Most robotics launches rely on polished demo videos that hide how the behavior was made. Pollen Robotics changes that by shipping the full training loop for Microduck, a 25 cm biped that you can buy for $399. Every motion—walking, sitting, kicking, roller‑skating, standing up after a fall—is a neural policy trained in a physics simulator and exported to the robot. The simulator settings, reward functions, domain‑randomization parameters, and the sim‑to‑real recipe are all public on GitHub, so you can see exactly how the policy was built and reproduce or modify it yourself.
Hardware wise, Microduck packs 15 motors, a front camera, dual IMUs, a compact LiDAR, an 8×8 time‑of‑flight matrix, microphones, a speaker, two NFC antennas, Wi‑Fi, Bluetooth, and a removable 2600 mAh battery good for about an hour. The compute board is a Rockchip RK3566 with an AI accelerator, 1 GB RAM and 32 GB storage. Out of the box you get seven pre‑trained behaviors that you can trigger with a bundled game controller before writing any code.
Because the observation vector (61 dimensions) and the ONNX policy file are shared across all behaviors, you can hot‑swap policies at runtime without rebooting. The onboard Rust daemon runs the control loop at 50 Hz and handles policy swaps over a JSON‑RPC socket, making experimentation fast and safe.
If you have been frustrated by black‑box demos, expensive kits, or locked‑in software, Microduck gives you an open, affordable platform to learn reinforcement‑learning‑to‑real‑world pipelines, tweak reward functions, add new sensors, or design entirely new motions—all with the same hardware you already own.
#AI #Robotics #OpenSource #STEMLearning #Sim2Real #DIYRobotics