Microduck Sandbox

Microduck Sandbox

Lets you steer a simulated Microduck robot in the browser with trained walking, sitting, rolling and kicking policies, plus a wheeled variant. For robotics and reinforcement learning fans.

Made by pollen-robotics on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Microduck Sandbox is a robot playground that runs in your browser. It runs trained reinforcement learning policies for the Microduck robot. MuJoCo compiled to WebAssembly handles the physics, and onnxruntime-web runs the policy network 50 times a second. No server or backend is involved.

Who it suits: It suits people curious about robot control and reinforcement learning who would rather try trained policies than read about them.

Industry | Who uses it | How they use it | Example
Robotics | Hobbyists | Drive a simulated robot with a keyboard or gamepad | Press M to switch from legs to rollers
Research | RL researchers | See trained policies running in a physics simulation | Watch the sit and roll policies work
Education | Students | Learn how a policy network drives a robot body | Read the game code under app/src/game
Web development | Front-end developers | Study physics and a 3D scene in the browser | Look at how React, three.js and MuJoCo fit together

Skip it if you need to train new policies or test on a real robot, since this page only plays the ones it ships.

Where to use it:

• A web browser, on the Space's page at Hugging Face, with keyboard, mouse or gamepad
• Docker on your own Windows, macOS or Linux computer

Use case | Who | What they get back
Play with the robot | Hobbyists | A walking or skating duck you steer yourself
Watch policies at work | Researchers | Sit, roll, kick and crouch behaviours in a live simulation
Meet other visitors | Everyone | Translucent ghost ducks of people on the page at the same time

How to set it up: The online page needs nothing. A local copy needs Docker, and Python with pip to fetch the files.

• Docker
• Python and pip, to fetch the files
• Windows, macOS or Linux

1. Install the Hugging Face hub library.
2. Download the Space's files into a folder named microduck-simulator.
3. Move into that folder.
4. Build the Docker image.
5. Run the container with port 8080 published, then open that port in your browser.

pip install -U huggingface_hub
hf download pollen-robotics/microduck-simulator --repo-type space --local-dir microduck-simulator
cd microduck-simulator
docker build -t microduck-simulator .
docker run -it -p 8080:8080 microduck-simulator

How to use it: Open the page and steer the duck. Walking on legs is the default. A second variant on wheels loads the first time you switch to it.

docker run -it -p 8080:8080 microduck-simulator

Run that line for a local copy, then browse to port 8080 on your machine. Each mode is a separate trained policy.

Tool | What it does
Run (legs) | Walking policy that follows your steering, the default mode
Sit | Sits down on its hull and stands back up
Roll | Rolls over and recovers
Kick | A one-shot kick with the left or right leg, over a window of half a second
Drive (rollers) | Skating on four passive wheels, with a top speed of 0.6 m/s
Crouch (rollers) | A one-shot glide that sinks low over about 3.5 seconds, then stands back up

Keyboard and mouse controls:

• Arrow keys or WASD: move forward and back, and turn
• M: switch between legs and rollers
• Q and E: kick left or right, on legs only
• F: alternate left and right kicks, on legs only
• R: sit or stand on legs, crouch-glide on rollers
• G: ground pick, on legs only
• C: toggle the chase camera, which starts on
• Space: reset
• Drag to orbit and scroll to zoom
• Colour dots: repaint the duck, and it quacks

Example: You open the page and press the arrow keys to walk the duck around the square arena. You press Q to kick the ball with the left leg, then M to switch to the wheeled variant and drive into it.

For newcomers and experts: Open the page, walk with the arrow keys and press R to make the duck sit. That shows two different trained behaviours in under a minute.

The app is a Vite and React shell with MUI around a game core written without a framework. React draws the menus and overlays, react-three-fiber owns the canvas, and a zustand store joins the two. The core loads the robot model, adds a floor, arena walls and a ball, and compiles it with the MuJoCo WebAssembly bindings. A policy then runs through onnxruntime-web on every control step.

Load the robot model and policies
MuJoCo steps the physics in WebAssembly
The policy network runs at 50 Hz
React and three.js draw the scene
Your keys or gamepad set the command

Good to know:

• Arena: A square play area of 3 by 3 meters keeps the ball and the duck from wandering off.
• Ghosts: Up to 3 other visitors show as see-through ducks. Wheels on a ghost do not spin, and a ghost in roller mode needs your tab to have loaded the roller rig.
• License: apache-2, with a LICENSE file in the repository.
• Cost to run: none named.
• Data it reaches: Your duck's pose, colour and ball position go to other visitors, peer to peer over WebRTC with signaling on public Nostr relays.
• Not for: Training new policies or driving the real robot.

On Hugging Face, 13 November 2026

• Likes: 804
• Built with: docker
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/pollen-robotics/microduck-simulator
• Files: https://huggingface.co/spaces/pollen-robotics/microduck-simulator/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.