This duck will teach you reinforcement learning — and pick up your socks
The New Stack Frederic Lardinois ● Covered by 4 sources
Hugging Face’s Pollen Robotics opened preorders for a $399 robot duck that walks, skates and grabs things. It ships before Christmas, and it’s built for reinforcement learning, not just desk-top gimmicks.
Based on reporting by The New Stack, Frederic Lardinois — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Hugging Face’s Pollen Robotics has started taking preorders for Microduck, a $399 introductory robot that looks a lot like a duck and a little like a lab demo that escaped onto the floor. The company says the first units should arrive before Christmas, and sales are set to begin in North America and Europe.
Microduck is meant to do more than sit there and blink. It can walk, waddle, use roller skates, pick things up with its beak, and right itself after a fall. Pollen is framing it as a platform for action, a step up from Reachy Mini, the smaller desktop robot the company and Hugging Face launched last year. Reachy Mini stayed put and talked to people. Microduck is supposed to move.
The software side is where the real pitch lives. Microduck is open source, with an SDK, a virtual training environment, and reinforcement learning scripts. The code is under Apache 2.0, while the hardware design files are non-commercial, so the door is open for experimentation but not for cloning and selling copies. There’s also a simulator for anyone who wants to play with the robot without buying one, plus a game controller for manual control if you do.
Pollen says developers can train movement policies in simulation and transfer them to hardware. On GitHub, the team says walking can be trained across 4,096 virtual ducks in parallel, with a usable gait in one to two hours. The repo lists 13 task families, including a forward roll and six tasks built around passive wheels that fit under the feet. Training needs an Nvidia GPU, though Hugging Face’s own infrastructure is an option too.
The hardware is modest and specific. Microduck weighs about 800 grams and runs on a Rockchip RK3566 with AI accelerator, 1GB of memory, and 32GB of storage. Its sensors include a front camera, a small LiDAR unit with an 8x8 time-of-flight matrix, and two IMUs for balance and orientation. The team hasn’t locked down the final camera resolution or LiDAR range yet. Accessories are already lined up too, including a $39 Charger Pack and a $119 Dev Pack with spare motors, cables, batteries, tags, credit, screws, and a screwdriver.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of silly: a toy that happens to be a serious training platform. The hardware may be small, but the point is big — if you want robot research to escape the lab, you need something cheaper and stranger than another polite humanoid. Also, duck-shaped robots are objectively better than most corporate demo props, which have all the charm of a printer error.
Read more about this at: The New Stack
