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

Hierarchical World Models as Visual Whole-Body Humanoid Controllers

arXiv 24.5 2024 51.5 method, application

TLDR

Hierarchical world model for visual whole-body humanoid control using RL, achieving performant policies in 8 simulated tasks without reward design.

Reasoning

The paper presents a novel hierarchical world model approach that avoids hand-crafted rewards and skill primitives, showing strong results in simulation. However, it lacks real-world validation and comparison to baselines, and the abstract does not detail the world model's internal dynamics or generative capabilities.

Read-first score

Read-first score 51.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 32.

Recency 8%
75.1

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

Reproducibility 25%
65

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

Topical relevance 42%
45.7

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

Methodology quality 25%
40

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

Field roles

Candidate

Rank sensitivity

Stability: volatile; rank range: 229.

Keyword Scores

world model
10
model-based reinforcement learning world model
7
world dynamics prediction
5
interactive world model
4
world simulator
3
generative world model
2
video world model
1

Deep Analysis

Innovations

  • Hierarchical world model for visual whole-body humanoid control
  • Fully data-driven approach without simplifying assumptions, reward design, or skill primitives
  • Both high-level and low-level agents trained with rewards

Methodology

The paper proposes a hierarchical world model where a high-level agent generates commands based on visual observations for a low-level agent to execute. Both agents are trained using reinforcement learning with rewards, without any simplifying assumptions, reward design, or skill primitives.

Key Results

The approach produces highly performant control policies across 8 tasks with a simulated 56-DoF humanoid, and the synthesized motions are broadly preferred by humans.

Tags