Skip to content
Getting startedDesktop entry, movement and reset checked September 8; public official reference load and unload checked October 12; other behaviors source reviewed

MicroDuck Simulator: First Walk and Public Move Loading

Open the official simulator and choose Waddle In to try the built-in walk. A ?move=org/repo URL can load a public move; the Playing label identifies what is loaded, not whether it is driving. Its X button restores the official slot. We checked the public sprint reference load and unload on October 12.

Your first browser task

  1. Open the official simulator; wait for the scene and choose Waddle In (or Enter).
  2. Click inside the scene if needed, then hold ↑ briefly to move forward.
  3. Release it and try ← / → to turn.
  4. Press Space to reset, then choose whether to explore more moves or learn the training route.

Done when you have entered the scene, moved, turned and reset. This task uses existing policies.

Open the official simulator ↗
What was checked: On September 8, 2026, we entered the official desktop simulator, moved forward and reset. On October 12, we opened the public official sprint reference link below, chose Waddle In, saw PLAYING / SPRINT-2M-REFERENCE BY POLLEN-ROBOTICS, and used its ✕ button to remove the label. We did not measure its gait or speed, test other models, or run a physical robot. Public move loading was implemented in September; the simulator’s October 9 latest commit added a license, not this feature. September 7 change ↗ October 9 commit ↗.

First, walk with the built-in move.

  1. Open the official simulator ↗ in a desktop browser and wait for the scene and policy assets.
  2. Choose Waddle In or press Enter. Click the scene if the keyboard has not taken focus.
  3. In Feet mode, briefly hold ↑ to walk, release it, then try ← or → to turn. These are turns, not sideways steps.
  4. Press Space to reset after a fall or when comparing another attempt.
Keyboard controls for this browser simulator
KeyEffect
↑ / ↓ or W / SForward / backward command while held.
← / → or A / DTurn left / right.
RIn Feet mode: built-in sit / stand, or the loaded episodic trick, script or custom sitstand action. In Rollers mode, R takes the crouch-glide path first, even with a custom move loaded.
Q / E; F; GIn Feet mode: left / right kick; alternating kicks; ground pick.
MSwitch Feet / Rollers. The first Roller switch may download extra assets.
C; drag; scrollChase camera; orbit; zoom.
SpaceReset the simulated robot.

The source reads physical key positions, so printed letters can differ on another keyboard layout. Arrows are the simplest first check. The local Python viewer has its own input mapping; follow this browser map only inside the Space. Keyboard code ↗ Python viewer ↗.

Open this official sprint reference in the simulator ↗ and choose Waddle In. The address uses ?move=pollen-robotics%2Fmicroduck-sprint-2m-reference; the encoded slash is the Hub repository separator. The interface may show LOADING MOVE first, then a PLAYING chip naming the loaded move. PLAYING is a load state, not proof that the move is currently driving: a perpetual walk drives in Feet walking, while Feet-mode episodic, script and sitstand moves need R. In Rollers mode, R takes the crouch-glide path first. This public official example needed no account in our October 12 load/unload check. The policy’s fixed manifest ↗ identifies a MicroDuck perpetual walk with 61 observations and 14 actions. Its name is not proof of a measured 2 m/s browser result.

For another public single-policy repository, replace the value after ?move= with its owner/repo ID. The startup parser ↗ also accepts ?policy= as an alias. The loader ↗ accepts a direct browser-reachable https://…/policy.onnx URL and probes a sibling manifest.json when available. Access restrictions or browser cross-origin rules can still prevent retrieval. These are URL imports; the reviewed UI has no local file-upload control. Private Academy session links are a separate authenticated path, not a free public move link.

What a manifest can mount.

Supported move kinds in the reviewed simulator
Manifest kindBrowser behavior
perpetual + walk (default)Replaces the walking policy in Feet mode while active.
perpetual + sitstandReplaces the Feet-mode sit/stand move, triggered with R.
episodicMounts as a one-shot trick on R in Feet mode, rather than replacing every built-in action.
scriptIn Feet mode, R toggles a timed 13-command sequence over the built-in walker; it is not an ONNX policy.

The browser accepts manifest model_api up to 2, expects robot.model to be MicroDuck when supplied, and rejects a policies array because it cannot load a whole multi-policy set at once. A missing manifest falls back to a single perpetual walk. For ONNX moves, the manifest must not claim a non-61 observation length or non-14 action length; the loaded model itself must have one 61-input and one 14-output tensor and produce finite, nonconstant smoke-test actions. Those checks catch some incompatible files, not every difference in observation meaning, accessories, action scaling or behavior. Match the policy’s mode and hardware assumptions as well; do not treat a Rollers model as a Feet walk merely because its tensor shape matches. Validation source ↗.

Unload or recover the official move.

When the PLAYING chip appears, choose its ✕ button to explicitly unload before trying another move. This restores the official policy in the occupied slot; our public reference check confirmed the chip disappeared. On a fresh public link, a failed initial download, rejected manifest or ONNX check shows an error while the built-in policy remains available. If an already loaded custom move is followed by a failed second load, the code does not automatically clear the old move: unload with ✕ before a replacement attempt, or return to the clean simulator URL to start again. The source also reverts a driving custom move if its output becomes nonfinite or repeated abnormal falls/explosions trigger its watchdog. These are browser recovery paths, not safety claims about a physical machine. HUD ↗ Slot and recovery code ↗.

  • Keys do nothing: close the welcome overlay, click the scene, try an arrow, then open the direct Space if an embed has focus problems.
  • The move will not load: check that the public repository contains policy.onnx; read the error for HTTP, manifest or tensor shape failures. Try the official reference link first to isolate your own repository.
  • Base scene does not finish loading: open the official ?boot=1 diagnostic view ↗ and record its last successful step. README ↗.

The browser temporarily plays an existing model; it does not train one or install it onto a robot. A local Python replay and a physical robot’s policy management are separate workflows. If you are building a model, continue with training routes; if you have hardware, use the hardware policy guide for its version and recovery checks.

CONTINUE THE CONVERSATION

What would you like to understand next?

Keep the source, your environment and your observations together.

Before you continue

Does loading a move in the browser train a new policy?

No. The browser runs an existing compatible movement. Training a new policy requires a separate training environment.

Applies to: Pollen Robotics MicroDuck. Keep the model, image, software and driver versions together. The dated sections below retain their original verification limits.

Source checked:

Sources and review (12)
  1. huggingface.co/spaces/pollen-robotics/microduck-simulator
  2. huggingface.co/spaces/pollen-robotics/microduck-simulator/blob/ad9d41a763f30dc770817c33d92dcfec0205db1e/README.md
  3. huggingface.co/spaces/pollen-robotics/microduck-simulator/blob/ad9d41a763f30dc770817c33d92dcfec0205db1e/app/src/App.jsx
  4. huggingface.co/spaces/pollen-robotics/microduck-simulator/blob/ad9d41a763f30dc770817c33d92dcfec0205db1e/app/src/game/custom-policy.js
  5. huggingface.co/spaces/pollen-robotics/microduck-simulator/blob/ad9d41a763f30dc770817c33d92dcfec0205db1e/app/src/game/game.js
  6. huggingface.co/spaces/pollen-robotics/microduck-simulator/blob/ad9d41a763f30dc770817c33d92dcfec0205db1e/app/src/ui/Hud.jsx
  7. huggingface.co/spaces/pollen-robotics/microduck-simulator/blob/ad9d41a763f30dc770817c33d92dcfec0205db1e/app/src/game/controls/keyboard.js
  8. huggingface.co/pollen-robotics/microduck-sprint-2m-reference/blob/62afcc461f1d4a5ac22bfbff55ce4900c1afba32/manifest.json
  9. huggingface.co/spaces/pollen-robotics/microduck-simulator/commit/cb0312619f09138069130328c3bef05736e1730c
  10. huggingface.co/spaces/pollen-robotics/microduck-simulator/commit/850e9430c7d4451ca35a69f8600cca74ce51814f
  11. huggingface.co/spaces/pollen-robotics/microduck-simulator/commit/ad9d41a763f30dc770817c33d92dcfec0205db1e
  12. github.com/pollen-robotics/microduck_rl/blob/ceca01706143d729305baa94cbd8ab990979e325/scripts/infer_policy.py

Next useful guides

MicroDuck training to deployment: a complete reading path

Follow preparation, local training, export, simulation replay and deployment checks in order. A successful export or upload is not evidence that the robot can execute the policy.

MicroDuck Local Training: Setup, Checks & First Run

Prepare the official MicroDuck RL environment with a pinned revision, Python 3.12, a CUDA check and a short training run before a longer experiment.

MicroDuck troubleshooting: updates, pairing and policies

MicroDuck 0.20.1: check access before recovery. Wi-Fi SSH stays a local decision. Check daemon/API versions and Setup Code before treating a refused connection as an update failure.