microduck-step-up

step-up

Experimental Microduck policy for ascending a 25 mm square-edged household step; simulation-validated, hardware-unvalidated.

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

[!WARNING] This is an experimental, simulation-validated and hardware-unvalidated policy. It does not detect the step itself. For a first physical run, align the robot before the edge and use a safety tether, soft floor, conservative gains, and an emergency stop.

Run it on a robot

sudo robotctl policy load walk Nupr-Haokun/microduck-step-up

This temporarily replaces the normal walking slot. The intended controller switches to this policy only after an external camera or operator has aligned the robot with the threshold, then restores the normal walking policy after the crossing. Do not use it as a general household walking policy.

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
• task: Mjlab-StepUpFull-MicroDuck
• checkpoint: 1000
• base commit: d424a0c
• upstream contribution: microduck_rl PR #25

The ONNX observation normalizer is baked into the graph. The official publisher validates the [1,61] -> [1,14] contract and smoke-runs the policy for finite, non-constant output.

Simulation evaluation

The complete two-policy experiment (crossing plus stand recovery) was tested on an 800 mm-wide, 25 mm-high square-edged step. It completed 13 of 14 deterministic alignment cases. The known failure started 20 mm off-center at -10 degrees: the robot reached the upper surface and then exited through the side. These results are not a sim-to-real guarantee.

The full research package, recovery policy, checkpoint, source snapshot, videos, checksums, and detailed safety notes remain available in Nupr-Haokun/microduck-step-up-head-brake.