What Microduck Is
Microduck is a bipedal robot that you teach by reinforcement learning instead of by scripting its motions. In the guides we publish here, the machines worth covering past launch week are the ones a person can change.
Hugging Face and Pollen Robotics opened pre-orders on 27 August 2026 at $399, with first deliveries targeted before Christmas. Pollen Robotics designed the machine. Hugging Face acquired Pollen Robotics in April 2025 and now sells the robot through its own store. You can check the current price and stock on the Microduck product page before ordering, because pre-order pricing and ship dates move.
The robot weighs under 800 g and stands about as tall as a stack of two paperbacks. It has one camera eye, an articulated beak that can grip small objects, and 15 motors spread across its legs, neck, and head.
| Spec | Published figure | What it means for you |
|---|---|---|
| Height and width | 25 cm tall, 14 cm wide | Small enough to run on a desk, tall enough to fall over properly |
| Weight | Under 800 g | Light enough that a fall onto a hard floor rarely ends the session |
| Actuation | 15 motors plus an articulated beak | Enough joints for a real gait, and a beak that grips small objects |
| Vision and range | Front camera, compact lidar, 8x8 time-of-flight matrix | The robot can see obstacles and measure how far away they are |
| Motion sensing | Two IMUs, one in the body and one in the head | Balance data the training loop needs to keep the gait upright |
| Audio | Microphones and a speaker | Voice input and sound output without adding hardware |
| Compute | Rockchip RK3566 with an AI accelerator, 1 GB RAM, 32 GB storage | Trained policies run onboard, so no laptop tether at runtime |
| Power | Removable NP-F550 battery, 2600 mAh, about one hour | A spare battery doubles your session before a recharge |
| Connectivity | Wi-Fi, Bluetooth, two NFC antennas | Deploy new policies over the network rather than by cable |
The Seven Skills Microduck Ships With
Microduck arrives with seven trained behaviours already loaded, so it does something useful on the first evening. The box holds the robot, a battery, a USB-C cable, and a game controller. Each one is a neural policy trained in simulation and exported to the hardware. None of them is a hand-written animation.
Roller-skating is the odd one out and the most revealing. Wheels change the balance problem completely, so shipping a skating policy alongside a walking one shows the training stack generalises past a single gait. The rollers themselves come in the $39 Accessory pack, not in the box.
Self-recovery is the one worth pausing on. A small biped that cannot stand back up turns every fall into a trip across the room, and that single behaviour makes an untethered training session practical.
- Walk, the base locomotion policy.
- Sit, the rest position it settles into.
- Stand, returning from sitting to the walking posture.
- Kick, a single-leg balance behaviour.
- Grab, using the articulated beak.
- Roller-skate, a second locomotion mode that needs the rollers from the $39 Accessory pack.
- Self-recovery, standing back up after a fall.
How Training Microduck Works
Training happens in a physics simulator on your own machine, and only the finished policy goes to the robot. The published stack uses mjlab, built on MuJoCo Warp. The learning algorithm is Proximal Policy Optimization, usually shortened to PPO.
The reported figure is roughly one to two hours on a CUDA graphics card to reach a usable gait, running 4,096 simulated robots in parallel. That assumes you own one, or rent one. Pollen Robotics says training runs on your machine or on Hugging Face Jobs, so a local card is the cheap path rather than the only one.
The trained policy exports to ONNX and runs on the robot under a Rust runtime driving a 50 Hz control loop. Fifty corrections per second keep a small biped upright, so the policy has to run onboard rather than stream from your laptop.
New to reinforcement learning? Our guide to reinforcement-learning robots covers the loop in plain terms. Our LeRobot guide covers the teach-by-demonstration approach Hugging Face publishes alongside it.
What Is Open and What Is Not
The software is Apache-2.0 licensed and the mechanical and electronic designs are not published under an open licence. That distinction decides whether you can repair and fork the machine or only reprogram it.
In practice you get three things on GitHub: the SDK, the simulation environment, and the full reinforcement-learning training stack. That is enough to fine-tune a policy, retrain from scratch, and redeploy it. You do not get the CAD to print a replacement shin. You do not get the board files to respin the electronics.
For buyers who care about repairability more than programmability, that gap matters. Pollen Robotics sells spare motors in a $119 Dev pack, and publishes no price for a replacement leg or shell. Our open source robots guide covers where the fully open projects sit, including the ones that publish their bill of materials.
What Hugging Face Has Not Published Yet
Several things a buyer needs to know are still unpublished three days after the announcement. Name them rather than guess. Each one changes how the robot fits a given desk or household.
The launch material does not state which floor surfaces the walking policy was trained on. A gait tuned on flat lab flooring behaves differently on deep carpet. New owners usually discover that first. If your floors are carpeted, ask about training surfaces in the Pollen Robotics Discord before you order.
Noise level, shipping regions, warranty terms, and the price of a structural spare are also absent from the announcement. Pollen Robotics does sell spares: a $39 Charger pack, a $119 Dev pack with three spare motors and a screw pack, and a $39 Accessory pack. What it has not published is a price for a replacement leg or shell, which is the part a falling robot breaks most. Ask Pollen Robotics directly through the Microduck product page before ordering if any of these decide it for you.
What we see readers get wrong most often after a launch is treating an unlisted spec as a missing feature. Usually the figure just was not measured for the press kit. Read silence as silence.
When Reinforcement Learning Is the Wrong Tool Here
Reinforcement learning suits continuous control problems like balance and gait, and it is a poor fit for tasks with a single correct sequence. Microduck sits squarely in the first category, which is why walking and self-recovery are the headline behaviours.
Take a fixed routine, such as a repeatable pick-and-place at set coordinates. Training a policy for that is slower and less reliable than writing the motion directly. Every hour in the simulator buys generalisation you did not need.
The other failure mode is reward design. A policy learns exactly what the reward function measures, so a badly shaped reward produces a robot that shuffles, or hops, or drags one leg while scoring well. Reward debugging takes more hours than the one-to-two-hour training run does.
Manipulation is the third boundary. A single articulated beak is a gripper with one degree of freedom, so object handling stays limited to small, light items. Builders who want real manipulation should look at the arm-based route in our SO-101 robot arm guide.
Who Microduck Is Not For
Microduck is a poor fit for anyone who wants a robot that does chores. It does not vacuum, fetch, monitor a room, or answer household questions, and nothing in the shipped skill set points that way. Buyers after a working home assistant should read our home robot assistant guide instead.
It is also a poor fit if you have no CUDA graphics card and no wish to pay for hosted runs. Pollen Robotics points at Hugging Face Jobs for that, and without either route you are left with seven pre-trained tricks.
Children are the third mismatch. A 15-motor biped with an exposed training workflow is a developer tool. Families shopping for a play robot get more from the picks in our robots for kids guide.
What would change our answer is the hardware licence and the parts supply. Publish the mechanical files and sell spare legs, and Microduck stops being a sealed product with open software. It would then beat most DIY builds on time to a first walk.
Our Read on the $399 Price
At $399 Microduck is priced against DIY biped kits rather than against toys. The closest comparison is Open Duck Mini v2, a fully open project targeting a bill of materials under $400, which reaches a similar place through self-assembly and sourcing.
Paying the $399 buys assembly, calibration, a supported battery, and a training stack someone else already made work on this exact hardware. Sourcing your own parts saves nothing on paper and costs a build.
We walk through the full figure, including the spare battery and the graphics-card assumption, on our Microduck price breakdown. To decide between the two Hugging Face robots, start with Microduck versus Reachy Mini.
Bottom Line
Microduck is a $399 developer robot whose real product is the training loop rather than the seven tricks it arrives with. Buy it if you want to train your own locomotion policies on hardware that already works. Skip it if you want a household helper, have no graphics card, or need a machine you can repair from published files.
Compare the full cost on our Microduck price breakdown, then check Microduck against Reachy Mini before you commit to a pre-order.
FAQs
How much does Microduck cost?
Microduck costs $399 as published at launch on 27 August 2026. Pre-orders opened the same day and first deliveries are targeted before Christmas. Pre-order pricing and ship dates can change after launch.
What can Microduck do out of the box?
Microduck ships with seven trained behaviours: walk, sit, stand, kick, grab with its beak, roller-skate, and stand back up after a fall. Each is a neural policy trained in simulation rather than a scripted animation. Roller-skating needs the rollers from the $39 Accessory pack.
Is Microduck open source?
The software is Apache-2.0 licensed, including the SDK, the simulator, and the reinforcement-learning training stack. The mechanical and electronic designs are not published under an open licence, so you can reprogram the robot but you cannot fabricate replacement parts from official files.
Do you need a GPU to train Microduck?
Training a new walking policy assumes a CUDA graphics card. The reported time is roughly one to two hours for a usable gait at 4,096 parallel simulated environments. Without a suitable card, you are limited to the seven behaviours that ship on the robot or to renting cloud compute.
How long does the Microduck battery last?
Microduck uses a removable NP-F550 battery rated at 2600 mAh, good for about an hour of use. Because the pack is removable and the NP-F550 is a common camera battery format, a charged spare extends a session without waiting for a recharge.
Does Microduck arrive assembled?
Hugging Face has not published an assembly requirement for Microduck, and the launch material describes it as a finished product rather than a kit. Its sibling Reachy Mini is explicitly sold as a kit that takes two to three hours to build, so a ready-to-run Microduck is an assumption worth testing before ordering.
How tall is Microduck?
Microduck is 25 cm tall and 14 cm wide, and weighs under 800 g. It is built to walk on a floor and get back up after falling rather than to sit on a desk.