wyhuai / HumanX

HumanX: Toward Agile and Generalizable Humanoid Interaction Skills from Human Videos 
**CoRL 2026**
English | 简体中文
---
# What You Can Reproduce
The released code and example data support reproducing the following
humanoid-object interaction skills in simulation:
- 🏀 Basketball catch-and-throw
- 🏀 Basketball pump-fake shot
- ⚽️ Football kicking
- 📦 Bin pick-and-place
- 💡 New interaction skills learned from your own videos
# Overview
HumanX is a full-stack framework for turning human videos into agile and
generalizable humanoid interaction skills.
The pipeline consists of two complementary components:
- **XGen**: a motion and HOI editor that turns compatible robot motions into
 physically plausible humanoid-object interaction motions.
- **XMimic**: a unified imitation learning framework for training and
 evaluating interaction skills in simulation.
The interface between the two components is motion data. XGen edits and
exports robot and HOI `.npz` files, while XMimic converts compatible XGen
outputs into the HumanX `.pkl` format used for training and evaluation.
For new skills, the recommended video-to-policy pipeline is:
1. **GVHMR** recovers world-grounded human
 motion from monocular video.
2. **UMR** retargets the recovered motion to
 the G1 humanoid.
3. **XGen** edits and synthesizes physically plausible
 humanoid-object interaction motions, then exports HOI `.npz` files.
4. **XMimic** converts the XGen HOI `.npz` files into HumanX `.pkl` files and
 trains the corresponding interaction policy.
## Repository Structure
 [code] 
## Documentation
Use the README for the component you want to run:
- XGen README: XGen editor, motion formats,
 HOI synthesis, visualization, and conversion tools.
- XMimic README: installation, training, evaluation,
 XGen-to-XMimic conversion, and motion playback.
## End-to-End Workflow
1. Use XGen to inspect robot motion, synthesize
 HOI data, and export motion `.npz` files.
2. Follow the XMimic data pipeline
 to convert XGen HOI data into HumanX `.pkl` motion data.
3. Use the XMimic training and evaluation instructions
 to train, evaluate, and visualize the resulting policy.
For the standard interface, keep source XGen data and generated XMimic motion
files under the corresponding `data/` directories.
# Citation
If you find this repository useful, please cite:
 [code] 
# Acknowledgements
This repository builds upon prior open-source efforts, including:
- ASAP
- HumanoidVerse
- BeyondMimic
- SkillMimic
We thank these communities for their excellent work.
# License
This project is licensed under the Apache-2.0 WITH Commons-Clause