Microduck had reached 270 points on Hacker News when our brief captured it, but the useful number for developers is 50 Hz. That is how often its onboard neural policy drives 15 servos, according to the robot's published software repository. Pollen Robotics is taking preorders for the 25 cm biped at $399 before tax and shipping, with first deliveries targeted before Christmas 2026. The cute shell is doing the attention work. Under it sits a compact, inspectable test bed for moving reinforcement-learning policies from MuJoCo onto a physical machine.
The launch comes from Pollen Robotics, the French robotics company Hugging Face acquired in April 2025. Its Microduck product page describes a robot that can walk, sit, stand, kick, grab, roller-skate and recover after falling. A game controller is included, so those behaviors do not require a training run. The more consequential promise is that owners can retrain each policy, deploy it to the robot and publish the result.
That makes the $399 price more than a consumer-gadget figure. Pollen's larger Reachy Mini currently starts at $399 for a computer-tethered Lite version and $499 for a wireless model, according to TechCrunch's launch report. Microduck puts a self-contained computer, sensors, actuators and battery into the lower price point. It is a preorder rather than a reviewed shipping product, however, and Pollen calls the price introductory.
A small robot with a readable control path
Microduck runs on a Rockchip RK3566 with an AI accelerator, 1 GB of RAM and 32 GB of storage. Its sensors include a wide-angle front camera, an 8 by 8 time-of-flight LiDAR unit and two inertial measurement units. Pollen's store specification also lists microphones, a speaker, two NFC antennas, Wi-Fi and Bluetooth. A removable NP-F550 battery is rated for about one hour. The store gives the weight as 780 grams, while the main launch page rounds it to 800 grams.
The articulated beak is both a gripper and part of the robot's character. Pollen says it can lift up to 800 grams, roughly its own stated weight. That figure is a manufacturer specification, not an independent payload test. Videos published for the launch show the robot collecting light objects, kicking a ball, following a laser pointer and moving on clip-on rollers. The Verge's account also notes that each unit generates an individual voice when first activated and keeps that audio identity.
The software is unusually legible for a small consumer robot. The main repository is a Rust workspace with separate daemons for the control loop, updates, configuration, Bluetooth, gamepad input, camera streaming and depth sensing. They communicate through a JSON-RPC contract over Unix sockets. Pollen says an app, console, controller or user script reaches the robot through the same calls rather than through separate hidden control paths.
The update design deserves attention because a failed update can turn a mobile robot into a desk ornament. Pollen's repository says updaterd installs signed releases, checks the robot's health and rolls back when startup fails. Owners can install, roll back or pin updates. The documentation also exposes branch builds over SSH for developers working on the robot itself. Those mechanisms are in public code today; how well they survive interrupted power, damaged storage or incompatible community policies will need testing on production units.
The training stack is part of the product
Pollen has split execution and training into two repositories. The main Microduck code runs the hardware. A linked microduck_rl project contains the MuJoCo environment, Proximal Policy Optimization training, domain randomization and the export path to ONNX. The robot loads those exported policies for its 50 Hz loop. This division gives developers a concrete route from a simulated behavior to the embedded runtime.
The company describes a four-part cycle: train in simulation, deploy to the robot, adjust the simulation from what happened on hardware, then train and deploy again. Pollen also offers Hugging Face Jobs as an alternative to local compute, though it does not say that cloud training is required. The $119 developer pack includes Hugging Face credit along with three spare motors, motor cables, batteries, a charger, NFC tags, screws and a screwdriver.
Seven trained moves ship with the robot. They include a velocity-tracking walk, sitting and standing, kicking, grabbing, skating and self-recovery. Pollen says every shipped policy will be published and retrainable. This matters because an open SDK alone can leave the behavior that makes a robot useful sealed behind model binaries. Publishing the policy files and their training recipe gives developers something they can reproduce, inspect and alter.
The claim still has boundaries. The launch materials call the whole software stack Apache-2.0, and the main repository reports that license. They do not establish that the mechanical CAD, circuit-board designs or every hardware component are released under an open license. Calling Microduck an open-source robot is reasonable shorthand for Pollen's marketing, but the evidence available at launch is strongest for the SDK, runtime, simulator and reinforcement-learning stack.
There is another practical limit: simulation transfer is never automatic merely because the command path is short. Contact forces, motor wear, floor friction, battery voltage and small assembly differences can change a learned gait. Pollen's use of domain randomization is meant to expose a policy to variations during training. Its own workflow still includes refining the simulator after a real deployment, which acknowledges that the first simulated result may need work.
What $399 includes, and what it does not
The base box contains the robot, one battery, a USB-C cable and a game controller. Extra batteries require either the $39 charger pack, which includes two batteries and a dual charger, or the larger developer pack. A separate $39 accessory pack supplies the rollers, ball, laser pointer and NFC items shown around the launch behaviors. Taxes and delivery charges sit outside the advertised $399 price.
Pollen lists initial availability in the United States, Canada, the European Union, the United Kingdom, Norway, Switzerland, Japan and South Korea, with other countries planned later. The product page uses the firmer phrase "ships before Christmas 2026," while the store says first deliveries are "targeted" before Christmas. Buyers should treat the latter as the more precise description of a preorder schedule. Neither page gives production volume or a country-by-country delivery date.
The low price also brings clear computing constraints. One gigabyte of RAM is enough for a focused robot runtime, but it leaves little room for large local vision or language models. The published architecture instead looks built around compact policies and services with narrow jobs. Developers who want speech recognition, richer perception or a conversational layer may need an external computer or network service. Pollen has not claimed that Microduck runs a general-purpose language model onboard.
That distinction keeps the product's scope sensible. The robot is an embodied policy platform with a camera, shallow depth sensing and whole-body motion. Its 8 by 8 LiDAR grid is useful for coarse range information, not detailed 3D reconstruction. Its one-hour battery estimate also sets a ceiling on untethered experiments unless owners keep charged packs nearby. Both constraints are visible in the specifications rather than buried behind an undefined claim of autonomy.
Open code does not settle the privacy question
A camera and microphones on a mobile device raise a different issue from the training workflow. Open code lets an owner examine Pollen's runtime and replace parts of it. It does not guarantee that every future app, community policy or network service will keep recordings local. TechCrunch made the same distinction in its report: software added later could use the sensors and send data to an outside service.
The architecture provides some useful inspection points. mediad handles the camera stream over WebRTC, while configuration and Bluetooth live in their own daemons. Shared JSON-RPC calls should make it easier to trace which client requests a capability. The value depends on documentation remaining current and on shipped binaries matching public source. Pollen has published the design and update machinery, but independent buyers have not yet been able to compare retail hardware with the repositories.
Physical safety is narrower here than with a large robot, though it is not absent. The robot has 15 moving motors, can carry an object in its beak and is sold with software intended for modification. The developer pack's spare motors and cables suggest that repair and experimentation are expected. Pollen's public material does not provide enough evidence to judge motor torque limits, pinch protection or the durability of repeated falls, so those questions should wait for hardware tests.
Microduck is therefore most interesting at the seam between a toy-like product and a reproducible robotics kit. The controller and seven supplied behaviors offer an immediate baseline. Rust services, signed reversible updates, MuJoCo training and ONNX deployment give developers specific pieces to inspect. The 270-point Hacker News response measures community attention, not product quality, and the launch videos measure curated demonstrations, not reliability.
The next evidence should come from shipped units: whether Pollen meets its delivery target, whether the published policies reproduce their advertised movements, how long the battery lasts under continuous walking, and how much work a new sim-to-real behavior needs. Repository activity can show whether outside contributions arrive. Retail teardowns can answer the remaining hardware-license and repair questions. Until then, Microduck is a well-documented preorder with an unusually visible software path, not a finished verdict on affordable physical AI.