Collect
Record human demonstrations with dual egocentric views and dynamics-guided feedback.
Learn from human demonstrations alone.
Walk, carry, and interact over long distances.
EgoAlign
Bridging the Human–Humanoid Gap
for Long-Range Loco-Manipulation
1 Beihang University 2 Shanghai Innovation Institute 3 Sichuan University
4 Alibaba Group
* Work done during an internship at Alibaba Token Hub (ATH), Alibaba Group.
01 / The gap
Beyond stationary tasks.
Learn to move and interact.
Long-range loco-manipulation connects walking, reaching, and interacting across a scene. People demonstrate these skills naturally, but differences in body proportions and controller response make their motions difficult for a humanoid to realize.
EgoAlign bridges this gap by turning egocentric human demonstrations into robot-compatible action and state supervision, enabling task learning without physical-robot demonstrations.
02 / The method
From human demonstrations
to robot-compatible data.
We capture egocentric views, body motion, and hand commands, then adapt the motion and reconstruct corresponding robot states. Using only these human task demonstrations, we fine-tune a VLA for long-range, closed-loop humanoid execution.
Adapt
Combine kinematic scale alignment with controller-in-the-loop refinement.
Reconstruct
Pair pre-action robot states with same-tick motion tokens from the final rollout.
03 / Demonstration collection
Capture human skill.
See the robot’s response.
A PICO headset and five trackers capture body motion, while two GoPro cameras record downward and level views. Corresponding human and robot cameras have matched ground-relative heights.
Feedback while collecting
Real-time SONIC–MuJoCo feedback lets demonstrators see the robot’s realized motion and adjust their movements accordingly.
More demonstrations, less on-site effort
The collection example below shows six human demonstrations in the time used to collect one teleoperated demonstration.
04 / Scale alignment
Align the motion.
Account for the controller.
EgoAlign preserves global travel references while adapting upper-body interaction geometry. Kinematic alignment and controller-in-the-loop refinement correct hand-position errors; during deployment, visual feedback guides the robot’s walking toward its goal.
Final causal replay constructs robot-state and action supervision for the adapted references. At deployment, the VLA predicts motion tokens, while SONIC decodes them using live robot state history.
05 / Physical deployment
Human-only task training.
Long-range robot behavior.
With no physical-robot task demonstrations, the fine-tuned VLA deploys zero-shot on a Unitree G1. Through SONIC’s continuous whole-body interface, it performs long-range object relocation, position-generalized navigation, and foot interaction.
Navigation & Foot-Based Interaction
The robot approaches a pedal bin and opens it with its foot. Further demonstrations show navigation to different bin locations and responses to a moving target.
Long-Range Object Relocation
Across an approximately 10-m task, the robot approaches a basket, picks it up, transports it, and places it on a table—all learned from human demonstrations. The videos also show heading adjustments toward a relocated table.
Bridging the human–humanoid gap
Train on human demonstrations.
Go the distance on a humanoid.
By aligning egocentric data with the target robot, EgoAlign enables human-only task training for zero-shot, long-range loco-manipulation.
Reference
Citation
@misc{jiang2026egoalign,
title = {{EgoAlign}: Bridging the Human-Humanoid Gap for Long-Range Loco-Manipulation},
author = {Yiming Jiang and Jin Chen and Chongyang Xu and
Yilun Chen and Aimin Hao and Yisheng He},
year = {2026},
eprint = {2609.38046},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.38046}
}