microduck-walk-recover

walk-recover

v2: the v1 walk-and-recover policy fine-tuned with +-1 deg of gear play in every servo (Mjlab-VelStand-Flat-Backlash-MicroDuck), simulation only: v1's 20,000 iterations plus 3,000 with the slop, 4096 envs, 62 min on one RTX 5090. With the slop it stayed on its feet for 20 s in 6 of 7 standing takes (v1: 6 of 7); from lying on its belly or back it did not get up (0 of 3 each). Not yet tested on a real Microduck.

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 witcheer/microduck-walk-recover

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).

Training

• repo: pollen-robotics/microduck_rl
• branch: develop
• commit: 53b8971b6
• exported from a checkout with uncommitted changes

Results in simulation (v2, 2026-09-26)

Proof takes with headless_play in Mjlab-VelStand-Flat-Backlash-MicroDuck (±1° of gear play in every servo), 20 s each, trunk height and tilt read from the simulator. A fall is the trunk reaching the floor or a mid-episode environment termination.

• From standing: v2 stayed on its feet for the full 20 s in 6 of 7 takes (one fell at about 4 s). v1 in the same gear-play sim: 6 of 7.
• From lying on its belly: 0 of 3 get-ups for v2 and for v1.
• From lying on its back: 0 of 3 get-ups for v2 and for v1.
• Training, last 100 iterations: time fallen and velocity-tracking error unchanged against v1, mean reward 3% lower.

So ±1° of gear play does not break this walk in simulation, and v2 did no worse than v1. It does not get up once it is down. The training checkout had one local change (the task registration for an unrelated get-up task in tasks/__init__.py); the backlash task itself is unmodified at the commit above.

v1

The first release: 20,000 iterations on Mjlab-VelStand-Flat-MicroDuck, no gear play (326 min on one RTX 5090). Still installable:

sudo robotctl policy load walk witcheer/microduck-walk-recover@v1