Robot-Factored World Models via Robot Rendering

arXiv 2026 41.6 method

TLDR

Proposes robot-factored world models that render robot geometry via URDF and use nominal trajectories to avoid action-realization learning and future-state leakage.

Reasoning

The paper introduces a novel factorization of robot-specific factors from video world models, with a clear methodological contribution. However, the abstract does not mention real-world experiments or benchmarks, so empirical validation is not evident from the visible text.

Read-first score

Read-first score 41.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 6%

100

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

Methodology quality 18%

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

Topical relevance 29%

61.4

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 impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 289.

Keyword Scores

world model

video world model

world dynamics prediction

generative world model

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Robot-factored world models that move action realization and robot rendering outside the model: actions are converted into nominal trajectories via the robot's own controller and kinematics, then rendered using the URDF, avoiding both action-realization learning and future-state leakage.
Rendered robot geometry as a visual interface, factoring out the robot's geometry, kinematics, and appearance from the world model, and creating a shared representation that stays consistent across viewpoints and robot embodiments.
Depth disambiguation by pairing end-effector depth with scene depth, providing geometric cues for contact and occlusion beyond image-plane overlap.
Cross-embodiment transfer from human demonstrations: robot manipulation videos can be generated by retargeting and rendering human hand motion as robot geometry, demonstrating generalization to unseen embodiments.

Methodology

The world model receives a static RGB/depth scene context and rendered robot geometry images derived from actions. Actions are realized into nominal trajectories through the robot's controller and kinematics, then rendered via the URDF into visual robot geometry, removing the need to learn the robot's appearance or kinematics. End-effector depth is paired with scene depth to resolve depth ambiguity. The model is trained to predict future video frames, and its performance is compared against vector-conditioned baselines and tested on unseen robot embodiments.

Key Results

The rendered interface outperforms vector-conditioned baselines and generalizes to unseen robot embodiments at inference. The model also generates robot manipulation videos from human demonstrations by retargeting and rendering hand motion as robot geometry.

Resources

Paper