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

SCMA: Self-Consistent Model-based Adaptation for Visual Reinforcement Learning

IJCAI 25 2025 47.9 method

TLDR

SCMA uses a denoising model and pre-trained world model to adapt visual RL agents to distractions without policy modification.

Reasoning

Strengths include a novel unsupervised adaptation method with theoretical optimality and real robot experiments. Weaknesses are reliance on a pre-trained world model and potential limited generalization to unseen distractions.

Read-first score

Read-first score 47.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 18.

Methodology quality 25%
90

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

Recency 8%
86.7

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

Reproducibility 25%
30

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

Topical relevance 42%
25.7

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 505.

Keyword Scores

world model
6
model-based reinforcement learning world model
5
generative world model
2
world dynamics prediction
2
world simulator
1
interactive world model
1
video world model
1

Deep Analysis

Innovations

  • Proposes Self-Consistent Model-based Adaptation (SCMA) that uses a denoising model to transfer cluttered observations to clean ones without modifying the policy, acting as a plug-and-play enhancement.
  • Derives an unsupervised distribution matching objective with a theoretical analysis of its optimality for training the denoising model.
  • Presents a practical algorithm that estimates the distribution of clean observations using a pre-trained world model to optimize the unsupervised objective.

Methodology

SCMA employs a denoising model to transform cluttered visual observations into clean ones, enabling robust adaptation for visual reinforcement learning agents without altering the policy. The denoising model is trained via an unsupervised distribution matching objective, which is theoretically justified, and the distribution of clean observations is estimated using a pre-trained world model. The method is evaluated on multiple visual generalization benchmarks and real robot data, comparing against baselines that use hand-crafted augmentations.

Key Results

SCMA effectively boosts performance across various visual distractions and exhibits better sample efficiency compared to existing methods on visual generalization benchmarks and real robot data.

Limitations

  • The method relies on a pre-trained world model, which may not be available or easy to obtain in all environments.
  • The unsupervised distribution matching objective may not perfectly handle all types of visual distractions, potentially limiting generalization to unseen distortions.

Resources

Paper

Tags