Microduck RL playground

Real trained RL policies for the Microduck robot, running fully in the browser: MuJoCo compiled to WebAssembly steps the physics, onnxruntime-web runs the policy network at 50 Hz. No server, no backend.

Two locomotion variants of the same robot are included: legs (walking, the default) and rollers (the wheeled skating variant). Press M (or hold D-pad up ~1 s on a gamepad) to switch; the roller model, meshes and policies are lazy-loaded on the first switch.

Policies

Policies and MJCF model from pollen-robotics/microduck and pollen-robotics/microduck_rl.

Controls

• Arrows or WASD / ZQSD: forward / back + turn
• M: switch legs <-> rollers
• Q / E (A / E on AZERTY): kick left / right (legs only)
• F: alternate left / right kicks (legs only)
• R: sit / stand (legs) / crouch-glide (rollers)
• G: ground pick (legs only)
• C: toggle the chase camera (on by default; dragging detaches it)
• Space: reset
• Drag to orbit, scroll to zoom
• Colour dots: repaint the duck (it quacks)

In roller mode the legs-only actions (kicks, sit) are disabled and their hints fade out; play the ball by driving into it.

URL parameters

• ?boot=1: skip the welcome modal and land straight on the BIOS console, which then shows the real loading progress live (honest loader) before the normal entrance plays. Boot failures (missing asset, policy fetch error...) freeze the console on a SYSTEM HALTED screen with the error detail - handy for debugging.

The ball's physics are local to your tab, but its pose is broadcast with your duck's, so other visitors see a translucent copy of your ball next to your ghost (and you see theirs). A square arena (3 x 3 m) fences the play area so neither the ball nor the duck can wander off.

Gamepad

Plug in a controller and the same mapping as the real robot runtime applies:

• Left stick: forward / back + turn
• Right stick: orbit the camera (detaches the chase cam)
• R3 (right stick click): toggle the chase cam back on
• X or D-pad down: sit / stand (legs) / crouch-glide (rollers)
• A: ground pick (legs only)
• Y: toggle head control
• RB / LB: right / left kick (legs only)
• D-pad up: hold ~1 s to switch legs <-> rollers (the real robot uses a 3 s hold)
• Right trigger: mouth (analog) + quack
• Left trigger: hold for wheee
• A: enter / resume the game while the title or pause overlay is up
• Menu/Start (or Select): exit back to the title overlay while in-game

Multiplayer ghosts

Other people visiting the Space at the same time show up as translucent ducks, live. Peer-to-peer WebRTC via Trystero (serverless signaling over public Nostr relays), so it works from a static Space with no backend. Each tab broadcasts its duck's pose (trunk + 14 joints + jaw + colour + locomotion variant) at 10 Hz; up to 3 ghosts are rendered, extra peers stay connected but invisible. The camcorder-style OSD in the top-right corner shows "N ONLINE" when peers are around.

Ghost limitations in v1: a peer in roller mode renders with the roller rig only if your tab has already loaded it (otherwise it falls back to the leg rig), and ghost wheels don't spin (passive wheel joints aren't broadcast). Old clients simply ignore the new variant flag.

How it works

The app is a Vite + React + MUI shell around an imperative game core: React/MUI renders the UI chrome (title menu, BIOS console, HUD, touch overlay), react-three-fiber owns the canvas/lights/environment, and the physics/policy/rig loop lives in framework-agnostic modules under app/src/game/. A zustand store bridges the two (game state out, UI intents in).

• app/src/game/game.js fetches the MJCF (robot_allcollisions.xml, or robot_allcollisions_rollers.xml for the roller variant), strips the visual geoms, injects a floor, arena walls, a ball, a collision box for the arcade cabinet row and a STAND keyframe, and compiles it with the official @mujoco/mujoco WASM bindings.
• Both variants share the exact same policy interface: 61D observation (gyro, projected gravity, 14 joint pos/vel, last action, 13D command) and 14 position-targets, matching microduck_rl/scripts/infer_policy.py. The roller variant adds 4 passive wheel hinges that appear in qpos (zeroed in the keyframe) but not in the observation.
• Rendering is a three.js rig built from kinematics.json / kinematics_rollers.json + decimated STL meshes, driven directly from MuJoCo qpos (including the passive wheel spin).
• Switching variants swaps the compiled model + data + rig + ONNX sessions wholesale; both stay resident after the first load so toggling back and forth is instant.

Development

Large binary assets (robot meshes, GLB models, ONNX policies) are stored with Git LFS. Clone with LFS installed, or pull the real files after a plain clone - otherwise the app boots against text pointer files and fails with errors like SyntaxError: Unexpected token 'v', "version ht"... is not valid JSON (that string is the start of an LFS pointer, not your model):

git lfs install   # once per machine
git lfs pull      # fetch the actual binaries in an existing clone

cd app
npm install
npm run dev     # dev server on http://localhost:5173
npm run build   # production bundle in app/dist/

The Space builds with the included Dockerfile (Vite build, served by nginx-unprivileged on port 8080). Static assets (meshes, policies, audio, images) live in app/public/ and keep their historical URLs.