robotic Manipulation

arXiv 26.2 2026 53.6 method

TLDR

MVISTA-4D proposes a view-consistent 4D world model for robotic manipulation that generates multi-view RGBD from single view and uses test-time action optimization.

Reasoning

The paper introduces a novel 4D world model with cross-view and cross-modality fusion for geometrically consistent scene generation, and a test-time action optimization to address ill-posed inverse dynamics. Strengths include strong experimental results on multiple datasets and practical ablations; weaknesses include reliance on single-view input and potential computational overhead of test-time optimization.

Read-first score

Read-first score 53.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.

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=ablation,baseline,dataset,experiment,metric

Topical relevance 29%

74.3

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=dataset

Citation impact 18%

18.3

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.18304881

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: 610.

Keyword Scores

world model

generative world model

world dynamics prediction

video world model

model-based reinforcement learning world model

world simulator

interactive world model

Deep Analysis

Innovations

Geometrically consistent, arbitrary-view RGBD generation from a single-view RGBD observation, enabling complete 4D scene dynamics prediction.
Cross-view and cross-modality feature fusion that jointly enforces RGB-depth consistency and geometric alignment across multiple viewpoints.
Test-time action optimization via backpropagation through the generative model to infer a trajectory-level latent, combined with a residual inverse dynamics model for accurate action execution.

Methodology

The proposed MVISTA-4D is an embodied 4D world model that takes a single-view RGBD observation as input and generates multi-view, cross-modality RGBD frames across time. It employs explicit cross-view and cross-modality feature fusion to ensure geometric consistency and alignment. For action inference, a test-time optimization strategy backpropagates through the generative model to infer a trajectory-level latent, which is then refined by a residual inverse dynamics model to produce executable actions. The model is trained and evaluated on three datasets, with baselines and metrics implied but not detailed in the abstract.

Key Results

The model demonstrates strong performance on both 4D scene generation and downstream robotic manipulation tasks, with ablation studies providing practical insights into key design choices.

Resources

Paper