3DFlowAction: Learning Cross-Embodiment Manipulation from 3D Flow World Model

arXiv 25.6 2025 68.1 method, application

TLDR

Proposes 3DFlowAction, a 3D flow world model for cross-embodiment manipulation, predicting object motion from video diffusion to guide robot actions.

Reasoning

The paper introduces a novel 3D flow world model that leverages video diffusion to predict object motion, offering an embodiment-agnostic representation for manipulation. Strengths include the large synthetic dataset and flow-guided planning, but weaknesses are reliance on synthetic data and lack of real-world validation in the abstract.

Read-first score

Read-first score 68.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=dataset,github

Topical relevance 42%

61.4

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 25%

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

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 176.

Keyword Scores

world model

world dynamics prediction

generative world model

video world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Learning a 3D flow world model from both human and robot manipulation data to predict future movement of interacting objects in 3D space, which is embodiment-agnostic.
Synthesizing a large-scale 3D optical flow dataset (ManiFlow-110k) via a moving object auto-detect pipeline.
Using a video diffusion-based world model to generate 3D optical flow trajectories conditioned on language instructions.
Proposing a flow-guided rendering mechanism that renders the predicted final state and leverages GPT-4o to assess alignment with the task description for closed-loop planning.
Using predicted 3D optical flow as constraints for an optimization policy to determine a chunk of robot actions for manipulation.

Methodology

The authors synthesize a large-scale 3D optical flow dataset (ManiFlow-110k) using a moving object auto-detect pipeline. A video diffusion-based world model learns manipulation physics from these data, generating 3D optical flow trajectories conditioned on language instructions. A flow-guided rendering mechanism renders the predicted final state and uses GPT-4o to assess whether the predicted flow aligns with the task description, enabling closed-loop planning. Finally, the predicted 3D optical flow serves as constraints for an optimization policy to output a chunk of robot actions.

Key Results

Extensive experiments demonstrate strong generalization across diverse robotic manipulation tasks and reliable cross-embodiment adaptation without hardware-specific training.

Resources

Paper Code