Pollen Robotics, the robotics team under Hugging Face, released the bipedal robot Microduck last week at $399. In the first 24 hours after pre-orders opened, total order value passed $2.6 million — about 6,500 units — with sales peaking at one unit every four seconds. Working backward from that price, the order value and unit count line up, meaning the total wasn't inflated by pricier add-ons.
The robot stands 25 centimeters tall and weighs under 800 grams. It packs 15 Dynamixel XL330 servos, a camera, a depth sensor and two IMUs, running on a Rockchip RK3566 as its main controller. The duck bill on its head is a movable joint with roughly 800 grams of grip force, letting it carry objects while walking.
A brain made of seven Rust processes
Microduck's software stack isn't built on any web framework. The entire runtime consists of seven Rust daemons that talk to each other over Unix sockets using JSON-RPC 2.0, with messages passed line by line as NDJSON.
The division of labor is clean: robotd has exclusive access to the hardware and runs the 50Hz control loop; configd manages WiFi, identity and controller pairing; updaterd handles OTA updates, checksums and rollbacks; btd passes through Bluetooth GATT; padd forwards controller input; mediad manages the camera, inference and WebRTC; and tofd handles the head's depth sensor on its own. The whole project comes to around 30 dependencies.
Keeping motor control inside a single process is a common survival tactic in embedded robotics — if the control loop ever gets preempted by another task, the robot simply falls over on the spot. Writing this layer in Rust avoids the jitter that comes with garbage-collection pauses.
Walking was trained, not scripted
More than a dozen motion skills — walking, self-righting after a fall, kicking a ball, roller-skating — were never written as state machines. They were trained through reinforcement learning in simulation and then transferred to the physical robot. The training framework is mjlab, built on MuJoCo Warp, using PPO at a 50Hz control frequency with a 61-dimensional observation vector covering proprioception, velocity and posture commands. The final policy is exported to ONNX and runs on the robot itself.
The simulation is more detailed than most open-source projects bother with. Voltage control law, back-EMF, gear backlash (±1°) and Coulomb friction — terms usually left out — are all built into the model, since most sim-to-real transfer failures trace back to exactly these unmodeled effects. Running 4,096 parallel environments for one to two hours is enough to produce a usable policy, a workload that fits on a single consumer GPU.
Both repositories are public: the Rust runtime is at pollen-robotics/microduck, and the reinforcement-learning Python code is at pollen-robotics/microduck_rl.
Supply can't keep up with orders
Manufacturing is handled by Seeed Studio in Shenzhen. The first batch of deliveries has already slipped to after Christmas 2026, and new orders now face a four-to-six-month wait. The company's stated sales target is 20,000 units. At $399 each, that puts revenue in the ballpark of $8 million (a rough estimate) — not a business that moves the needle for Hugging Face as a whole.
It reads more like a launch campaign. Programmable bipedal research platforms like this have typically cost anywhere from a few thousand to tens of thousands of dollars, putting embodied AI out of reach for most individual developers on hardware cost alone. At $399, that barrier drops into the range of everyday personal spending. What Pollen wants is for people to take the stack and modify the policies, publish their own weights, and submit pull requests — the way things played out after model weights went open source, except this time the thing being reproduced is a piece of physical behavior.
The first deliveries won't arrive until the end of the year. Whether the community can sustain the momentum will come down to how many other people's trained policies have piled up in the repository by then.
Sources: Pollen Robotics project repository, CocoLoop, Tony Bai's technical breakdown, Sina Technology; hardware specs and software architecture follow the official repository, order value and day-one unit count cross-checked against public reporting.