IgaPro/microduck-veldrop

veldrop

VelStand fine-tuned to land drops of 20-60 cm: servo-housing impacts -80% stepping off 60 cm (simulation). Not yet tested on a real robot.

A perpetual policy for the microduck (61-D observation, 14 actions, 50 Hz). Runs until told otherwise — a gait for the walk slot.

Run it on a robot

sudo robotctl policy load walk IgaPro/microduck-veldrop

The observation normalizer is baked into policy.onnx; feed raw observations. manifest.json follows schema 2 of the microduck policy manifest (docs/policy-manifest.md in the daemon repo).

A remix of pollen-robotics/microduck-policies.

Training

• task_id: Mjlab-VelDrop-Flat-MicroDuck
• repo: pollen-robotics/microduck_rl
• branch: backroll
• commit: b02fdea49
• checkpoint: 6300
• exported from a checkout with uncommitted changes

Install, and how to go back

sudo robotctl policy load walk IgaPro/microduck-veldrop   # replaces the walking gait
sudo robotctl policy reset walk                           # back to the stock VelStand

Read this before trying it

• Simulation only. Trained and evaluated in MuJoCo (mjlab), checked on a second physics path only partially. It has not been run on a real Microduck yet. Start with the robot on its stand, then try low drops (~20 cm) onto a carpet.
• Keep limp_fall ON (the default). It takes over when the robot tips past ~26°, i.e. on step-offs and tumbles. VelDrop matters most where limp_fall never triggers: drops that stay upright (put down from a height, feet first) — and on step-offs it is at least as gentle as going limp in simulation (table below).
• The pickup detector may pause the policy while the feet are off the floor during a drop; that is runtime behaviour, not something this policy can change.

What changed vs. the stock VelStand (simulation)

Warm-started from Pollen's VelStand and fine-tuned with drops from 20–60 cm (feet-down, stepping off an edge, tumbling), linoleum/carpet floor, and impact costs on servo housings, head and trunk raised ×10 during drops only. Walking, turning and getting up are trained as before.

Mean of 128 drops per row, 2 s after the spawn (higher-is-better only for "on feet"):

Drop	Policy	Servo-housing impact	Head impact	Joint jerk (rad/s²)	Landed on feet
40 cm, stepping off	VelStand	11 N	52 N	660	38 %
	VelDrop	2 N	54 N	478	52 %
60 cm, stepping off	VelStand	15 N	75 N	665	21 %
	VelDrop	4 N	65 N	502	30 %
	limp (limp_fall)	13 N	80 N	329	23 %
60 cm, feet first	VelStand	4 N	10 N	451	59 %
	VelDrop	4 N	8 N	443	59 %
60 cm, tumbling	VelStand	11 N	66 N	457	22 %
	VelDrop	11 N	65 N	444	21 %

The video's split screen is the clearest of 128 identical 60 cm step-offs (same pose and physics on both sides, randomisation off); its end card shows the average over all 128. Servo-housing contact is what strips gears — that is where the gain is. Tumbling drops are unchanged, and going limp still has lower joint jerk.

Training code: task Mjlab-VelDrop-Flat-MicroDuck (a VelStand extension), evaluation with scripts/drop_eval.py, video with scripts/drop_video.py.