EgoHumanoid-V2: Human-to-Humanoid Transfer of Coordinated Whole-Body Skills for Loco-Manipulation

The first egocentric human-to-humanoid skill transfer framework for coordinated whole-body loco-manipulation

Full Paper Code

Explore the research

Highlights

The first egocentric human-to-humanoid skill transfer framework for coordinated whole-body loco-manipulation.
Coarse-to-fine action alignment combines kinematic reference correction with dynamics-aware refinement.
Zero-shot skill transfer on four real-world tasks, with analyses of accuracy, policy scores, and collection cost.

EgoHumanoid-V2 overview: human demonstrations undergo visual and action alignment, then supervise a VLA policy for whole-body robot control.

Results

Human Data Collection

Unaligned Data · Replay

Aligned Data · Inference

Scaling Human Demonstrations

Close Laptop

Open Fridge: task scores at 25, 50, 100 and 200 human demonstrations. — Open Fridge

Pull Curtain: task scores at 25, 50, 100 and 200 human demonstrations. — Pull Curtain

Close Drawer

Task scores by demonstration count

Human demos / task	Close Laptop	Open Refrigerator	Pull Curtain	Close Drawer
25	35.00	21.67	40.00	56.67
50	31.67	33.33	38.33	36.67
100	48.33	28.33	50.00	53.33
200	53.33	38.33	53.33	61.67

Action Alignment

H: Initial SONIC rolloutK: Kinematic reference correctionD: Dynamics-aware refinement

Human motion is initialized through SONIC, corrected by offline inverse kinematics, and refined using an MLP response model and iterative execution feedback.

Mean score ↑

21.25%

H + K

47.08%

H + K + D_s

51.67%

Endpoint error ↓

Stage	Position (cm)	Orientation (°)
H	11.21	21.22
H + K	4.45	13.77
H + K + D_s	3.41	6.71

Task scores

Action alignment	Close Laptop	Open Refrigerator	Pull Curtain	Close Drawer	Mean
H	0.00	0.00	28.33	56.67	21.25
H + K	46.67	31.67	51.67	58.33	47.08
H + K + D_s	53.33	38.33	53.33	61.67	51.67

Visual Alignment

Close Laptop

Open Fridge

Pull Curtain

Visual alignment results

Training images	Close Laptop	Open Refrigerator	Pull Curtain	Close Drawer
Human RGB	31.67	16.67	35.00	50.00
+ Arm rendering	51.67	23.33	41.67	—
+ Augmentation	61.67	33.33	33.33	61.67
+ Both	53.33	38.33	53.33	—

Task scores with 200 human demonstrations. Arm rendering is not evaluated for foot-operated drawer closing.

BibTeX Citation

@article{chen2026egohumanoidv2,
    title={{EgoHumanoid-V2}: Human-to-Humanoid Transfer of Coordinated Whole-Body Skills for Loco-Manipulation},
    author={Chen, Jin and Jiang, Yiming and Xu, Chongyang and Shi, Modi and Peng, Shijia and Chen, Li and Li, Tianyu and Xu, Mu and Chen, Yilun and Hoi, Steven and Li, Hongyang},
    journal={arXiv preprint arXiv:2609.37181},
    year={2026}
}

Contact

Jin Chen [email protected]