deep crouch hold
"a duck that squats down into a deep crouch and holds it perfectly still, staying exactly where it stands without stepping"
A Microduck move trained by tfrere with
Microduck Academy. Family velocity
(walking styles (gaits)), tier None, kind perpetual.
Judge (code, on the numbers): PASS (score 1.0) - {"height_ratio": 0.673, "height_m": 0.0773, "speed_mps": 0.0, "displacement_mps": 0.0, "pitch_deg": 7.1, "max_tilt_deg": 7.4, "yaw_rate_rps": 0.004, "head_yaw_ptp_rad": 0.003, "knee_left_rad": 1.31, "contact_fraction": [1.0, 1.0], "travel_m": 0.001, "heading_drift_rad": 0.021, "foot_lifts": [0, 0],
Eye (VLM, on the frames): agrees (The robot drops smoothly into a deep squat by bending its knees during the first second.) - on the frames of checkpoint 5000, the one that ships
Commands (the battery, telemetry only): answers 2 of 2 commands. (command area 100 %; see docs/TRAINING.md, the command battery)
Score shown in Discover: 100 % (the judge's score, capped at 75 % when the eye is not sure and at 50 % when it disagrees; see docs/TRAINING.md section 8).
| round | judge | judge score | eye (VLM) | score |
|---|---|---|---|---|
| 2 | pass | 1.0 | agrees (The robot drops smoothly into a deep squat by bending its knees during the first second.) | 100 % |
Files: policy.onnx (play it), model.pt (remix = fine-tune from it), rollouts/*.traj (trajectory.v1
recordings), checkpoints/r<round>-<iter>.traj (the training steps of every round, listed in
manifest.checkpoints), manifest.json (schema 2 + judge_score / vlm / score + academy block with
prompt, family, judge, lineage), video.mp4 + poster.jpg (the card clip).
Run it on a robot
sudo robotctl policy add deep-crouch-hold tfrere/microduck-move-deep-crouch-hold-2
robotctl robot do deep-crouch-hold
deep crouch hold
a duck that squats down into a deep crouch and holds it perfectly still, staying exactly where it stands without stepping