MicroDuck Reference Policies: Sprint, Rollers & Sit/Stand

Choose by behaviour and hardware, then check the slot.

Each manifest declares schema 2, model API 1, 61 observations, 14 actions and 50 Hz. Those common numbers do not make the behaviours interchangeable: the roller controller and the sit flag reinterpret command fields. Preserve robot model, servo family, accessories, entry pose and command encoding in your test plan. Official policy manifest schema and compatibility rules ↗

Use one pinned candidate and retain the reset command.

First save robotctl version, policy list and health. Confirm your runtime accepts the manifest and prepare a supported test area with access to power. The following loads the reviewed sprint snapshot, changes the walk slot and persists across reboot; its final line removes that override.

sudo robotctl policy load walk pollen-robotics/microduck-sprint-2m-reference@62afcc461f1d4a5ac22bfbff55ce4900c1afba32
robotctl policy list
robotctl health
sudo robotctl policy reset walk

A reset resolves the official default; it does not restore a previous custom override. Save that override's repository and revision before the trial. The model card says observation normalization is already inside the ONNX model: a custom runner should feed raw observations without applying the same normalization twice. Sprint reference: loading, normalization and training record ↗

For the talent model, use its documented sitstand slot, not walk. Revert that choice with:

sudo robotctl policy reset sitstand
robotctl policy list
robotctl health

Public weights are available; full training reproduction needs more.

The manifests record task, source commit, checkpoint, seed and the training command. The cards point to pollen-robotics/microduck-challenges, but that GitHub repository returned 404 to our anonymous check on October 8. We could inspect public weights and metadata, not the referenced training source. Talent reference: loading and training recipe ↗

The talent card also names Mjlab-GotATalent-MicroDuck in its training summary while its recipe uses Mjlab-SwagContest-MicroDuck. Confirm the task in an accessible source checkout before spending compute. Repeating a seed is not a promise of identical GPU-trained weights.

Report observations with their limits.

Record the Hub revision, daemon version, previous slot, support conditions, start/stop behaviour and reset result. A published reference is a useful baseline, not a current ranking or proof of compatibility with HD-1910, Open Duck Mini or another reconstructed robot. Use the related discussion to compare records with the same model and conditions.