Microduck Swing
A simulation-trained Microduck policy that pumps its head and legs from rest to swing through a measured 173.20-degree full span.
hf download HannesVonEssen/microduck-swing policy.onnx config.json manifest.json --local-dir policies/swinggit clone https://github.com/Vottivott/microduck-playground.git && cd microduck-playground && git checkout c5fcc50219fef01ac9931d0079c583ccbb29b689 && uv syncuv run python scripts/evaluate_swing_checkpoint.py experiments/swing/checkpoints/alpha050.pt --output /tmp/swing-seed27.json --device cpu --duration 36 --seed 27uv run train Mjlab-SwingPump-MicroDuck --agent.resume True --agent.load-run release-alpha050 --agent.load-checkpoint model_3500.pt --agent.max-iterations 2561 OBS / 14 ACTIONS / 50 HZ REQUIRES RETAINED SEAT AND TWO 380 MM ELASTIC CORDS TWIST-COMMAND SLOTS ARE REPURPOSED AS AN IMU-DERIVED SWING-PLANE CUE STANDARD WALKING RUNTIME REQUIRES AN OBSERVATION ADAPTER SIMULATION VERIFIED / HARDWARE UNVERIFIED
MODEL CARD / SWING PUMP
Microduck Swing
Microduck Swing starts motionless at the bottom of a retained two-cord swing and learns to pump with coordinated head and leg motion. The released video and ONNX policy correspond to the best strict seed-27 simulation rollout.
Project Overview
The release packages the deployment ONNX graph, its manifest, an exact continuation checkpoint, checksums, a deterministic evaluator, and printable retained-seat resources. The actor uses robot-available IMU, encoder, and action-history signals, but requires a task-specific observation adapter.
Task
| Field | Value |
|---|---|
| Task | Pump a retained two-cord swing from a motionless start |
| Robot | Microduck hardware revision 1, 14 Dynamixel XL330 servos |
| Mechanism | Retained seat with two independent 380 mm elastic tension-only cords |
| Entry state | Bottom of arc, motionless |
| Policy horizon | Trained episodically; evaluated for 36 seconds |
| Observation / action | 61 float observations / 14 servo actions at 50 Hz |
| Best published result | Seed 27, 173.20-degree full span |
| Verification | Simulation only |
Intended Use
Use this release to reproduce the published simulation evaluation, inspect or continue the PPO training lineage, and conduct carefully controlled hardware experiments with the specified retained seat, cords, observation adapter, safety tether, current limits, emergency stop, and exclusion zone. It is not a drop-in walking policy and should not be treated as hardware validated.