Distill neutral-head Flamingo controller

[abhinav3]

Summary

This experiment distills the published Microduck Flamingo balance actor into a
neutral-head policy while preserving its mature leg controller.

• Keeps the teacher observation normalizer and hidden layers.
• Fits only the final linear layer against fixed-head teacher rollouts using
  ridge regression.
• Constrains all four head action outputs to the neutral HOME target.
• Includes the ONNX policy, source dataset, distillation/evaluation scripts,
  raw telemetry, checksums, and media.

Demo

Download the full-quality 30-second MP4

Independent simulation result

Measurement	Result	Acceptance
Left-foot single-support fraction	100%	>= 95%
Right-foot single-support fraction	99%	>= 95%
Raised right-foot median height	9.67 cm	>= 5 cm
Raised left-foot median height	7.64 cm	>= 5 cm
Final two-foot contact fraction	100%	>= 95%
Non-foot ground-contact samples	0	0
Maximum trunk tilt	27.93 deg	< 60 deg

The command timeline holds the left-foot and right-foot poses for 10 seconds
each, with two seconds excluded from each scored window as transition time.

Validation

• Fresh evaluation reproduced the committed metrics.
• The committed dataset reproduces the ONNX byte-for-byte with ridge 25.
• ONNX SHA-256:
  ca1d39539074655ad344bcf95cffb352215ba1d25b444c315cebf8a792a63e5f
• Machine-readable results:
  results/summary.json

Provenance and limitation

The teacher is Pollen Robotics / Remi Fabre's published
microduck-flamingo-cycle
policy and is attributed under Apache-2.0.

"Neutral head" means neutral actuator targets, not mechanically locked
joints. Small passive head deflection remains under load; the README reports
the measured joint ranges. This is a simulation result and does not claim
hardware validation.