Robots you can actually buy have always come with a catch: they cost more than a used car, or they do nothing but sit on your desk and blink. Hugging Face just split the difference with a duck.

The number that matters is $399. That's the pre-order price of the Microduck, a 25-centimeter open-source biped that waddles, crouches, picks things up with its beak, falls over, and gets back up on its own (TechCrunch). For context, the cheapest humanoid robot most people can order today, Unitree's G1, starts at $13,500 (Theresarobotforthat). The Microduck costs less than 3% of that.

It ships before Christmas. It roller skates. And the software that makes it move is on GitHub for anyone to fork.

The pitch is simple: a walking robot with a real training stack, priced like a mid-range drone.

🧠 Why This Matters

Legged robots are hard in a way that stationary arms aren't. Balance, falling, and recovery are the problems that eat entire research budgets. Until now, if you wanted to touch that problem yourself, your options were a five-figure humanoid or a simulation running on your laptop with nothing physical at the end of it.

The Microduck drops the price of the physical part to $399 and hands you the whole pipeline. Pollen Robotics, the Bordeaux team Hugging Face acquired in 2025, open-sourced the runtime and the reinforcement-learning training code under an Apache-2.0 license (MarkTechPost). You can train a new gait in simulation, then push it straight onto the hardware sitting on your desk.

"An open-source robot you can teach new tricks with reinforcement learning." — Clem Delangue, CEO, Hugging Face (TechCrunch)

That last sentence is the whole game. Cheap toys don't let you retrain them. Research platforms don't cost less than an iPhone. This does both.

📊 Deep Dive

Inside a body under 800 grams is a surprising amount of hardware. The Microduck runs 15 motors across its legs, neck, head, and beak, and carries a front camera, an 8×8 time-of-flight LiDAR matrix, two inertial measurement units, microphones, and a speaker (Engadget). A Rockchip processor with an AI accelerator handles the on-device work, backed by 1GB of RAM and 32GB of storage. A removable 2,600mAh battery gives it roughly an hour of waddling before it needs a swap.

The training side is where the price tag stops making sense in a good way. Policies are trained with PPO inside a MuJoCo-based simulator, spinning up 4,096 parallel environments and producing a usable walking gait in one to two hours on a single GPU (MarkTechPost). Sim-to-real transfer uses actuator modeling with randomized voltage, delay, and friction so the policy survives contact with the real world.

Here's how it lands next to the robots you might compare it to:

  • Microduck — $399, 25cm biped, walks and self-recovers, fully open-source software
  • Reachy Mini — $499 ($399 for the Lite), Hugging Face's earlier desk robot that stays put; it has already shipped 10,000+ units (MarkTechPost)
  • Unitree G1 — from $13,500, full humanoid, closed development kit at the education tier (Theresarobotforthat)

The Reachy Mini comparison is the telling one. Same company, same rough price, opposite philosophy. The Mini was built to sit on a desk and interact. The Microduck was built to leave the desk, fall down, and figure out how to stand up again.

⚠️ The Catch

Open-source has an asterisk here. The software is Apache-2.0 and public, but the mechanical and electronic design files stay closed (MarkTechPost). You can retrain the brain all you want; you can't fab your own duck. So "open" describes the intelligence, not the hardware.

Then there's the runtime. One hour on a charge is fine for tinkering and terrible for anything sustained, which is why the battery pops out for swaps. And a 25cm duck with an articulated beak is a genuine research and education tool, not a household helper. It picks up socks in demos; it is not folding your laundry.

The last catch is the oldest one in hardware: a pre-order is a promise. "Ships before Christmas" (Axios) is a date every robotics startup has missed at least once.

🎯 What Happens Next

Watch the fork count. The entire bet is that a cheap physical platform plus an open training stack pulls in developers who were priced out of legged robotics. If the GitHub repos light up with community gaits and behaviors over the next few months, that's the signal the strategy is working.

Watch the deliveries too. Reachy Mini's 10,000-plus units prove Hugging Face can actually ship consumer hardware, not just publish models. If the Microduck moves in comparable numbers before the holidays, expect a second, larger, more capable body to follow the same playbook.

🧩 Bigger Picture

Hugging Face built its name giving away model weights and letting a community do the rest. The Microduck is that instinct pointed at atoms instead of tokens: put a capable, retrainable body in as many hands as possible and let the ecosystem discover what it's for.

The economics are the story. When the frontier of physical AI runs $13,500 and up, robotics stays a lab sport. Drop a real training platform to $399 and the pool of people who can experiment grows by orders of magnitude. That's the same curve that turned machine learning from a PhD pursuit into something a teenager runs on a laptop.

A duck won't change the industry by itself. A hundred thousand people teaching one to walk might.

The cheapest way into humanoid robotics used to cost as much as a car. Now it fits on your desk, waddles, and asks you to teach it a trick.

Sources