Microduck is a 25 cm tall, two-legged robot that stands, walks, kicks, grabs and picks itself back up. It is open source: the Python SDK, the simulator and the reinforcement-learning stack behind every move are published, so you can retrain the robot instead of only driving it.
You will also see the name written as Micro Duck, or misspelled as Microduk, Mikroduck or Microdack. It is the same robot.
What is on the robot
- 25 cm tall and about 800 g, small enough to keep on a desk and carry in one hand
- 15 motors driving the whole body
- A camera, LiDAR and two IMUs for the sensing the walking and balancing policies need
- A 50 Hz onboard policy loop, so movement is generated on the robot rather than streamed from a server
- Seven moves shipped trained: standing up, sitting and standing, walking, turning, kicking, driving and grabbing
What ships in the box
- The Microduck robot, in one of four colourways
- A battery, a USB-C cable and a game controller
- The seven trained moves preinstalled
- Access to the open-source SDK, the simulator and the training stack
Pre-order pricing, the colourways and the optional packs are on the buy page and the pre-order page.
The open-source part
Microduck is not a closed toy with an app. The point of the robot is that you can read the code that makes it move and change it:
- Python SDK - talk to the robot, read its sensors, and command joint positions or whole policies
- Simulator - the same robot in simulation, so a policy can be trained and broken without risking the hardware
- Reinforcement learning stack - the training setup the shipped moves were produced with, ready for your own reward function
Everything lives in the open on GitHub, and the simulation runs in a hosted Hugging Face Space if you want to try it before installing anything. The documentation covers the SDK entry points, the simulator and the training loop.
What you can actually do with it
- Play it out of the box. Seven moves, a controller, and no setup beyond charging the battery.
- Program it. Drive the joints from Python and script your own sequences.
- Retrain it. Take a shipped policy, change the reward, run the training, and see the robot walk differently.
That last step is where Microduck stops being a gadget. Sim-to-real reinforcement learning on a robot you can hold in one hand is the interesting part, and it is the reason the code is public.
Microduck compared with a toy robot
A toy robot plays back animations. Microduck runs learned policies, exposes its sensors and motors, and lets you replace the policy with one of your own. It is closer to a small research platform than to a remote-controlled toy: same size, different expectations.
How to pre-order
Pre-orders are open. Microduck is $399 before taxes and shipping, with the four colourways - Cream, Graphite, Lavender and Sky - and optional Charger, Dev and Accessory packs. Taxes and shipping are calculated at checkout.
- Start on the buy page for the purchase-oriented answers
- Read the pre-order guide for the process in detail
- Or go straight to the pre-order page and pick your pack
Where to go next
- Pre-order MicroDuck - the robot and the packs
- Buy Microduck - price, payment, colourways and buying FAQ
- Documentation - SDK, simulator and training
- GitHub - the source
- Discord - the community
- Updates - what has been shipped so far
