beak-throw

Wind up, throw a 24 mm ball from the beak, and recover to a two-foot stand.

The linked 4K preview shows three separately reset episodes from the released checkpoint at half speed with low, locked cameras: the ball comes toward the lens, then crosses to the right and left of it.

• Weights and model card: q2p/microduck-beak-throw
• Complete tester bundle: v0.1.0-sim release

Simulation-validated, hardware-unvalidated. The policy completed 50/50 randomized simulator trials only with the mandatory runtime anatomical clamp. It failed the strict raw-output gate. Read SAFETY.md before any supervised hardware experiment; the first powered run must be empty-beak.

Command — one-shot skill: robotctl robot do beak-throw. The 14-action network controls the body and head. The patched runtime holds the mouth at +20°, opens it to +30° over phase 0.30–0.34 of the 2.4-second motion, then hands off to the supplied standing policy.

What it does — throws a 24 mm / 3 g simulated ball about 0.53 m forward. The ball is physically retained by compliant pads, a rounded front lip, a rear stop, and side cheeks rather than by a weld or teleport.

Limits — 50/50 randomized trials passed the release, landing, ≥0.50 m forward, ≤0.20 m lateral, and hybrid-recovery checks. The raw network exceeded the strict target limit in 50/50 trials (median 0.751 rad), making the supplied runtime clamp mandatory. Full results: docs/ACCEPTANCE.txt.

Try it in simulation

git clone https://github.com/llama/microduck-beak-throw.git
git clone https://github.com/pollen-robotics/microduck_rl.git && cd microduck_rl
git checkout d424a0c899f6b33cbd3daeb279913134349c0b63
git apply ../microduck-beak-throw/training/microduck_rl-shared-files.patch
rsync -a ../microduck-beak-throw/training/overlay/ ./
uv sync
uv run hf download q2p/microduck-beak-throw checkpoints/model_2150.pt --local-dir policies/beak-throw

Linux:

uv run play Mjlab-BeakThrow-Hardware-MicroDuck \
  --checkpoint-file policies/beak-throw/checkpoints/model_2150.pt --num-envs 1

macOS:

ln -sf "$(.venv/bin/python -c 'import sys; print(sys.base_prefix)')/lib/libpython3.12.dylib" .venv/libpython3.12.dylib
.venv/bin/mjpython .venv/bin/play Mjlab-BeakThrow-Hardware-MicroDuck \
  --checkpoint-file policies/beak-throw/checkpoints/model_2150.pt --num-envs 1

Try it on the robot

This requires the included runtime patch; it is not a simple policy swap.

git clone https://github.com/pollen-robotics/microduck.git && cd microduck
git checkout 590b986bd8c0d50ae02cb3ea2f59c463b6828168
hf download q2p/microduck-beak-throw beak_throw.onnx alpha_stand.onnx runtime/microduck-runtime.patch --local-dir beak-throw
git apply beak-throw/runtime/microduck-runtime.patch
cp beak-throw/beak_throw.onnx policies/beak_throw.onnx
cp beak-throw/alpha_stand.onnx policies/alpha_stand.onnx
scripts/dev-push.sh --docker radxa@<robot>

Then add the values in runtime/robotd-beak-throw.toml to the robot's existing [policy] section, restart robotd, and follow hardware/INSTALL_AND_TEST.md. Measure and adapt the reference liner before any loaded test.

robotctl health
robotctl monitor
robotctl robot enable
robotctl robot do beak-throw     # EMPTY BEAK FIRST

Repository map

• training/: reproducible patch and overlay for the exact microduck_rl base.
• runtime/: mandatory Microduck runtime patch and opt-in configuration.
• hardware/: supervised test protocol and editable compliant liner.
• docs/: provenance, acceptance, and software-validation reports.
• media/preview.mp4: three 4K, half-speed, low locked views—toward, right, and left.

Contract obs[1,61] f32 → actions[1,14] f32, normalizer baked in, 50 Hz, kind episodic, duration 2.4 s, entry pose standing.

Provenance source base pollen-robotics/microduck_rl@d424a0c; runtime base pollen-robotics/microduck@590b986; selected checkpoint model_2150.pt.

Apache-2.0. Upstream robot assets retain their original notices and licenses.