FlowMPC: Improving Flow Matching policies with World Models

arXiv 2026 55 method, application

TLDR

FlowMPC combines flow matching policies with a learned world model for test-time planning, improving manipulation task success.

Reasoning

The paper clearly demonstrates that adding a world model (TD-MPC2) to an imitation-learned flow matching policy improves performance in simulated manipulation tasks. Strengths include a clean integration and empirical gains; weaknesses include limited task scope and no modification of the FM training objective.

Read-first score

Read-first score 55, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 31.

Recency 6%

100

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

Citation impact 18%

94.7

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.94737915

Methodology quality 18%

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

Topical relevance 29%

44.3

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=none

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: 388.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

interactive world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

Combining Flow Matching (FM) policies with a learned world model for test-time planning via Model Predictive Path Integral (MPPI)
Introducing FlowMPC, a framework that integrates an imitation-learned FM policy with a TD-MPC2-based world model without modifying the FM training objective
Demonstrating that world-model-based planning can effectively complement flow-based imitation policies in manipulation tasks

Methodology

FlowMPC uses a pretrained Flow Matching policy for behavior cloning and a learned world model (based on TD-MPC2) to perform MPPI planning over candidate action sequences proposed by the FM policy at test time. The approach is evaluated on ManiSkill manipulation tasks (PickCube and PickSingleYCB) against the FM policy alone, with success rate as the primary metric.

Key Results

Adding the world model improved performance over the FM policy alone on both PickCube and PickSingleYCB, with especially clear gains in end-of-episode success.

Limitations

Evaluation is limited to only two manipulation tasks (PickCube and PickSingleYCB), so generalization to other tasks is unknown
Requires a separately learned world model, which adds complexity and may suffer from model inaccuracies
The approach does not modify the FM training objective, potentially limiting the extent of improvement achievable

Resources

Paper