RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation

arXiv 2026 44.7 method, application

TLDR

RynnWorld-4D generates future RGB, depth, and optical flow from a single RGB-D image and language instruction for robotic manipulation.

Reasoning

The paper introduces a novel 4D world model that jointly predicts appearance, geometry, and motion, supported by a large-scale dataset. However, the abstract lacks real-world experimental validation and relies on pseudo-labels, limiting evidence of practical effectiveness.

Read-first score

Read-first score 44.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 47.

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=dataset,experiment,metric

Topical relevance 29%

67.1

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%

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

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: 365.

Keyword Scores

world model

generative world model

world dynamics prediction

video world model

interactive world model

model-based reinforcement learning world model

world simulator

Deep Analysis

Innovations

Proposes synchronized RGB, depth, and optical flow (RGB-DF) as a physically grounded representation that captures 4D scene dynamics, bridging world prediction and policy learning.
Introduces RynnWorld-4D, a unified diffusion model that jointly generates future RGB frames, depth maps, and optical flow from a single RGB-D image and a language instruction.
Develops a tri-branch architecture with cross-modal attention and frame-wise 3D RoPE to ensure consistent evolution of appearance, geometry, and motion.
Curates Rynn4DDataset 1.0, a large-scale dataset of over 254.4 million frames of egocentric human and robotic manipulation videos with pseudo-labels for depth and optical flow.
Proposes RynnWorld-4D-Policy, an inverse dynamics head that uses the frozen internal 4D representations to directly output robot actions in a single forward pass, bypassing multi-step denoising.

Methodology

A generative diffusion model co-produces future RGB, depth, and optical flow from a single RGB-D image and a language instruction, using a tri-branch architecture with cross-modal attention and 3D RoPE. Training leverages Rynn4DDataset 1.0, a massive dataset of 254.4 million frames of egocentric manipulation videos with pseudo-labeled depth and flow. An inverse dynamics policy head is then trained on the model's frozen 4D representations to predict closed-loop robot actions without iterative denoising.

Key Results

RynnWorld-4D generates temporally and spatially coherent 4D predictions, and RynnWorld-4D-Policy achieves state-of-the-art performance on real-world dexterous bimanual manipulation tasks, outperforming others especially on tasks requiring spatial precision and temporal coordination.

Resources

Paper