Make the duck play sports.

Microduck mini-golf sports build: real recessed MuJoCo cup, frozen Pollen 61→14 locomotion/kick primitives, and a goal-conditioned megapolicy that observes ball, cup, friction, slope, and physical gates to select kick heading/force. PPO continuation WIP. Public repo: https://github.com/ob1-s/microduck-mini-golf. Microduck Sports challenge tagging was rejected by the upload API; this is the public ready fallback.

Bruno Oliveira·Microduck

Built a 3v3 RL Soccer Microduck trained on 80 GPU-hrs

Mikul Rai·Microduck

Rendered Microduck outfit concept - The Flash: scarlet suit, gold boots, gold cowl blades and a chest roundel struck with a scarlet bolt, worn through a simulated 5 m dash in MuJoCo. The kit is geom_rgba plus six zero-density, non-colliding geoms; duck_outfit.py compiles the robot dressed and bare and proves masses, inertias, actuators and contacts identical. Rendered only; no hardware.

Steve Pederzani·Microduck

Iron Man dressed as a Duck! Wait... flip that

Mikul Rai·Microduck

Trained simulated MicroDuck sprint policy (MuJoCo): a 5 m dash with every joint clipped to the XL330 servo's real 1.75 A current limit, 50 Hz, 40 ms actuator delay, sensor noise, domain randomization. Fastest of 240 seeded runs: 5.00 m in 3.55 s, then a commanded brake and a fall, all closed-loop, nothing keyframed. The Flash costume is zero-mass paint, verified by script. Simulation only; not validated on hardware.

Steve Pederzani·Microduck

Maze tag for 5 real Microducks in MuJoCo (Pollen's BAM actuators). Trained here: a stand-up policy (PPO) and the Duck-Man strategy net (imitation, then evolution in-sim); ghosts scripted, gaits Pollen's. 60 held-out seeds: learned 406, planner 360, disabled 0. Sim only.

nyle-prosal·Microduck

Rendered 3D Microduck surf outfit: turquoise vest with palm/wave embroidery, cream piping, coral trim, gold zip, open visor and wooden longboard. The selected AI concept is rebuilt on genuine Pollen geometry. Editable Blender project and 12-second multi-angle video included. Concept only, not physical surfing or manufactured clothing. Source: https://github.com/jrubiosainz/microduck-him-2026-submissions

jrubiosainz·Microduck

Two teams of 3 Microducks play paintball: hits with paint-filled balls eliminate opponents. Win by carrying the rival flag home in the beak. MuJoCo simulation; scripted control with pretrained motion. Full 97.68s game. Contact-triggered ideal grip; estimated jaw dynamics. Costume graphics only; no hardware claim. Source: https://github.com/jrubiosainz/microduck-him-2026-submissions

jrubiosainz·Microduck

A Microduck that is told which third of the goal to aim at, and hits it. Before every shot the policy is commanded left, centre or right, and it scores inside the commanded zone 58 times out of 60, holding 94.7 percent across 300 further shots. After a goal, a second learned policy takes over and celebrates. What to notice: it is aiming, not just kicking.

Scott Middlecamp·Microduck

Six Microducks play Duck Duck Goose in colorful overalls and propeller hats. Scripted strategy, gaze and beak-tag gestures use Pollen Robotics’ pretrained walking/sit-stand networks; no new training. Three scored rounds with seated spectators and role swaps. Disabling runner navigation produces a physical catch; disabling reach prevents the opening tap. Free-base MuJoCo physics, no balance assistance. Source and limitations linked in the README at https://github.com/maddiedreese/duck-duck-goose

Maddie D. Reese·Microduck

I love Gritty, so I rendered a Gritty-inspired Microduck costume: orange shag, googly eyes and a hockey sweater. A head-bob dance uses a small imitation policy trained on HIM, with Pollen Robotics’ pretrained balance. A separate cake-toss take uses a HIM-optimized neck gesture and a beak-mounted serving paddle. Cake flight comes from MuJoCo contact and gravity. Two staged takes; cosmetic fur/clothing, a rigid cake, no hardware validation. Repo at https://github.com/maddiedreese/gritty-microduck

Maddie D. Reese·Microduck

Striker Microduck vs keeper Microduck in MuJoCo. The keeper shuffles along its line to cut down the angle; it concedes 2 of 20 where a stationary duck concedes 14, and saves 8 of the 9 genuine shots it faces. All learned skills are Pollen Robotics' pretrained networks used for inference — we trained nothing. Built on Soren Larson's Double D (Apache-2.0). Simulation only, no hardware validation.

aoalp·Microduck

Trained policy for just the action of a microduck kicking a simulated ball into a net

AdamBangash·Microduck