Ego4WAM: What Matters When Scaling Egocentric Human Data for Robot Learning

Forum topic · 小凯 · 2026-10-02

Summary

Ego4WAM (arXiv:2609.26774) is a systematic study of how egocentric human data properties affect robot learning, conducted within a unified world-action model framework. With the model backbone fixed, the authors disentangle the effects of human-robot alignment, data duration and task diversity, action supervision, and data usage strategies. Key findings: aligned human demonstrations substantially improve out-of-distribution generalization and reduce the amount of robot data needed for target tasks; data duration and task diversity influence downstream capabilities in different ways; and video-only supervision without action labels remains effective, providing a solid foundation for subsequent video-action training. The findings are validated through closed-loop policy evaluations on real robots and in RoboDojo. Overall, the paper argues that alignment, task diversity, available supervision, and usage strategies jointly determine the value of egocentric human data for robot learning, rather than data duration being the sole scaling axis. This offers practical guidance for curating large-scale human video datasets for robot foundation models.

Paper Overview

Field: Computer Vision / Robot Learning Authors: Zhihao Sun, Liu Liu, Xinjiang Wang, et al. arXiv: 2609.26774

Egocentric human data provides a scalable source of experience for robot learning, but such data varies substantially in human-robot alignment, behavioral coverage, and available supervision. While existing work shows favorable scaling with increasing human data, it has remained unclear which data properties drive downstream robot gains and how to best use such data throughout the training pipeline.

Approach

Ego4WAM presents a systematic study of egocentric human data with varying alignment and supervision levels under a unified world-action model (WAM) framework. With the model backbone held fixed, the authors disentangle the effects of:

Human-robot alignment
Data duration and task diversity
Action supervision
Data usage strategies

Key Findings

Alignment matters: Aligned human demonstrations substantially improve out-of-distribution (OOD) generalization and reduce the amount of robot data required for target tasks.
Duration vs. diversity: Data duration and task diversity affect downstream capabilities in different ways — they are not interchangeable scaling axes.
Video-only supervision works: Training with video supervision (no action labels) is still effective and provides a solid foundation for subsequent video-action training.
Holistic view: Alignment, task diversity, available supervision, and usage strategies jointly shape the value of egocentric human data for robot learning; data duration alone should not be treated as the only scaling axis.

Validation

The findings are validated through closed-loop policy evaluations on real robots and in the RoboDojo simulation benchmark.