flamingo

Stand on one foot, either side, on command, and come back to a two-foot stand: twist = [flag, side, 0].

⚠️ Simulation only

Never tested on a real robot. Trained and checked in simulation only (MuJoCo with the XL330 servo model, and duck-sim with the real robot software). If you try it on hardware anyway, keep a hand under the duck: standing on one foot, it falls on pushes of about 0.2 m/s or more.

Credits

This is a retrain of RemiFabre/microduck-flamingo-cycle by Rémi Fabre: same task (Mjlab-FlamingoCycleHard-Flat-MicroDuck on the flamingo branch of pollen-robotics/microduck_rl), same command layout, and control.py is his driver script, unchanged. Only the training run is ours. All credit for the task design goes to him.

Command — twist slots of the 13-D command block (head / body slots: unused (zeros), unused (zeros)); idle = [0, 0, 0]

slot meaning
twist[0] flag: 0 = stand on two feet (HOME), 1 = stand on one foot
twist[1] side: +1 = right foot down / left leg lifted, -1 = left foot down / right leg lifted; 0 allowed while flag = 0
twist[2] unused (0)

Results

test result
model card timeline (right foot, then left, 8 s holds), CPU MuJoCo never falls
duck-sim (real robotd vs MuJoCo body), both sides one-foot hold at gz −0.89, no fall verdicts
32-push grid (0.15–0.3 m/s × 4 directions × both feet) 5/32 falls (the reference flamingo: 6/32)

media/preview.mp4: the published policy under real robotd in duck-sim, front view, driven by control.py --timeline "0:0,1;2:1,1;10:0,1;14:1,-1;22:0,-1" — right foot down for 8 s, both feet, left foot down for 8 s, both feet.

Contract obs[1,61] f32 → actions[1,14] f32, normalizer baked in, 50 Hz, targets around HOME × 1.0. Kind perpetual, entry pose standing. Provenance Mjlab-FlamingoCycleHard-Flat-MicroDuck — pollen-robotics/microduck_rl @ 4757d6ae on branch flamingo (not merged yet), run local 2026-10-25_22-40-46_ours-hard, model_2999 (3000 it, 4096 envs, RTX 5080, 1h44m); exported with scripts/export.py. Files policy.onnx · manifest.json (all of the above, machine-readable) · media/preview.mp4 (sim rollout) · control.py Format: Microduck policy sharing.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading

Collection including nocrab/microduck-flamingo