allen73/microduck-physical-ai

Policies

The ONNX policies robotd runs. All of them are obs[1,61] -> actions[1,14]; robotd checks that at load rather than discovering it mid-stride.

This is a temporary home

These belong on the Hugging Face Hub, delivered as a model updater component that versions independently of the daemon — a gait retrain should not need a daemon release, and a daemon fix should not re-download 6 MB of unchanged weights. deploy/updater.toml already describes that component and deliberately leaves it unconfigured until the repos exist.

They are vendored here because they were not on the Hub yet and the daemon cannot move without them. Committing them makes a release self-contained, which is the property that makes the update path testable end to end: one robotctl update apply turns a standing robot into a walking one.

Removing this directory later is the whole migration: point the [policy] paths in deploy/robotd.toml at wherever the model component installs, and drop the --include lines from .github/workflows/dev.yml, .github/workflows/_build-release.yml and scripts/dev-push.sh — all three, and xtask's every_policy_in_the_repo_is_packaged test is what keeps the three lists honest until then.

Provenance

Copied from apirrone/microduck_runtime at commit 5f3b314 (roulade.onnx at 7e4ab6d, where it first appeared), dereferencing the symlinks that repository uses to give stable names to specific training runs:

here	there	role
alpha_walking.onnx	BEST_alpha_walking_rough.onnx	walking / velstand
alpha_stand.onnx	BEST_alpha_stand_body_control.onnx	standing + body-pose
alpha_sitstand.onnx	BEST_alpha_sitstand.onnx	sit ↔ stand (posture flag)
alpha_ground_pick.onnx	alpha_ground_pick.onnx	ground pick (phase command)
ball_kick_left.onnx	ball_kick_left.onnx	left-leg kick
ball_kick_right.onnx	ball_kick_right.onnx	right-leg kick
roller.onnx	BEST_roller.onnx	roller-mode locomotion
roller_crouch.onnx	BEST_roller_crounch.onnx	roller-mode crouch (ground-pick slot)
roulade.onnx	roulade.onnx	forward roll (Mjlab-Roulade-MicroDuck)

(roller_crouch also fixes the upstream file name's typo.)

The names here are the roles — what deploy/robotd.toml asks for — not the training runs. That indirection is deliberate and worth keeping: swapping which run is "the walking policy" should not mean editing config on every robot.

Why the 61-D family only

The prototype also ships a 51-D family (3 gyro + 3 gravity + 42 joints + 3 command, the legacy [vx, vy, vtheta] command). 61 is the same sensors with the unified 13-value command ([vel(3), head(4), body(6)]) this daemon builds — the only observation it builds. A 51-D file fails at load with one precise line naming both widths:

policy unavailable: .../walking.onnx: observation width is 51, expected 61

That shape check earned its place: it turned a wrong-policy mistake into a diagnosis instead of a robot moving in ways nobody could explain.

Trying your own

No release needed — deploy/robotd.toml takes absolute paths:

[policy]
walk  = "/home/radxa/my_walk.onnx"
stand = "/home/radxa/my_stand.onnx"

Then sudo systemctl restart robotd. A policy that fails to load is reported through robot.health as policy unavailable: <reason> while the loop keeps ticking and holding its pose, so a bad file is visible without putting the robot on the floor.