---
language:
- en
- ko
tags:
- robotics
- reinforcement-learning
- mujoco
- onnx
- humanoid
- bipedal
license: mit
pipeline_tag: robotics
---

# 🦆 MicroDuck: Bipedal Robot Reinforcement Learning Policies & 3D Viewer

[![Hugging Face Space](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-3D%20Web%20Viewer-yellow)](https://huggingface.co/spaces/kyoungsim/microduck-viewer)

This repository contains pre-trained Deep Reinforcement Learning (RL) policy models, MuJoCo simulation assets, and the interactive 3D Web Viewer for the **MicroDuck** small bipedal robot.

### 🌐 Live Interactive 3D Demo
👉 **[Open MicroDuck 3D Web Viewer](https://huggingface.co/spaces/kyoungsim/microduck-viewer)**  
*(Play with the robot directly in your browser: forward walking, duck waddle, and physical soccer ball kicking!)*

---

## 📦 Repository Contents

- `index.html`: Standalone Interactive 3D Web Viewer & Soccer Simulator (Three.js)
- `policies/`: Pre-trained ONNX policy checkpoints
  - `alpha_walking.onnx`: Dynamic omnidirectional bipedal locomotion
  - `alpha_stand.onnx`: Stable posture hold and disturbance rejection
  - `alpha_sitstand.onnx`: Continuous sit and stand-up transition
  - `alpha_ground_pick.onnx`: Ground object picking motion
  - `ball_kick_left.onnx`: Precision left foot ball kick
  - `roulade.onnx`: Dynamic forward roll (roulade) motion recovery
- `robot/`: MuJoCo MJCF XML definitions and 3D mesh assets (`scene.xml`, `scene_ball.xml`, `robot_allcollisions.xml`, `assets/*.stl`)
- `infer_engine.py`: Standalone policy inference engine
- `app.py`: Offline simulation renderer

---

## 🎮 Features
- **3D Robot Anatomy**: Realistic duck bill, neck linkage, and webbed duck feet
- **Soccer Ball Mechanics**: Physical contact-based ball kicking & dribbling
- **Real-time Control**: Keyboard WASD/Arrow driving & motion toggling
- **Reinforcement Learning**: 50 Hz control frequency trained with MuJoCo

## 🚀 Quick Start
```bash
pip install mujoco onnxruntime numpy imageio
```