Hugging Face Launches $399 Open Source Microduck with 15 Motors, Camera, and Mini LiDAR
Pollen Robotics, a subsidiary of Hugging Face, has opened pre-orders for the Microduck, starting at $399, excluding tax and shipping. The goal is to ship to North America, Europe, and the UK before Christmas, with color options in cream, graphite, lavender, and sky.
This armless bipedal robotic duck stands 25 cm tall and weighs less than 800 grams. It features 15 motors, a camera, a mini LiDAR, two sets of inertial measurement units, a microphone, a speaker, Wi-Fi, and Bluetooth, and can grasp objects weighing around 800 grams with its beak. The battery capacity is 2600 mAh, providing about one hour of runtime. It has a 50 Hz onboard strategy loop, a Rockchip RK3566 processor, 1GB of RAM, and 32GB of storage.
It comes pre-loaded with seven trained actions: swaying walk, sitting and standing, kicking a ball, pecking at objects, wheel sliding, and self-righting after common falls. It includes a game controller, can chase a laser pointer, and responds to its surroundings. Each unit generates a unique sound upon first startup. Demonstrations show it pecking up dirty socks and markers, and resetting after being pushed over.
A software development kit, MuJoCo simulation, and complete reinforcement learning training stack are available under the Apache 2.0 license on the code platform, allowing training in nearly 4096 parallel environments either locally or on Hugging Face's computing tasks, with the ability to export to ONNX for deployment on real hardware. Pollen notes that mechanical and electronic design files are not open source. It is ready to play out of the box, targeting developers who want to train new actions through trial and error.
Pollen was founded in 2016 by former researchers from Bordeaux, France, and was integrated into Hugging Face in April 2025, becoming part of its robotics team. Its previous desktop robot, Reachy Mini, was priced at $499 for the full version and $399 for the lightweight version, focusing on interaction; Microduck shifts towards mobility and fall recovery. CEO Clem Delangue stated this is an affordable open-source robot that can learn new tricks through reinforcement learning, aimed at lowering the barriers for physical AI and world models, targeting around 50,000 units and approximately $20 million in revenue.
Pre-order funds will go towards the complete machine, accessory packages, and subsequent cloud training time. Buyers are developers and geeks looking to transfer simulations to reality on the actual machine, while sellers aim to extend the model community to desktop hardware. The event-driven product rollout is not a large laboratory order. Beneficiaries include low-cost data collection and strategy sharing networks; those under pressure are the quadrupedal and humanoid complete machines and closed training kits that cost tens of thousands of dollars. The cost of falling has been reduced to desktop impacts, moving trial and error from laboratories into living rooms.
Source: Public Information
ABAB AI Insight
Hugging Face first established itself as a model hosting and weight distribution platform, then acquired Pollen in April 2025, integrating the Bordeaux open-source robotic arm team into its internal production line. After creating the Reachy Mini as a desktop interactive robot, the next model was intentionally designed to fall: first selling a breakable body, then sellable strategies. Pollen has been pursuing an open-source complete machine route since 2016, which is not the same capital path as Boston Dynamics' high-priced closed machines or UBTECH's industry exhibition machines.
The $399 price targets developer wallets, not household cleaning budgets. Funds are directed towards motors, sensors, and Christmas production capacity, with profit expectations tied to subsequent Jobs training, strategy uploads, and accessory sliding kits. The motivation is to move "practice in simulation, run on real machines" from university labs to personal workshops, using the same community distribution method for gait weights. Whoever owns the falling data will own the next round of world model scraps.
The benchmark is not factory humanoids like Optimus or Figure, but programmable desktop pets like Anki Cozmo from the 2010s, and low-cost quadrupeds from Yuzhu that spread motion strategies into the geek community. The industry phase is shifting from laboratory control to consumer-level expansion: while humanoids are still burning financing to prove they can climb stairs, this duck first proves that it can recover from falls and that strategies can branch. Closed-source structural components and open-source brains are isomorphic to the mobile phone industry's "molds retained, systems released" approach.
The essence is technological substitution. High-priced real machines make reinforcement learning trial and error so expensive that only a few labs dare to let them fall; by reducing the body to under 800 grams, the cost of failure shifts from repair bills to a sound on the desktop. The migration from simulation to real machines no longer requires proprietary production lines, and pricing power shifts from complete machine brands to reusable strategy files. When open-source stacks can train deployable actions on personal graphics cards or platform computing power, the rental costs of closed robotic operating systems begin to be dismantled.
ABAB News · Cognitive Laws
Only if it can fall, can it be trained.
What is open-sourced is the strategy, what is locked is the mold.
Where the threshold is lowered, data begins to explode.