A $399 Robot That Teaches Itself to Walk

A tiny robot that learns on its own

Pollen Robotics — now part of Hugging Face — just put a small bipedal robot called Microduck up for pre-order. It's 25 centimetres tall, weighs less than a water bottle, has 15 motors, a camera, a depth sensor, and a little articulated beak that can pick things up. Price: $399. Delivery target: before Christmas 2026.

What makes it genuinely interesting isn't the hardware. It's how it learns. Instead of someone writing out every movement step by step, Microduck is trained by running millions of simulated attempts until it figures out how to walk, crouch, recover from a fall, even roller-skate. The whole recipe for doing that training is free and open for anyone to use or modify.

Hugging Face's plan is to let people share the movement patterns they train — the same way a chef might share a recipe — so the whole community gets smarter together.

For most business owners this is still something to watch rather than buy today. But the signal matters: physical AI — machines that learn by doing rather than being programmed — is becoming cheap enough to experiment with. The gap between software agents and robots you can touch is closing faster than most people expect.

Words worth knowing

Hugging Face — A platform where researchers share AI models, a bit like GitHub but for trained AI brains. They acquired Pollen Robotics in 2025.

Open-source — The plans, code, and training recipes are public. Anyone can read them, copy them, or improve them. The opposite of a locked black box.

Neural policy — Instead of a robot being told "lift left foot, then right foot," it develops its own internal sense of balance through practice — like a toddler learning to walk, but in a computer simulation running at high speed.

Simulation-to-real — Training a robot inside a virtual world (cheaper, faster, no broken parts) and then transferring what it learned into the physical one.