Pollen Robotics, a team under Hugging Face, launches mini bipedal robot Microduck
Pollen Robotics, a team under Hugging Face, has launched the mini bipedal robot Microduck, with its software stack and reinforcement learning training code open-sourced under Apache 2.0. The robot is available for pre-sale at $399 (excluding tax and shipping). The company emphasizes that the open-source scope is limited to software, and mechanical and circuit design files have not been released, thus it cannot be classified as open-source hardware.
The robot stands 25 cm tall and weighs approximately 800 grams (the store lists it as 780 grams), featuring 15 servos covering the legs, head, and neck. It is equipped with a Rockchip RK3566, an on-chip accelerator, 1GB of RAM, and 32GB of storage. Sensors include a front camera, an 8×8 time-of-flight LiDAR, one inertial measurement unit each for the body and head, a microphone, and a speaker, with NFC on both the head and beak. The detachable NP-F550 battery provides about one hour of runtime. It comes with seven retrainable strategies out of the box, with a control loop at 50 Hz.
The physical robot can walk using a controller, stand up after falling, pick up objects with its movable beak, sit down, squat, kick a ball, and roll forward. By adding wheels to the feet, it can switch to another strategy to glide, turn, go downhill, and stand up while in skating mode. Each unit generates and binds a lifetime voiceprint upon first boot, communicating with a short sound resembling biological rather than assistant-like calls, and can also follow a laser pointer and react to the environment. The training environment microduck_rl is based on MuJoCo/MuJoCo Warp with PPO, exporting to ONNX after 50 Hz training; domain randomization includes motor friction, battery voltage, control delay, and servo backlash. Without a local GPU, remote training can be done via Hugging Face Jobs.
During operation, it is split into independent daemons using Rust: robotd is responsible for 50 Hz control, strategies, and safety gates, while updaterd handles signature verification and failure rollback for over-the-air upgrades. Other services include configuration, Bluetooth, controller, camera audio, WebRTC, and depth sensors, communicating via Unix Socket and JSON-RPC 2.0. Only robotd can drive the motors; phones, controllers, and custom programs only send action intentions. The upgrade path involves downloading the release package, verifying signatures, switching versions, rebooting, and health checks, with automatic rollback to the previous version in case of exceptions to avoid bricking. The main repository has about 1200 stars.
Pollen was founded in France in 2016 by former Inria researchers and will merge into Hugging Face's robotics team in April 2025, having previously launched the desktop interactive robot Reachy Mini. Microduck is the second consumer-oriented product, shifting its positioning from "interactive desktop pet" to "a reinforcement learning platform that can simulate to real machines on the desktop." It comes in four colors: cream, graphite, lavender, and sky, with pre-sales starting on August 27. The initial plan was to deliver the first batch in North America, Europe, and the UK before Christmas; however, the store now indicates that there are too many community orders, and new year goods cannot be guaranteed to arrive by Christmas, with new orders estimated to take 4 to 6 months, with earlier orders prioritized. Camera resolution, radar range, wireless versions, and software development kit languages are still marked as tentative.
Buyers are developers, classrooms, and enthusiasts looking to practice bipedal control, simulate to real machines, and develop embodied strategies, exchanging $399 for a physical closed-loop system that can walk, fall, and retrain. The seller is Hugging Face: transferring the collaborative model from the model center to physical products, with software open-sourced to encourage community strategy feedback, while hardware remains closed-source to protect BOM and gross margins. Funding shifts from cloud inference subscriptions to desktop robot pre-sales; beneficiaries include developers who can exchange ONNX strategies on the Hub, as well as the RK3566 supply chain. The pressure point remains on educational bipedal platforms still priced in the thousands of dollars, and buyers who interpret "open-source" as being able to build entire machines—since the designs are not public, what is replicated is the strategy, not the whole duck. The event-driven aspect is clear: after the open-source model war over weights, the next battle is for affordable real machine entry.
Source: Public information
ABAB AI Insight
Pollen's path is the commercialization of laboratory machines, not a toy factory suddenly understanding reinforcement learning. The Inria background team first developed the Reachy series of open-source humanoid upper bodies, selling to Hugging Face in 2025, exchanging for distribution channels and the habit of "models as community": weights can be forked, and strategies must also be forked. Microduck compresses this logic to 25 cm—actions are not hardcoded into a gait library but are seven retrainable strategies, with training that can be offloaded to Hugging Face Jobs. This aligns with the company's earlier efforts to turn LeRobot into an open-source robot dataset and model library: first occupying the developer workflow, then selling boxes that can run these workflows.
Investment is directed at the last centimeter from simulation to real machines. The $399 covers servos, batteries, cameras, and a board capable of running ONNX at 50 Hz, not complete open-source hardware. Closed-source structural components mean that BOM, mold, and channels are locked by the company, while the open-source aspect is the strategy files that can compound on the Hub. The motivation is to bring the experimental costs of "physical AI" down from laboratory machine time to consumer electronics pre-sales: developers use their own cards or platform computing power to produce new skills, which flow back to the platform, while hardware sales absorb cash. Reachy Mini focuses on interaction, while Microduck emphasizes movement, with product lines segmented by scenarios to avoid a single machine taking on the entire narrative of both a desktop pet and bipedal research.
The benchmarks are clear: Boston Dynamics' Spot sells industrial reliability, Unitree brings quadrupeds and humanoids into a price range, while Hugging Face's path is "model community + desktop real machine," closer to the early TurtleBot for ROS and Donkey Car for autonomous driving interest groups—not the strongest movement, but a replicable experimental unit. The industry phase has shifted from laboratory demonstrations to control: whoever can randomize the simulation domain down to backlash and voltage can turn strategies into transportable goods. Closed-source mechanics combined with open-source brains represent the current mainstream compromise in robotics, with more expensive reference bipeds like HopeJR remaining in another tier.
Structural changes reflect the shift in pricing power under the overlay of technological substitution. Substitution occurs in controllers: traditional gait engineering is replaced by PPO and domain randomization, with what engineers deliver shifting from code trajectories to exportable ONNX. Pricing power shifts from "who has the laboratory" to "who can connect the $399 machine with cloud training tasks." The mechanism is distribution: once a model platform has a base of real machine installations, the marginal distribution cost of new skills approaches zero, making hardware a reader for the strategy market. The Christmas season has been filled with pre-sales, with delivery times pushed back to four to six months, indicating that demand has first hit the channels, while production capacity and open-source hardware expectations have been intentionally left on the company's books.
ABAB News · Cognitive Law
The open-source aspect is a replicable brain, while the deliverable body is locked down.
If strategies can be forked, pricing power shifts from the laboratory to the installation base.
Affordable real machines are readers, while community weights are deposits.