EGOHUMANOID | Human-to-Humanoid Transfer for Loco-Manipulation

HIGHLIGHT

Research Summary

Human demonstrations offer rich environmental diversity and scale naturally, making them an appealing alternative to robot teleoperation. We present EGOHUMANOID, the first framework to co-train a vision-language-action policy using abundant egocentric human demonstrations together with a limited amount of robot data.

To bridge the embodiment gap, we introduce a systematic alignment pipeline with two key components: view alignment reduces visual discrepancies; action alignment maps human motions into a unified action space for humanoid control.

Extensive real-world experiments demonstrate that incorporating robot-free egocentric data significantly outperforms robot-only baselines by 51%, particularly in unseen environments.

Task Showcase

Real World Loco-Manipulation Tasks

Spanning varying levels of difficulty for large-space movement and dexterous manipulation

Robot Data Collection

Teleoperated demonstrations in laboratory environments

Laboratory

Pillow Placement

Laboratory

Trash Disposal

Laboratory

Toy Transfer

Laboratory

Cart Stowing

Human Data Collection

Egocentric demonstrations in diverse real-world environments

In-the-Wild

Pillow Placement

In-the-Wild

Trash Disposal

In-the-Wild

Toy Transfer

In-the-Wild

Cart Stowing

Loco-Manipulation Deployment

Real-world deployment comparing different training configurations

Robot-only

0% Gen.

Human-only

95% Gen.

Co-training
97.5% Gen.

Robot-only

55% Gen.

Human-only

80% Gen.

Co-training
87.5% Gen.

Robot-only

37.5% Gen.

Human-only

35% Gen.

Co-training
78.8% Gen.

Robot-only

31.3% Gen.

Human-only

0% Gen.

Co-training
65% Gen.

Results

Performance Comparison

Co-training consistently improves performance across both in-domain and generalization settings

In-Domain Performance

Laboratory environments

Robot-only

Co-training

Generalization Performance

Diverse real-world scenes (+51% improvement)

Robot-only

Co-training

Transfer Analysis

Subtask Performance Breakdown

Navigation transfers effectively; manipulation depends on precision requirements

Data Config	Pillow Placement		Trash Disposal		Toy Transfer				Cart Stowing
	s1	s2	s1	s2	s1	s2	s3	s4	s1	s2	s3	s4
Robot-only	0	0	65	45	100	50	0	0	100	15	5	5
Human-only	100	95	100	80	100	100	45	35	100	5	0	0
Co-training	100	95	100	75	100	100	60	55	100	60	50	50

Locomotion

Manipulation

Failure Analysis

Understanding Failure Modes

Systematic breakdown of failure cases across different training configurations

Robot-only

Human-only

Co-training

Scaling Analysis

Human Data Scaling Behavior

Performance consistently improves as human demonstrations accumulate

Data sampling ratio:

Robot:Human 1:2

Robot:Human 1:1

Robot:Human 2:1

Average

Pillow Placement

Trash Disposal

Toy Transfer

Cart Stowing

Method

Human-to-Humanoid Alignment Pipeline

A systematic approach to bridge the embodiment gap between humans and robots

Human Demonstrations

Diverse Environments

Robot Demonstrations

Laboratory Environments

(a) View Alignment

Depth Estimation

Inpainting

Point Cloud Transform

🖱 Drag to rotate

Point Cloud Reprojection

(b) Action Alignment

✨ Unified Action Space

💪

Upper Body

→

6-DoF Delta End-Effector

X

Y

Z

Rx

Ry

Rz

🦵

Lower Body

→

Discrete Velocity Commands

⬆️

⬇️

⬅️

➡️

↩️

↪️

🖐️

Dexterous Hand

→

Binary Open/Close

✋

Open

/

✊

Close

BibTeX Citation

@article{shi2026egohumanoid,
    title={EgoHumanoid: Unlocking In-the-Wild Loco-Manipulation with Robot-Free Egocentric Demonstration},
    author={Shi, Modi and Peng, Shijia and Chen, Jin and Jiang, Haoran and Li, Yinghui and Huang, Di and Luo, Ping and Li, Hongyang and Chen, Li},
    journal={arXiv preprint arXiv:2602.10106},
    year={2026}
}

EgoHumanoid: Unlocking In-the-Wild Loco-Manipulation with Robot-Free Egocentric Demonstration

Research Summary

First Human-to-Humanoid Transfer

Principled Alignment Pipeline

Comprehensive Evaluation

Real World Loco-Manipulation Tasks

Robot Data Collection

Human Data Collection

Loco-Manipulation Deployment

Performance Comparison

Subtask Performance Breakdown

Understanding Failure Modes

Human Data Scaling Behavior

Human-to-Humanoid Alignment Pipeline

Human Demonstrations

Robot Demonstrations

(a) View Alignment

(b) Action Alignment