Lifting Embodied World Models for Planning and Control

arXiv 2026 43.4 method

TLDR

Proposes lifting embodied world models by training a policy to map high-level actions to low-level joint actions, improving planning efficiency.

Reasoning

The paper addresses high-dimensional action spaces in embodied world models by introducing a lightweight policy for action abstraction, showing significant performance gains. Strengths include a practical solution with interpretable waypoints and empirical validation; weaknesses include limited embodiment scope and lack of real-world experiments.

Read-first score

Read-first score 43.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 29.

Recency 6%

100

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

Citation impact 18%

63.5

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.63533609

Methodology quality 18%

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

Topical relevance 29%

41.4

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

Reproducibility 18%

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

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierBridge

Rank sensitivity

Stability: volatile; rank range: 231.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

generative world model

video world model

world simulator

Deep Analysis

Innovations

Lifted world model that predicts future observations from high-level actions via a lightweight policy composed with a frozen world model
High-level action space defined as a small set of 2D waypoints on the current observation frame for leaf joints (pelvis, head, hands), which are low-dimensional and visually interpretable
Demonstration that searching in high-level waypoint space substantially outperforms searching directly in low-level joint space (3.8× lower mean joint error) while being more compute-efficient and generalizing to unseen environments

Methodology

The authors train a lightweight policy that maps high-level actions (2D waypoints) to sequences of low-level joint actions. This policy is composed with a frozen world model to create a lifted world model that predicts future observations from a single high-level action. The framework is instantiated for a human-like embodiment, with high-level actions defined as near-term goal positions for leaf joints. Evaluation compares the lifted world model against direct search in low-level joint space using mean joint error to the goal pose.

Key Results

The lifted world model achieves a 3.8× lower mean joint error to the goal pose compared to searching directly in low-level joint space, while remaining more compute-efficient and generalizing to environments unseen by the policy.

Resources

Paper