Microduck Is a $399 Open Source Robot You Can Retrain

Most robots you can buy for under $500 are demonstrations. They run the behaviours the manufacturer shipped and nothing else, because the training stack that produced those behaviours stays in-house. Microduck, announced on August 27, 2026 by Hugging Face and its Pollen Robotics subsidiary, ships the training stack in the box — and that is the whole point of it.

• Microduck costs $399, stands 25cm tall, weighs 780 grams, and ships before Christmas
• It carries 15 motors, a wide-angle camera, LiDAR, microphones, a speaker, two IMUs, NFC, Wi-Fi and Bluetooth
• Seven pre-trained behaviours ship on the device: walking, sitting and standing, kicking, grabbing, roller skating, and recovering from a fall
• The SDK, the MuJoCo simulation environment and the full reinforcement learning training stack are Apache 2.0 licensed

What the $399 Hardware Actually Includes

For a robot this size, the sensor list is the surprise. Fifteen motors is enough for genuinely articulated biped locomotion rather than a shuffling approximation, and the articulated beak doubles as a gripper rated to pick up objects weighing up to 800 grams — slightly more than the robot itself.

Perception comes from a wide-angle camera, LiDAR, and two inertial measurement units, with microphones and a speaker for audio interaction. NFC, Wi-Fi and Bluetooth handle connectivity. At 780 grams and 25cm, it is small enough to live on a desk and robust enough to fall over, which it is explicitly trained to recover from.

That last detail matters more than it sounds. A robot that cannot get up on its own needs supervision, and a robot that needs supervision never gets left alone long enough to learn anything interesting.

Why the Apache 2.0 Training Stack Is the Real Product

The seven behaviours Microduck ships with can be inspected and retrained. Hugging Face has published the SDK, a MuJoCo simulation environment, and the reinforcement learning training stack on GitHub under Apache 2.0, so the workflow is: train a policy in simulation, deploy it to the physical robot, then fine-tune on hardware.

CEO Clem Delangue described it as an open-source robot you can teach new tricks with reinforcement learning, and the licensing choice backs that up. Apache 2.0 permits commercial use, which puts Microduck in reach of classrooms, research groups, and small companies that cannot build a sim-to-real pipeline from scratch.

This is the same strategy that produced Reachy Mini, the $299 Raspberry Pi CM4 desktop robot in 2026, and it echoes the broader open robotics push behind NVIDIA's GR00T 1.7 and LeRobot. Microduck extends the line from a stationary desktop unit into something that moves around a room.

How Does Microduck Compare to Reachy Mini?

Different jobs. Reachy Mini is a desktop companion — expressive head movement, camera, audio, sitting in one place. Microduck is mobile, bipedal, and built around locomotion policies you are expected to modify.

The price gap is small enough that the choice comes down to what you want to study. If the interesting problem is conversation, vision, or desktop interaction, the stationary unit is simpler. If the interesting problem is balance, gait, manipulation, or sim-to-real transfer, a walking robot with a gripper and an open RL stack is the one that teaches you something. Pollen Robotics, the French startup Hugging Face acquired in 2025, built both.

For anyone assembling a small-hardware bench, the rest of our compact computing coverage lives on the mini computers page, and our edge AI dev board buyer's guide covers the compute side of the same problem.

Pre-orders opened on announcement day with shipping promised before Christmas. As always with a first production run of a new form factor, the specs are the manufacturer's until independent units are in hand.

Sources: TechCrunch — August 27, 2026; Hackster.io — August 27, 2026; Pollen Robotics — August 27, 2026.