microduck-stilts

Eight MicroDuck forward-walking policies specialized to blend-0.50 stilts from 10 cm to 2.0 m. Every row has its own complete silent 10-second video, ONNX deployment graph, and continuation checkpoint.

The root policy.onnx, manifest.json, and replay.mp4 are convenient aliases for the 10 cm default variant. The height-specific directories remain the authoritative files for all eight variants.

Stilt hardware gallery

All videos show alternating stilt support for the full horizon without reset or auxiliary body contact. These are simulation experiments, not hardware-validated policies. The 50 cm–2.0 m variants are extreme controller research results, not recommendations for monolithic printed hardware.

The simulation mass law is 12 g + 1 g/cm per stilt. The 10 cm policy was trained at 22 g per stilt; the prototype slicer estimate is about 29 g per stilt. The unchanged actor survived an initial 64-environment, 10-second simulation battery at 29 g, but broader mass-randomized continuation and tethered hardware testing are still required. See the exact mass audit.

Matching geometry and source

Training code, the executed curriculum, and left/right/paired STLs for every released height are in Vottivott/microduck-playground at commit c5fcc50.

• Exact released STL index
• Executed training lineage
• Parametric generator

Run one policy

git clone https://github.com/Vottivott/microduck-playground.git
cd microduck-playground
git checkout c5fcc50219fef01ac9931d0079c583ccbb29b689
uv sync

hf download HannesVonEssen/microduck-stilts 25cm/policy.onnx config.json \
  --local-dir policies/stilts

MICRODUCK_STILT_HEIGHT_CM=25 MICRODUCK_STILT_BLEND=0.5 \
  uv run python scripts/infer_policy.py \
    --walking policies/stilts/25cm/policy.onnx \
    --new-cmd-obs --lin-vel-x 0.15

Continue training

Each checkpoint.pt preserves the exact released actor and normalizer. Because the original stilt critic and optimizer snapshots were not retained, it uses a fresh compatible critic scaffold, empty optimizer moments, 1e-5 learning rate, and 0.1 exploration standard deviation. It is an actor-exact warm start, not the original full PPO state.

mkdir -p logs/rsl_rl/stilt_locomotion/release-25cm
cp policies/stilts/25cm/checkpoint.pt \
  logs/rsl_rl/stilt_locomotion/release-25cm/model_2800.pt

MICRODUCK_STILT_HEIGHT_CM=25 MICRODUCK_STILT_BLEND=0.5 \
  uv run train Mjlab-Stilt-Flat-MicroDuck \
    --agent.resume True \
    --agent.load-run release-25cm \
    --agent.load-checkpoint model_2800.pt \
    --agent.max-iterations 100

See TRAINING.md for the full lineage, checkpoint caveat, and promotion procedure. PyTorch checkpoints use pickle internally; load them only from a repository and revision you trust.

Runtime contract

• input: obs, float32 [1, 61]
• output: actions, float32 [1, 14]
• control rate: 50 Hz
• action scale: 1.0 rad around MicroDuck HOME joint pose
• observation normalizer: baked into each ONNX
• command: twist=[0.15, 0, 0], zero head/body commands
• morphology: matching height, blend 0.50, 17 × 22 mm rounded tips

The 3.0 m demonstration is intentionally absent: it was a zero-shot failure of the unchanged 2.0 m policy, not a separately trained model.