Microduck: how a little robot learns to walk in a virtual world

A small robot and a visible learning process

Microduck is a two-legged robot from Pollen Robotics, Hugging Face’s robotics team. It was introduced on 27 August 2026. At 25 centimetres tall, it fits on a table; an included controller lets people control its movements. The manufacturer presents it as a platform for play and robotics experiments. [1, 4]

The current development is a pair of NVIDIA examples for developers. A tutorial published on 21 October explains how a walking controller learned in simulation can be transferred to the physical robot. A 23 October post adds training in a virtual obstacle course. These examples are neither a new sales launch nor proof that Microduck can already handle any course at home. [2, 3]

Practising in a world made of software

A digital copy of the robot moves inside the simulation. It tries movements and receives a score for outcomes such as making progress and keeping its balance. Many attempts produce a learned controller that selects suitable motor movements. Learning through this feedback is called reinforcement learning. People define which behaviours earn a reward. [2]

Many virtual robots can practise at once. Each attempt does not require setting up a physical device and picking it up after a fall. NVIDIA uses Isaac Lab, a training framework, and Newton for the physics calculations. The learned controller can then be exported and run on physical Microduck hardware. [2, 3]

Built-in movements and personal experiments

Pollen says Microduck can be operated with its controller without writing code. Advertised movements include walking, sitting, standing up and kicking. The robot is also meant to recover from many falls. This is a manufacturer claim, not a guarantee covering every surface or fall. [1, 4]

Training a new movement is more demanding. The NVIDIA tutorial requires a Linux workstation with a supported NVIDIA graphics card for training. The subsequent steps include evaluating and exporting the learned controller and preparing the physical robot. These requirements apply to that training workflow; they are not prerequisites for using the supplied movements with the controller. [3, 4]

Pollen makes the control, simulation and training software available as open source. It explicitly excludes the mechanical and electronic design files from that statement. [7]

Why the physical world remains harder

A movement can work in a virtual environment and still fail on real hardware. Ground friction, motor delays, mechanical play and battery voltage all affect behaviour. The documented training varies such properties deliberately. This is intended to make the learned controller less sensitive to differences, but transfer to hardware still needs checking. [3]

The new obstacle-course example has a clear boundary too. NVIDIA mainly describes building the environment, training and evaluation in simulation. Physical deployment needs further refinement. The example does not establish reliable obstacle avoidance in an arbitrary home. We have not tested a Microduck or run these training workflows ourselves. [2]

Price and delivery to Germany

As of 25 October 2026, Microduck is offered for pre-order at US$399 before taxes and shipping. The listed delivery regions include the European Union, and therefore Germany. The shipping policy says EU orders ship from France, with destination-country VAT and shipping costs calculated at checkout. The dollar price alone does not establish the final cost for a German buyer. [4, 5, 6]

For new orders, the shop now estimates a wait of four to six months and no longer promises Christmas delivery. Older announcements targeted first deliveries before Christmas 2026. The current shop notice is the relevant information for a new order, and it remains an estimate. Pre-order availability does not mean immediate delivery. We have not tested a completed purchase for delivery to Germany. [1, 5]

What makes it interesting

Microduck makes an abstract AI process visible: a movement learned in software can control a physical robot. The published examples offer a concrete way to explore that technology. The demonstrated movements and training workflows do not establish an all-purpose household assistant. [2, 3, 4]

Our interpretation: its appeal is as an experimental platform for understanding the step from virtual practice to physical movement. A purchase decision also needs to account for the technical work, total cost and currently long delivery estimate.