craigm26/microduck-policy-golden-vectors

Microduck policy golden vectors

Observation → action pairs recorded from Pollen Robotics' trained Microduck policies, so that anybody writing their own runner can check it against the same numbers instead of against a video.

This is a conformance fixture, not a model and not a dataset to train on. It contains no weights. If you want the networks, they are Pollen's, in pollen-robotics/microduck and pollen-robotics/microduck_rl.

What is in it

golden_policies.json, keyed by policy filename. Each entry carries the network's sha256 — so you can tell whether you are holding the same file these were recorded from — and a list of cases, each an observation vector and the action that policy produced for it.

Policies	alpha_walking.onnx, ball_kick_left.onnx
Recorded with	onnxruntime 1.29.0
Max reference error	2.38e-06

The observation layout is written into the file's meta block rather than left to be inferred:

gyro(3) gravity(3) jointpos-home(14) jointvel(14) lastaction(14)
cmd(13: [vx, vy, vyaw, neck_pitch, head_pitch, head_yaw, head_roll, 0, 0, z, roll, pitch, 0])

61 floats in, 14 out.

What it is for

Two things go wrong when you implement a policy runner, and neither shows up as a crash:

1. The observation is assembled in the wrong order. A swapped block still produces a plausible gait that is subtly wrong, and you find out on hardware.
2. The normaliser is missed. These networks bake actor(normalizer(obs)) into the graph, so a runner that normalises again gets a duck that almost walks.

Feeding these observations to your runner and diffing the actions catches both in a second. A match to ~1e-06 means your observation assembly and your forward pass agree with the reference.

What it does not tell you

It says nothing about whether a policy is safe, whether it transfers to hardware, or how the robot's runtime scales and filters what the network returns before it reaches a servo — all of which happen after the numbers in this file. See robotd in pollen-robotics/microduck for that half.

Provenance and licence

The policies are Pollen Robotics', Apache-2.0. These vectors are their outputs, generated with onnxruntime and published under the same licence with attribution. Recorded by craigm26 for Duck Studio, an iOS policy inspector, whose test suite asserts against this same fixture.