⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

MinD: Unified Visual Imagination and Control via Hierarchical World Models

arXiv 25.6 2025 58.6 method

TLDR

MinD proposes a dual-system hierarchical world model with asynchronous diffusion for real-time robotic planning, achieving high success and failure detection.

Reasoning

Strengths include a novel dual-system architecture using single-step latent features for efficiency and real-world validation on Franka tasks. Weaknesses are limited detail on the DiffMatcher co-training strategy and potential scalability concerns.

Read-first score

Read-first score 58.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.

Recency 8%
86.7

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

Topical relevance 42%
81.4

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%
40

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 25%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 639.

Keyword Scores

world model
10
generative world model
9
video world model
9
interactive world model
8
world dynamics prediction
8
world simulator
7
model-based reinforcement learning world model
6

Deep Analysis

Innovations

  • Dual-system world model with two asynchronous diffusion processes: low-frequency visual generator (LoDiff) and high-frequency diffusion policy (HiDiff)
  • Key insight: robotic policies can use low-resolution latents from a single denoising step instead of fully denoised frames
  • DiffMatcher: video-action alignment module with co-training strategy to synchronize the two diffusion models
  • Real-time risk-aware planning with failure prediction (74% detection rate)

Methodology

MinD proposes a hierarchical world model with two asynchronous diffusion processes: LoDiff for low-frequency visual generation of future scenes and HiDiff for high-frequency action output. It uses a co-training strategy with DiffMatcher to align video and action representations. The model is trained on RL-Bench and evaluated on real-world Franka tasks, operating at 11.3 FPS.

Key Results

MinD achieves a 63% success rate on RL-Bench, 60% on real-world Franka tasks, and operates at 11.3 FPS. It identifies 74% of potential task failures in advance.

Resources

Paper

Tags