OSCAR: Omni-Embodiment Action-Conditioned World Model for Robotics

arXiv 2026 70.1 method, application

TLDR

OSCAR is an action-conditioned video world model that generalizes across robot embodiments for policy evaluation, using a unified skeleton representation and large-scale data.

Reasoning

The paper addresses key challenges in robot world models with a novel data pipeline and conditioning method, showing strong results and correlation with real-world evaluation. However, it relies on finetuning an existing model and lacks details on data curation and limitations.

Read-first score

Read-first score 70.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 61.

Recency 6%

100

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

Citation impact 18%

90.9

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.90927581

Methodology quality 18%

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

Topical relevance 29%

87.1

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 18%

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 393.

Keyword Scores

world model

video world model

generative world model

interactive world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Large-scale standardized data pipeline that curates, filters, and deduplicates broad robotics and egocentric human datasets for joint training
2D kinematic skeleton rendering as a unified conditioning representation that generalizes across different robot arms and human hands
Finetuning the Cosmos-Predict2.5-2B model on a single GH200 GPU, achieving strong performance despite smaller model size and fewer GPUs
Demonstrating significant correlation between virtual policy evaluation in OSCAR and real-world evaluation using RoboArena

Methodology

OSCAR uses a large-scale standardized data pipeline to curate, filter, and deduplicate diverse robotics and egocentric human datasets. It conditions the video world model via 2D kinematic skeleton rendering, which serves as a unified representation across embodiments. The Cosmos-Predict2.5-2B model is finetuned on a single GH200 GPU, and evaluation is performed against baselines (larger models or more GPUs) and by deploying OSCAR to evaluate robot policies from RoboArena, measuring correlation with real-world outcomes.

Key Results

OSCAR achieves significant improvement on action following, appearance quality, and motion consistency compared to existing baselines. Virtual policy evaluation in OSCAR shows significant correlation with real-world evaluation, indicating its potential for replacing physical robot testing.

Limitations

Limited scenario diversity in current robot training datasets remains a challenge that the data pipeline only partially addresses
Imprecise action following is improved but may still have residual errors
Poor generalization across embodiments is mitigated but not fully solved, especially for unseen robot morphologies
The model is finetuned on a single GPU, which may limit scalability or reproducibility on different hardware
Correlation with real-world evaluation is significant but not perfect, leaving uncertainty in fully replacing physical evaluation

Resources

Paper