| --- |
| library_name: onnx |
| pipeline_tag: reinforcement-learning |
| tags: [microduck, microduck-policy, mjlab, robotics, swing] |
| license: apache-2.0 |
| --- |
| |
| # Microduck on a swing |
|
|
| A fully 3D-printable standing swing for Microduck, plus three policies that chain into a full ride: |
| **hop on → pump → slow stop → walk off → stand**. SIMULATION ONLY. |
|
|
| <video controls muted loop src="https://huggingface.co/nocrab/microduck-swing/resolve/main/media/ride_full.mp4" style="max-width:640px;width:100%"></video> |
|
|
| The video shows one ride in real time (about 28 s), then 4× slow-motion replays of the hop-on and the hop-off. |
|
|
| ## ⚠️ Simulation only |
|
|
| Nothing here has been printed or run on a real duck. The mount and pump policies also need an input |
| the robot software doesn't provide yet: the pose of the deck (see *Observations*). Treat this as a |
| sim result and a starting point, not a trick to install. |
|
|
| ## The swing |
|
|
| The duck's legs are about 10 cm long and an XL330 hop clears about 4–5 cm, so it **stands** on a |
| low deck (top 40 mm above the floor) instead of sitting. The deck is 220 mm wide between the side |
| rails and open at the front and back. It hangs from two bowed arms on 608 bearings. Every part fits |
| a 256 mm bed. The STLs are in `stl/` (`_xN` = how many to print). |
|
|
| | Printed (PETG) | Qty | ~g each | |
| |---|---|---| |
| | `deck_x1` (print as-is, ribs down) | 1 | 209 | |
| | `arm_x2` (print on its side) | 2 | 44 | |
| | `tower_upper_x2` (print flat) | 2 | 82 | |
| | `tower_lower_x4` (print flat) | 4 | 58 | |
| | `base_splice_x2` (print standing) | 2 | 6 | |
| | `crossbar_half_x2` (lap-jointed in the middle) | 2 | 42 | |
|
|
| Hardware: 2× 608 bearing, 2× M8×45 bolt + nyloc nut + 4 washers, about 18× M4×20–25 bolt + nut. |
| Screw or clamp the base down for big swings. The simulation assumes 3 walls, 4 top/bottom layers and |
| 25 % gyroid infill (55 % of solid PETG mass). Its empty-swing period is **1.078 s**, so time ten swings |
| of your print to check it matches. Fits (bearing press fit, nut pockets) are untested on a real printer. |
|
|
| ## The policies |
|
|
| All three use the standard 61-D Microduck contract (obs[1,61] → actions[1,14] at 50 Hz, observation |
| normalizer baked in), exported with `scripts/export.py`. |
|
|
| | File | Skill | Command (twist slots) | Head / body slots | |
| |---|---|---|---| |
| | `policies/mount.onnx` | walk up to the swing and step onto the deck, then stand | `[0, 0, 0]` | body = deck pose | |
| | `policies/pump.onnx` | pump the swing / ride calmly and brake it | `twist[0]` = 1 pump, 0 calm | body = deck pose | |
| | `policies/dismount.onnx` | walk off the open front or back of the deck and stand | walking command (vx, vy, wz) toward the nearer open end | zeros (walking contract) | |
|
|
| **Observations.** For mount and pump, the 6 body-command slots carry the deck pose in the duck's |
| yaw frame: `[3·x, 3·y, 3·z, sin(yaw), cos(yaw) − 1, hinge angle]`, where (x, y, z) is the deck-top |
| centre relative to the trunk and yaw is the deck's heading relative to the duck. On a real duck this |
| would come from a fiducial marker on the frame plus a hinge angle sensor (not built). The dismount |
| policy needs no deck input. It is a walking policy you steer off the deck with a stick-like command |
| of about 0.15 m/s toward the exit, turning toward it. |
|
|
| ## Results (simulation, MuJoCo Warp, domain randomized) |
|
|
| Pivot friction, deck friction, swing mass, robot mass, CoM, joint friction, encoder bias and IMU |
| mounting are randomized as in training. |
|
|
| **Full ride** (30 rides): mount, pump 10 s, `twist[0] = 0` until the swing is below 2.5°, dismount, stand. |
|
|
| | | Rides | |
| |---|---| |
| | completed every phase without a fall | 28 / 30 | |
| | …and nothing but the feet ever touched the swing | 25 / 30 | |
| | failures by phase | mount 4 (beak or shin brushing the frame), calming 1 fall, dismount 0 | |
|
|
| **Mount** (40 mm deck): 84 % clean (no non-sole contact), 75 % at 60 mm. |
|
|
| **Pump**: from 1° the swing reaches 20° in about 1.5 s and holds 23–28°. The duck pumps mostly by |
| throwing its head and trunk along the swing in a deep crouch, not by crouch-and-rise. |
| **Calm** (`twist[0] = 0`): it brakes actively and quietly, 25° → about 6° in 4 s and about 2° in |
| 16 s, where friction alone would leave 14–19°. Pump for 10 s then stop: 88 % ride the full 20 s. |
|
|
| **Dismount**, at the real hand-off from calm mode: 95–97 % clean (on the floor, clear of the deck, |
| no shin or beak touch), 2–3 % falls. The push-off kicks the empty swing back up to about 13–17°. |
|
|
| ## How it was trained |
|
|
| [mjlab](https://github.com/mujocolab/mjlab) + PPO, with tasks built on |
| [pollen-robotics/microduck_rl](https://github.com/pollen-robotics/microduck_rl) (commit `4757d6ae`) |
| and the swing modelled from the same numbers as the STLs: |
|
|
| - **Mount**: a deck-height curriculum from 10 to 80 mm. Any touch of the swing by something other |
| than a sole voids the episode. The arms are bowed outward at head height after the beak kept |
| brushing straight arms. |
| - **Pump**: starts from recorded mount-policy stances, sideways on the deck. Facing along the swing |
| turns the deck into a rocking board. The reward is swing amplitude, capped at 30°, with a |
| potential-based growth term paid only while riding cleanly; in calm mode the growth term flips sign. |
| - **Dismount**: three attempts that started from the mount policy never stepped off the deck. The |
| actor is instead **Pollen's released walking policy** (`alpha_walking.onnx`, policy release |
| seed-v8), fine-tuned with a deck-height curriculum and then on recorded real hand-off states, |
| which took clean dismounts at the hand-off from 54 % to 95–97 %. |
|
|
| ## Credits |
|
|
| Microduck, its training stack and the walking policy the dismount is fine-tuned from are by |
| [Pollen Robotics](https://huggingface.co/pollen-robotics) (Apache-2.0). Swing design and swing |
| policies by nocrab, made with Claude Code. |
|
|
