Your screen is the stadium.
Their phones are the gamepads.

Duck Party, Ball Frenzy and Pond King seat 24 on a pitch 1.5× bigger: NPCs hold every seat until someone scans in and hold a formation, and the big screen tags only the people. Lost in the crowd? FIND ME on your phone makes your duck power up in a random color for 3 seconds (once every 20 s). Pick Duck Cup for two semi-finals and a final between four NPC clubs, set the NPCs to easy, normal or hard, and watch the JEV coach (TypeSafe’s System One model) call each team’s plan every 12 seconds, with built-in rules whenever it is off. Phones only send a stick, KICK and QUACK. The big screen turns your stick into a walking command from your duck’s live heading and the trained walker does the moving. A trained kick fires when the ball is at your feet, and a stand-up policy recovers falls. When nobody is waiting, everybody plays on. When people join the line, the longest-playing duck hands over its seat with at least 15 seconds’ warning on their phone.

The browser check proves the interface and a finite output, not gait quality or hardware readiness. Arena matches are unranked browser simulation.

Your duck is waiting.

Start the playground, then start a run inside the scene.
Click inside to use its keyboard and on-screen controls.

Start playground

Checking verified assets…

ControllerWaiting to start

Model—

Control rate—

Policy interface—

Live readings appear after the simulator starts.

Keyboard control works while the game is focused. Policy buttons make bounded requests through the game’s own controls. If a brain or recovery state refuses a move, the reason appears here. Policy file selection stays inside the scene. Account, multiplayer and training features open the full app.

Every chore runs the real MicroDuck model in browser MuJoCo. A referee reads the simulated physics each control step and stops a simulated clock, so pausing freezes it and a slow device is not penalized. Two chores are playable now and six more are reviewed design records. Each card names the MJLab task it maps to and where that task honestly stands.

Coming soon · design records

Chores use the legs model and the shared policy family: 61 inputs, 14 servo actions, 50 Hz. A browser record is simulated evidence of a human-steered run. It is not a trained policy, an evaluation, owner approval or physical validation.

There is no global server leaderboard. Each playable chore keeps its ten best ranked runs in this browser’s storage. Every row names the brain, a short digest of the walker file and its lifecycle label, so a record always says what was driving.

After a clean, unassisted run you can opt in to save it as a local human demonstration: a teaching example for MJLab training. Nothing becomes a demonstration unless you choose it, and playing in this Space never starts training or a paid job.

1. 01

   Play clean

   An unassisted run lands on your device board.

2. 02

   Opt in

   Save that run as a local human demonstration.

3. 03

   Curate

   Build a reviewed dataset in Trainer HQ ↗.

4. 04

   Train in MJLab

   Native PPO on the chore’s task. Sock Sweep’s 3-seed pilot trained but shoves the sock, so it failed its gates and is not promoted. Paid GPU runs only after explicit cost approval.

5. 05

   Export ONNX

   scripts/export.py bakes the observation normalizer into the file.

6. 06

   Import a brain

   Equip it in Policies + keys, or in Trainer HQ for your account.

7. 07

   Beat your record

   Same chore, same referee. The board labels the walker you ran.

Hosted GPU training is currently disabled on the live app. Native training and browser rehearsal are available.

These are captures from actual two-player simulations: two independent browsers, two customized ducks and a shared room. Each player drives their own robot. Pick a setting, invite friends and take on its room quest, a Beak Dash race or Twin Beacon co-op.

Recorded local multiplayer simulations with test profiles. Enter a live room in the full app to play with other people.

Play as MicroDuck alongside other visitors. Join a Beak Dash heat, hold the Twin Beacon pads with a partner, or explore room quests. Make an unlisted room for your group, choose its layout and share its invite link from the full app.

Pick a starter loadout in Policies + keys, or build your own. A walking brain goes in one slot; supplied kicks, rolls, sitting and quacks go in the others. Equip keys, then start a run. Click an empty slot to add a move, right-click to swap it, drag to rearrange, and flip between three loadout pages. Swap the walking slot for your own compatible ONNX whenever you want.

WalkFind your way

KickPick your foot

BalanceLift and hold

RollTuck and tumble

GlideSwitch to wheels

Your policyBring your own walk

EXPLORE / FIND YOUR WAY

Scout

Trail walking, ground pick and a signal for your flock.

• 1 Trail brain
• 2 Ground pick
• 3 Sit / stand
• 4 Quack
• 5 Walk again

LINE UP / PICK YOUR FOOT

Striker

Build around balloon football and your next well-timed kick.

• 1 Striker brain
• 2 Kick left
• 3 Kick right
• 4 Quack
• 5 Walk again

ROLL / RECOVER / CELEBRATE

Trickster

A forward roll, a seated breather and a quack for the crowd.

• 1 Factory brain
• 2 Forward roll
• 3 Sit / stand
• 4 Quack
• 5 Walk again

Starter sets update keys 1–5 and keep 6–9. Freely reassign every slot. Matching models and the room’s move rules still apply. Uploaded networks replace walking; skill slots use the supplied controllers.

These networks turn robot observations into motor actions. They have no camera input. Object alignment and decisions about what to try come from you and the simulator’s task logic.

Open Policies + keys to explore the Community hub: pinned MicroDuck policies published on Hugging Face, with clear labels for ready moves and those that need a different setup. Add a ready move and give it a key, or build a combo that chains 2–6 moves on one shortcut. Compatible object-push exports become private skills on this device; walking exports become brains.

MORE · TAKE THE NEXT STEP

Try a policy. Understand it. Build on it.

Browser practice, training, evaluation and physical validation are separate steps. This Space demonstrates simulation and carries each policy’s evidence; it does not establish hardware readiness.