Files
VibemakerSandbox/README.md
T
tfrere a8ad7ec412 feat: live multiplayer ghosts over WebRTC (Trystero/Nostr)
Every open tab broadcasts its duck pose (trunk + joints + jaw + colour)
at 10 Hz over WebRTC data channels with serverless Nostr signaling, and
renders other visitors as translucent ducks (capped at 3). Peers online
count shows in the telemetry line.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-20 12:11:30 +02:00

2.2 KiB

title, emoji, colorFrom, colorTo, sdk, pinned
title emoji colorFrom colorTo sdk pinned
Microduck Sandbox 🐤 yellow gray static false

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.

Policies

Button Checkpoint What it does
Run BEST_alpha_walking.onnx Velocity-tracking locomotion (arrows / WASD to steer)
Sit BEST_alpha_sitstand.onnx Sits down on its hull, stands back up
Roll roulade.onnx Rolls over and recovers

Policies and MJCF model from apirrone/microduck_runtime and apirrone/mjlab_microduck.

Controls

  • Arrow up / down: forward / back
  • Arrow left / right: turn
  • A / E: strafe
  • Space: roll (one barrel roll, then back to running)
  • R: reset
  • Drag to orbit, scroll to zoom
  • Colour dots: repaint the duck (it quacks)

Gamepad

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

  • Left stick: forward / back + strafe
  • Right stick: turn
  • X: roll
  • D-pad down: sit / stand toggle
  • Right trigger: mouth (analog) + quack

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) at 10 Hz; up to 3 ghosts are rendered, extra peers stay connected but invisible. The header shows "N online" when peers are around.

How it works

  • rl.js fetches the MJCF (robot_allcollisions.xml), strips the visual geoms, injects a floor and a STAND keyframe, and compiles it with the official @mujoco/mujoco WASM bindings.
  • The 61D observation (gyro, projected gravity, joint pos/vel, last action, command) matches mjlab_microduck/scripts/infer_policy.py.
  • Rendering is a three.js rig built from kinematics.json + decimated STL meshes, driven directly from MuJoCo qpos.