Hugging Face's Pollen Robotics Opens Preorders for the $399 Microduck Bipedal Robot
Hugging Face's Pollen Robotics opened preorders on August 27, 2026, for the Microduck, a one-eyed bipedal robot standing 25 cm tall and priced at $399. The robot comes in cream, graphite, lavender, and sky blue, with shipping planned before Christmas 2026 (The Verge; Pollen Robotics; Axios).
The Microduck runs on a Rockchip RK3566 processor — a quad-core ARM chip more commonly found in single-board computers like the Raspberry Pi class of devices. It packs 15 motors driving articulated legs, head, and neck, along with a grasping beak. Its sensor suite includes a camera (housed in a single eye), motion sensors, and LiDAR, which uses laser pulses to measure distances and build a sense of the surrounding space. Audio input and output are handled by an onboard speaker and microphone, with Wi-Fi and Bluetooth for connectivity (The Verge; Pollen Robotics; Yahoo Tech).
The robot's software is open-source, and Pollen Robotics says it is trained using reinforcement learning — a technique where the robot learns behaviors through trial and error, rewarded for actions that work well. It arrives functional out of the box. Users can retrain the Microduck to perform new behaviors. Out of the box, it can react to its surroundings, follow a laser pointer, accept input from a game controller, and sing. It can also recover to a standing position on its own after falling (The Verge; Pollen Robotics).
Each unit generates a unique audio identity on first wake-up and retains that voice for the life of the device, giving every Microduck a distinct vocal signature (The Verge).
The Microduck is Pollen Robotics' second robot. The company also produces the Reachy Mini, an open-source expressive companion robot aimed at human interaction, creative coding, and AI experimentation (Hugging Face). Both platforms sit under the Hugging Face umbrella, which has been steadily expanding from its origins as a model hub — a shared repository for machine learning models and datasets — into physical, embodied AI.
The $399 price point is worth pausing on. Consumer and educational robotics have historically hit a wall between toys with limited programmability at the low end and research-grade platforms costing thousands of dollars at the high end. The Microduck's spec sheet, with LiDAR, 15 degrees of actuation (meaning 15 independently controlled joints), and reinforcement-learning-based locomotion, reads closer to the latter category than the former. Whether the RK3566 can sustain responsive real-time control under that sensor load is an open question that only hands-on testing will answer.
The open-source angle is where the Hugging Face lineage becomes more than branding. A robot whose software stack can be forked, retrained, and shared through the same collaborative infrastructure that the ML community already uses for models and datasets lowers the barrier for experimentation meaningfully. For researchers working on reinforcement learning for bipedal locomotion, or educators building curricula around embodied AI, a sub-$400 platform with a LiDAR-equipped sensor suite and a motor count in the teens is a genuinely accessible entry point. The grasping beak adds a manipulation dimension that pure-locomotion platforms at this price tier typically lack.
The unique audio identity feature is a small but deliberate design choice. Voice consistency across a device's lifetime has been a pain point in consumer robotics, where factory resets or firmware updates often wipe personalized characteristics. Generating the identity once, at first boot, and persisting it is a simple approach that sidesteps that problem entirely.
The broader question is whether a 25-cm biped can maintain the kind of stability and responsiveness that makes it compelling beyond a novelty. Bipedal locomotion is computationally expensive even at small scale, and recovery from falls, which the Microduck claims, is one of the harder problems in legged robotics. The combination of RL training and the relatively modest compute budget of the RK3566 suggests Pollen Robotics is leaning on simulation-trained policies — behaviors learned in a virtual environment and then transferred to the physical robot — that can run efficiently at inference time rather than requiring heavyweight online planning. That approach has matured rapidly in research settings over the past several years, and a consumer product shipping before Christmas 2026 would be an early test of how well it translates to a mass-produced, low-cost platform.
For now, the Microduck is a preorder page and a spec sheet. What it promises is a capable, hackable bipedal robot at a price that puts it within reach of individual developers, educators, and hobbyists who have previously been shut out of hardware with this kind of sensor and actuator density. Whether it delivers on that promise will become clear when units start reaching buyers.
