Wh0: Generative World Models as Scalable Sources of Egocentric Human Hand Manipulation Data

arXiv 2026 63.2 method, benchmark, application

TLDR

Wh0 uses generative video world models to produce scalable egocentric human-hand manipulation data, improving dexterous VLA model zero-shot success on real-world tasks.

Reasoning

The paper presents a novel framework that leverages generative world models to address data scarcity for dexterous manipulation, with strong empirical results on 18 real-world tasks. However, the approach is limited to egocentric human-hand data and may not fully address sim-to-real gaps inherent in generative video outputs.

Read-first score

Read-first score 63.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 47.

Recency 6%

100

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

Citation impact 18%

92.7

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.92724431

Methodology quality 18%

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

Topical relevance 29%

67.1

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=artifact,code,dataset,github

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: 411.

Keyword Scores

world model

generative world model

video world model

world simulator

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Using generative video world models as scalable and controllable sources of egocentric human-hand manipulation data
Converting generated videos into robot-trainable supervision through hand motion reconstruction and visual editing
Co-training with limited real robot data to adapt pretrained VLA models for dexterous manipulation deployment
Demonstrating significant improvement in zero-shot success on unseen tasks (from 8.3% to 38.9%) across 18 real-world tasks

Methodology

Wh0 employs a generative world model conditioned on language, objects, and scenes to produce WM-H, a 50k-episode dataset of egocentric human-object interaction videos. The generated videos are then converted into robot-trainable supervision via hand motion reconstruction and visual editing. Finally, pretrained VLA models are co-trained on WM-H and a limited amount of real robot data to adapt to dexterous manipulation tasks.

Key Results

Across 18 real-world dexterous manipulation tasks, Wh0 improves zero-shot success on unseen tasks from 8.3% (model post-trained only on robot data) to 38.9% (with Wh0). Ablation studies confirm that scalable generation and scene/embodiment alignment are key drivers of performance gains.

Limitations

Dependence on the quality and realism of the generative world model, which may produce artifacts or unrealistic interactions
Potential inaccuracies in hand motion reconstruction from synthetic videos, affecting downstream robot training
Requirement for a limited amount of real robot data for co-training, which may still be costly to obtain
Generalization may be limited to tasks and scenes similar to those in the generated dataset

Resources

Paper