Microduck brings robot learning to the desktop for $399

What this is about

Pollen Robotics, a Hugging Face company, began taking pre-orders for Microduck on August 27, 2026. The 25-centimetre biped costs $399 at launch before taxes and shipping. It is aimed at learners, developers and research teams that want to test motion policies on real hardware, not at factories.

The interesting part is not only the duck shape. Microduck links a physics simulator to a physical robot and publishes the software, simulation and training code under Apache 2.0. That makes part of robotics research more accessible than platforms that require expensive laboratory hardware.

What Microduck actually does

Microduck has 15 motors, a camera, a small LiDAR sensor and two inertial measurement units. The manufacturer says it weighs 800 grams and runs its motion policy at 50 hertz. Included behaviours let it walk, sit, stand, kick, pick up objects with its beak and recover after falling.

New motions are first trained in MuJoCo. A learning algorithm can try many virtual versions of the robot in parallel. The trained policy is then transferred to the physical machine. Developers can refine the digital twin, retrain the policy and publish their results.

The public GitHub repository contains the SDK, simulator and full reinforcement-learning stack. That does not make Microduck effortless to use, but it makes the system inspectable and modifiable.

Why it matters

Robot learning has a basic constraint: physical machines are expensive, slow and vulnerable to damage during training. Simulation reduces that risk, but transferring a policy into the real world remains difficult. Friction, motor variation and imperfect sensors do not behave exactly as their models do.

A small and comparatively affordable platform lets more people study that transfer. University groups can build courses around physical hardware. Developers can test whether a motion learned through thousands of simulated trials still works on a desk. Open code also makes it easier to inspect the assumptions embedded in training.

The price matters, but it is not the total cost. Accessories, replacement motors, shipping, taxes and computing may add expense. Microduck is a learning and research device, not a cheap substitute for an industrial robot.

In plain language

Training is like using a driving simulator. A learner can practise a thousand corners without damaging a real car. Only after the basics work does the learner go onto the road. Microduck does the same for walking, standing and grasping, except that its “road” is a desk.

A practical example

A university course with twelve students wants to study walking on smooth and rough surfaces. The group first creates 4,096 virtual environments with slightly different friction values. After training, it transfers the policy to one Microduck and records 100 physical walking trials.

If the robot slips on wood in 18 of 100 trials, the students adjust simulated friction and the reward function. They train again and compare the next series. The lesson comes from measuring the gap between simulation and reality, not from replaying a prepared demonstration.

Scope and limits

First, many specifications and demonstrations come from the manufacturer. Independent long-term tests of durability, battery life and replacement-part support were not available at launch.

Second, an 800-gram robot with a beak does not represent every physical task. Results cannot be transferred directly to heavy arms, vehicles or humanoid systems.

Third, custom training requires knowledge of Python, simulation and reinforcement learning. Buyers expecting only a finished toy may never use the platform's most important open component.

SEO & GEO keywords

Microduck, Pollen Robotics, Hugging Face, open-source robotics, reinforcement learning, MuJoCo, sim-to-real, biped robot, robotics education, GitHub