MetaWorld: Skill Transfer and Composition in a Hierarchical World Model for Grounding High-Level Instructions

arXiv 26.1 2026 42.2 method

TLDR

MetaWorld proposes a hierarchical world model with VLM-driven semantic planning and expert policy transfer for humanoid robot loco-manipulation, outperforming world model-based RL on Humanoid-Bench.

Reasoning

The paper introduces a novel hierarchical world model that bridges semantic planning and physical control via expert policy transfer, addressing sample efficiency and generalization issues. However, experiments are limited to a simulated benchmark (Humanoid-Bench) without real-world validation, and reliance on a pre-trained expert library may constrain adaptability.

Read-first score

Read-first score 42.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 20.

Recency 8%

100

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

Methodology quality 25%

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

Reproducibility 25%

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

Topical relevance 42%

28.6

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

Frontier

Rank sensitivity

Stability: volatile; rank range: 426.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Hierarchical world model integrating semantic planning (VLM) and physical control (latent dynamics model) for humanoid loco-manipulation
Dynamic expert selection and motion prior fusion mechanism leveraging a pre-trained multi-expert policy library as transferable knowledge
VLMs as semantic interfaces to map instructions to executable skills, bypassing symbol grounding

Methodology

MetaWorld decouples tasks into a VLM-driven semantic layer and a latent dynamics model operating in a compact state space. It uses a pre-trained multi-expert policy library as transferable knowledge, with dynamic expert selection and motion prior fusion for efficient online adaptation via a two-stage framework. VLMs serve as semantic interfaces to map high-level instructions to executable skills.

Key Results

On Humanoid-Bench, MetaWorld outperforms world model-based reinforcement learning in both task completion and motion coherence.

Resources

Paper Project