duckwing-v80-roller-skating

DuckWing V80 — MicroDuck roller-skating policy

Historical V80 release. DuckWing V89 is the recommended general-purpose skating release, with improved high-speed braking and the published 180°, 360° and 720° controller bundle. This V80 repository preserves its trained graph and weights, numeric evidence and interface, with a documented metadata-only privacy cleanup. V89 requires a 256-frame [1,15616] observation-history input; check the version-specific model contract and adapter before switching. V90 Laguna is a separate racing research bundle with failed combined-transition qualification. The GitHub workbench is the canonical source for these published releases.

DuckWing V80 is a 50 Hz ONNX control policy for the Pollen Robotics MicroDuck with passive roller skates. It is the historical V80 simulation-qualified DuckLab release under the official evaluation physics: CPU MuJoCo, a 1.75 A XL330 current limit, and 0.003 passive-wheel bearing friction.

V80 preserves the qualified DuckWing V67 policy exactly at commands up to and including 0.5 m/s. Above 0.5 m/s, it adds a small state-feedback residual to the left/right hip-pitch, knee, and ankle actions. The result is a modest speed gain with materially better heading, drift, tilt, and grounded contact.

Status: simulation-qualified hardware candidate. It has not been validated on a physical robot. Do not treat the simulation limits as a hardware safety certification.

Results

All values below use wheel friction 0.003, current limit 1.75 A, 50 Hz control, and 200 Hz physics.

The public comparison below uses the official Pollen roller policy as the baseline under the same Race5 evaluator and physics—not DuckWing V67.

V80 wins all six direct comparisons against the Pollen baseline and passes all 15 DuckLab Race5 qualification gates. In additional V80-only tests, its clean 20-second screen measured 2.574 mph mean world speed, and five randomized 20-second starts averaged 2.497 mph sustained speed with 100% survival.

The often-mentioned 5.405 mph donor result is not V80's official-friction speed. It was an instantaneous peak from an older policy at zero wheel friction and is included only as research context.

Policy contract

• Input: obs, float32, shape [1, 61]
• Output: actions, float32, shape [1, 14]
• Control rate: 50 Hz
• Action scale: 1.0 around the policy's embedded MicroDuck default pose
• Kind: perpetual velocity controller
• Entry pose: upright standing pose used by scene_rollers.xml

The observation layout is:

The speed residual activates only when twist[0] > 0.5. Official evaluations use twist = [0.8, 0, line_correction]; low-speed steering, braking, idle, and other V67 behavior remain unchanged by the V80 residual.

Download and reproduce

git clone --recurse-submodules https://github.com/jvpflum/microduck-lab.git
cd microduck-lab
hf download juicenv/duckwing-v80-roller-skating policy.onnx --local-dir policies/duckwing-v80

upstream/microduck_rl/.venv/bin/python tools/evaluate_speed_discovery.py \
  policies/duckwing-v80/policy.onnx \
  --episodes 5 --duration 20 --command-mps 0.8 \
  --current-limit 1.75 --wheel-friction 0.003 --race-line-control \
  --yaw-kp 0.70 --lateral-kp 0.22 --yaw-kd 0.07 --max-correction 0.15

For an interactive simulation after installing the upstream environment:

cd upstream/microduck_rl
uv run scripts/infer_policy.py --roller \
  --walking ../../policies/duckwing-v80/policy.onnx \
  --new-cmd-obs --current-limit 1.75

Provenance

• DuckLab repository: jvpflum/microduck-lab at a1f1f580994574860f33659fe257917442cb028c
• Evaluator/training fork: jvpflum/microduck_rl at 0f1a54b328090a59346e46c0ee72278433cf62bf
• Parent policy: immutable DuckWing V67 release
• Construction: command-gated dynamic residual with gains joint_position=-0.075, joint_velocity=-0.010, previous_action=-0.070
• Policy SHA-256: 18da0d664cc0de8e8bbcab6bea6811cd384413cfeda15094aaa5322a53c891e4

Included files

• policy.onnx — inference-complete policy with embedded normalizer
• manifest.json — machine-readable contract, provenance, and evaluation
• metrics.json — concise V80 leader metrics
• evaluation-summary.json — complete deterministic Race5 evidence
• noisy-start-evaluation.json — five randomized 20-second trials
• clean-screen-evaluation.json — isolated clean-start screen
• verification.json — ONNX/interface verification
• SHA256SUMS — checksums for every published file
• LICENSE — Apache License 2.0

Limitations

• Simulation only; no physical-robot thermal, electrical, impact, or fall test.
• Qualified for the pinned passive-roller model and evaluator physics, not arbitrary wheel geometry, floor material, payload, battery, or firmware.
• The improvement over V67 is real but small; this release prioritizes a strict all-around gate rather than maximum unconstrained peak speed.
• External line-hold steering is part of the published Race5 measurement.

Full research history and rebuilding tools are in microduck-lab.

Metadata privacy cleanup

Private training paths were removed from descriptive ONNX metadata and evaluation JSON. The trained graph, numeric results and single-frame [1,61] interface are preserved. privacy-normalization.json records graph equality, inference parity and the original/current download hashes. Historical model hashes in evaluator provenance still identify the originally evaluated bytes.