Microduck Skill

Microduck RL

Close the loop: spec → sim → package → Hub → localhost control → robot. You are training a 61→14 PPO policy, not an LLM. Hub publish is not the end — deploy is in scope. A published repo is a deliverable for whoever owns a duck, not a baked-in official weight.

Canonical playbook: AGENTS.md in pollen-robotics/microduck_rl. This skill is the job ticket; that file wins if they disagree.

$SKILL = directory that contains this SKILL.md. Always invoke scripts with that absolute path. The skill id is microduck-skill.

Preconditions

1. Run $SKILL/scripts/doctor.sh (add --clone --sync if there is no checkout). It must print doctor ok.
2. Default compute is Hugging Face Jobs. Require HF_TOKEN (or hf auth login) and MICRODUCK_HF_NAMESPACE (user or org). No namespace → Jobs prompts and the agent hangs.
3. Local train only if MICRODUCK_TRAIN_LOCAL=1 and CUDA is present. Smoke uses the same switch.
4. Stop if doctor fails. Do not invent another trainer (Isaac Sim, AutoTrain, Colab-only hacks).

Hard rules

• Never skip the 64-env × 5-iter smoke test ($SKILL/scripts/smoke.sh).
• Never hand-convert a .pt to ONNX. Only upstream scripts/export.py.
• Observation stays 61D; action stays 14D. Unused command slots are zero-padded, never deleted.
• Do not hardcode joint indices. Servo helpers in mdp.py only.
• Community publish is constant-command episodic/perpetual only. Phase / posture-flag stay official.
• Do not run robotctl / SSH unless the user explicitly says to install on a real robot. Print the commands.

Workflow (do not skip stages)

1. Spec — one-line behavior, kind (episodic | perpetual), closest template. references/TRAIN.md.
2. Sim — edit env → CPU pytest tests/ → $SKILL/scripts/smoke.sh → $SKILL/scripts/train.sh. If wandb is logged in, watch the run page (curves; Jobs cannot record --video — no GL in the image). Keep penalty Episode_Reward/* ≤ 0. If not, tensorboard + --no-wandb so Jobs do not hang. Replay a checkpoint with $SKILL/scripts/play.sh (Viser + local mp4).
3. Package — $SKILL/scripts/export_publish.sh. references/PUBLISH.md. After a tensorboard run, export from --checkpoint-file (or an already-gated --onnx), not a wandb path. Then try $SKILL/scripts/preview.sh --repo <this-run-repo> --task … (or --mp4 if a clip already exists). That is a sim checkpoint replay on this Hub card (preview.mp4), not ONNX control, not a robot, not a ranking. play.sh is a viewer and does not exit; preview starts it, waits for the clip, then stops it. Play/upload failure is not a failed publish — say preview skipped and continue. Do not put mp4 in the plugin git.
4. Sim control — you run $SKILL/scripts/control.sh yourself (localhost HTTP). Do not paste these commands for the human to type. search lists Hub policy.onnx (not a store). You pick a repo by name/README match, then start --repo / pull --as. Worked Hub examples (not bundled, not official Pollen): neil-jo/microduck-walk (perpetual; start --repo) and neil-jo/microduck-polite-bow (pull --as then do). Read status.command_notes before twisting — walk speed caps are not universal. Read status.body_state to say what the body did (tilt_deg, feet, joints vs HOME) — same card for every policy. Do not call it fallen. Episodic Hub graphs: pull --as name then do name (same swap as kick). Do not start --repo them. Exercise twist/head/do, and report from that card. references/CONTROL.md. status.skills is loaded or untrained. --record + dataset.sh is datacollect only. Bind stays localhost. Do not start a train from a search hit.
5. Deploy — always this stage after publish. Print the install lines from references/DEPLOY.md for this run's --repo (never a hardcoded test repo). If the user explicitly says to install on a Microduck they own, run those robotctl lines. If they have no robot or did not ask, stop after printing. Do not claim hardware success until they report it.

Scripts

Report back

Say what rolled out ("bows, then face-plants 1 in 3"), not "it works". Training is hours — submit Jobs, then wait.