Microduck biped robot opens pre-orders at $399
Pollen Robotics has opened pre-orders for Microduck, a 25 cm, 800 g open-source biped robot whose movements are trained through reinforcement learning rather than shipped as fixed programming. It is described as playable straight out of the box, and it comes with seven pre-trained movement policies, including walking with a velocity-tracking gait, sitting down and standing back up on its own, a one-shot kick that returns straight to walking, a beak-scoop grab motion, roller skating when skates are attached, and righting itself from flat on its back.
Every one of those behaviors is a policy the owner can retrain. The advertised workflow has four steps: train a behavior in physics simulation, on the owner's own machine or on Hugging Face Jobs; deploy it to the physical robot, described as one step from simulation to the real thing; refine the simulation and retrain; then publish the resulting policy for others to use. All training happens in the MuJoCo physics simulator, and the page describes a simulated twin of the robot that behaviors are trained on before being deployed to the real hardware.
The robot's hardware includes 15 motors, a camera, LiDAR, and two IMUs, with an onboard policy control loop running at 50 Hz. It ships in four colorways, and Pollen Robotics sells optional packs of play gear and spare parts on top of the base robot, which the page says is everything a buyer needs on day one.
The full software stack, the SDK, the simulation, and the RL training code, is open source on GitHub at pollen-robotics/microduck, under the Apache-2.0 license; the page states that what the robot runs is exactly what buyers can read, fork, and retrain. Robot control is shown through a command-line tool, robotctl, with commands to monitor, configure, and update the robot.
Pre-orders are open now at $399, an introductory price before taxes and shipping, with delivery promised before Christmas 2026. Pollen Robotics points buyers to a Discord community described as a place to show builds, swap trained policies, and get help when something goes wrong.
Key facts
- Pollen Robotics has opened pre-orders for Microduck, a 25 cm, 800 g open-source biped robot trained through reinforcement learning, at an introductory price of $399 before taxes and shipping.
- The robot ships with 15 motors, a camera, LiDAR, and two IMUs, plus an onboard policy control loop running at 50 Hz.
- Seven pre-trained movement policies come in the box, including walking, sitting and standing back up, kicking, grabbing, roller skating with skates attached, and righting itself from its back.
- Every policy is trained in the MuJoCo physics simulator, on the owner's own machine or on Hugging Face Jobs, then deployed to the robot and can be published for others to use.
- The SDK, simulation, and full RL training stack are open source under the Apache-2.0 license on GitHub at pollen-robotics/microduck; pre-orders ship in four colorways before Christmas 2026.
Why it matters
Microduck is a 25 cm, 800 g biped whose behaviors are trained through reinforcement learning rather than fixed in firmware, and Pollen Robotics has released its entire supporting stack, the SDK, the MuJoCo simulation, and the RL training code, open source under Apache-2.0 on GitHub. None of the seven shipped policies, including walking, sitting and standing, kicking, grabbing, roller skating, and self-righting, are permanent: the page's own workflow lets an owner retrain any of them on their own machine or on Hugging Face Jobs, deploy the result to the physical robot, and publish it for others to reuse.
Who it affects
The product page speaks directly to buyers who want to retrain the robot themselves, describing policies trained on the owner's own machine, and it offers Hugging Face Jobs as an alternative for anyone without local compute for the simulation step. Because the SDK, simulation, and training stack are open source under Apache-2.0 on GitHub, developers and researchers already working with reinforcement learning or MuJoCo-based simulation gain an inexpensive physical robot to test trained policies on, not just a simulated one. Pollen Robotics also points buyers to a Discord community described as a place to show builds, swap trained policies, and get help when a leg does something strange. Pre-orders are open now for delivery before Christmas 2026, so early buyers are the first to choose among the four colorways and add optional packs of play gear and spare parts.
How to use it
Pre-orders are open now at $399, described as an introductory price before taxes and shipping, for delivery before Christmas 2026. The robot ships in four colorways, and buyers can add optional packs of play gear and spare parts on top of the base robot, which the page says is complete on its own. Robot control runs through a command-line tool, robotctl, shown running monitor, configure, and update commands. Because the SDK, the MuJoCo simulation, and the RL training stack are released under the Apache-2.0 license on GitHub at pollen-robotics/microduck, buyers can read, fork, and modify what actually runs on the hardware: the intended loop is to train a policy in simulation, deploy it to the robot, refine and retrain, then publish the resulting policy for the Discord community to use. The onboard policy control loop runs at 50 Hz.
How solid is it
This account rests entirely on Pollen Robotics' own product page for Microduck, a pre-order page rather than an independent review or third-party test. Several of its concrete claims are independently checkable: the SDK, simulation, and RL training stack are published at pollen-robotics/microduck on GitHub under the Apache-2.0 license, so the code and its terms can be inspected directly rather than taken on faith, and the page also references a separate press kit with a full technical specification sheet. The source names no individual person, engineer, or founder behind the product; the maker is identified only by the pollen-robotics domain and GitHub handle, without a fuller company description.
Risks and caveats
The $399 figure is explicitly an introductory price that excludes taxes and shipping, and the source states no dollar amount for either, so the final delivered price is not knowable from this page. It also names no country or region restrictions for shipping. The retraining workflow depends on the owner supplying compute, either their own machine or a Hugging Face Jobs account, and the page does not say what hardware that requires. No battery life, charging time, or runtime-per-charge figures are given, so how long the robot runs between charges is unknown. What is confirmed is that pre-orders are for hardware that has not shipped yet: delivery is promised before Christmas 2026, not on order.
“The SDK, the simulation and the full RL training stack are on GitHub. What the robot runs is what you can read, fork and retrain.”
— the Microduck product page