AD3: Implicit Action is the Key for World Models to Distinguish the Diverse Visual Distractors

ICML 24 2024 41.7 method

TLDR

AD3 uses implicit actions of distractors to train separated world models, improving visual control under diverse distractors.

Reasoning

The paper introduces a novel method (IAG and AD3) to handle homogeneous distractors, which are often overlooked. Strengths include addressing a challenging problem and empirical validation. Weaknesses are the lack of real-world experiments and potential overfitting to simulated environments.

Read-first score

Read-first score 41.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 26.

Recency 8%

75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Methodology quality 25%

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

Topical relevance 42%

37.1

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 25%

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

Field roles

Candidate

Rank sensitivity

Stability: volatile; rank range: 238.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

video world model

generative world model

interactive world model

Deep Analysis

Innovations

Proposing Implicit Action Generator (IAG) to learn implicit actions of visual distractors
Introducing AD3 algorithm that leverages implicit actions to train separated world models for distinguishing both heterogeneous and homogeneous distractors
Addressing the previously unexplored challenge of homogeneous distractors that closely resemble controllable agents

Methodology

The paper proposes Implicit Action Generator (IAG) to infer implicit actions of visual distractors, and then uses the AD3 algorithm to train separated world models for task-relevant and task-irrelevant components. The agent's policy is optimized within the task-relevant state space. Evaluation is performed on various visual control tasks with both heterogeneous and homogeneous distractors, comparing against prior model-based methods.

Key Results

AD3 achieves superior performance on various visual control tasks featuring both heterogeneous and homogeneous distractors. The indispensable role of implicit actions learned by IAG is empirically validated.

Resources

Paper