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

Zhihao Sun; Liu Liu; Xinjiang Wang; Haoyi Jiang; Wei Feng; Huiqiang Zhang; Xiaosong Jia; Zhizhong Su

cs.ROSep 30, 2026

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

Authors: Zhihao Sun, Liu Liu, Xinjiang Wang, Haoyi Jiang, Wei Feng, Huiqiang Zhang, Xiaosong Jia, Zhizhong Su, +1 more

Organizations: Institute of Trustworthy Embodied AI, Fudan University · Horizon Robotics · Huazhong University of Science & Technology · Zhejiang University of Technology

Abstract

Egocentric human data provides a scalable source of experience for robot learning, but varies substantially in human-robot alignment, behavioral coverage, and available supervision. Existing work shows favorable scaling with increasing human data, but it remains unclear which data properties drive downstream robot gains and how to use such data throughout the training pipeline. We present a systematic study of egocentric human data with different alignment and supervision under a unified world-action model framework. With the model backbone fixed, we disentangle the effects of human-robot alignment, data duration and task diversity, action supervision, and data usage strategies. We find that aligned human demonstrations substantially improve out-of-distribution generalization and reduce target-task robot data requirements; data duration and task diversity affect downstream capabilities differently; and video-only supervision remains effective without action labels, providing a strong foundation for subsequent video-action training. We validate these findings through closed-loop policy evaluation on both real robots and RoboDojo. Rather than treating data duration as the sole scaling axis, Ego4WAM shows how alignment, task diversity, available supervision, and usage strategy jointly shape the value of egocentric human data for robot learning.

Figures & tables

Figure 1: Overview of the egocentric data pyramid.

Figure 2: Illustration of the world-action model framework and unified human-robot action space.

Figure 3: Real-robot evaluation on Place into the Basket with human data under OOD settings.

Figure 4: Real-robot evaluation on Fold Cloth with human data under different robot-data budgets.

Figure 5: RoboDojo evaluation under data-duration and task-diversity scaling. We visualize a subset of tasks with non-zero success rate, while all aggregate results are computed over the full evaluation task set. We provide complete per-task results in the appendix.

Training Setup			Genera.		Precision		Long		Memory		Open		Average
pre-train	mid-train	post-train	Score	SR	Score	SR	Score	SR	Score	SR	Score	SR	Score	SR
-	-	base	5.87	3.83	5.56	1.25	16.97	8.25	2.75	1.67	0.78	0.75	6.39	3.15
-	$\checkmark$	base	10.73	7.00	8.45	3.00	25.18	16.25	6.38	5.33	0.68	0.50	10.28	6.42
$\checkmark$	-	base	12.04	7.67	23.53	16.25	27.85	17.50	6.67	5.33	0.55	0.50	14.13	9.45
$\checkmark$	$\checkmark$	base	12.63	8.33	24.77	18.00	29.46	20.50	7.42	6.33	0.90	0.75	15.04	10.78
$\checkmark$	-	joint	17.19	12.17	31.16	22.25	44.16	29.75	8.37	7.33	0.55	0.50	20.29	14.35

Table 1: Ablation study of egocentric data utilization strategies on RoboDojo. Base and joint denote different shared-attention mechanisms used during post-training.

Model	Data Scale		Genera.		Precision		Long		Memory		Open		Average
Model	Robot	Ego	Score	SR	Score	SR	Score	SR	Score	SR	Score	SR	Score	SR
Pi-0.5	-	-	13.38	8.17	12.40	5.50	23.54	14.67	5.89	4.67	1.98	1.67	11.44	6.93
Spatial Forcing	-	-	14.12	9.34	17.32	10.58	23.26	14.58	5.43	4.11	1.78	1.58	12.38	8.04
Hy-0.5-VLA	-	-	11.78	8.39	13.81	8.00	25.74	14.92	13.37	12.11	0.65	0.58	13.07	8.80
Meituan-0	-	-	13.75	8.17	16.77	7.75	29.61	18.58	10.06	8.89	4.54	4.25	14.95	9.53
Xiaomi-1	7.2K	100K	23.54	17.00	26.69	18.83	38.39	23.67	7.81	6.56	3.94	3.58	20.07	13.93

Table 2: Performance comparison on RoboDojo. The scale of robot and egocentric training data is reported in hours. Results of existing methods are taken from the official leaderboard. The style of bold and underline denotes the best and second-best results, respectively.

Figure 6: Real-robot evaluation across diverse manipulation tasks.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Video Pre-train	Full Mid-train	Scaling Mid-train	Real-Robot Post-train	RoboDojo Post-train
Training Steps	150K	80K	40K	10K	40K
Nodes	8	8	4	1	2
Initial LR	2e-5	1e-4	1e-4	1e-4	1e-4
Final LR	1e-5	1e-6	1e-6	1e-6	1e-6
Warmup Steps	2K	2K	2K	1K	2K
Shared Attention	Base	Base	Base	Base	Base / Joint

Appendix

Table 3: Training configurations used throughout our experiments. All experiments within the same configuration use identical optimization settings unless otherwise specified.

Figure 7: RoboDojo evaluation under data-duration and task-diversity scaling.

Explore similar work

Ego-Pi: VLA Fine-Tuning for Ego-Centric Human and Robot Data
Jun 6, 2026Ji Woong Kim, Ke Wang, Zipeng Fu +4Egocentric Dataset Humanoid
1Stanford University · 2Meta
Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data
Aug 3, 2026Ye Wang, Pei Lin, Xiong-Hui Chen +12Large-Scale Robot Demonstration Datasets Robotic Data Acquisition
1AIM3 Lab, Renmin University of China · 2Qwen Team, Alibaba Inc. · 3ShanghaiTech University +2
EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World
Apr 8, 2026Ryan Punamiya, Simar Kareer, Zeyi Liu +37Egocentric Dataset Scalable Robot Learning
Georgia Institute of Technology · Stanford University · University of California San Diego +5

Jun 6, 2026, cs.ROJ/K move · Enter open · S save

Ego-Pi: VLA Fine-Tuning for Ego-Centric Human and Robot Data

Ji Woong Kim, Ke Wang, Zipeng Fu, Sirui Chen+3

1Stanford University · 2Meta

Egocentric Dataset Humanoid

Open paper AskPDF arXiv

Robotics faces a fundamental challenge of data scarcity. Unlike language or vision research, there is no internet-scale dataset for robotic manipulation. A promising path forward is to leverage egocentric human data, which can be collected more easily, with greater breadth, and at a larger scale. Towards this end, we investigate key design choices for learning across human and humanoid embodiments equipped with dexterous five-finger hands, using the

π_{0.5}

model as a foundation. Our results show that human data enables robots to learn new task semantics and compose existing skills into novel behaviors without corresponding robot data. The paper website is here: https://egopipaper.github.io/

Abstract

Explore similar work

Ego-Pi: VLA Fine-Tuning for Ego-Centric Human and Robot Data

Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data

EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World