The official Microduck Sandbox is hosted by pollen-robotics on Hugging Face. A simulator helps you explore behavior, but a virtual result is not a hardware reliability test.
Open the official sandbox
Use the direct link below to the manufacturer’s Hugging Face Space. Loading time and availability belong to that external service. If the Space is waking up or unavailable, return later or use the training repository’s documented local viewer workflow. This café does not present an animation as a physical simulation and does not claim to run the official simulator itself.
A three-minute observation exercise
First, watch the neutral pose before changing anything. Next, try one available control at a time and describe what changes: the direction of the body, the placement of the feet or the orientation of the head. Finally, repeat the same input. Your goal is not to find the most dramatic move; it is to recognize the connection between a command and a repeatable response.
Write down what you cannot infer
A simulation lets you observe a model in a modeled environment. It cannot by itself establish the real robot’s battery life, long-term wear, noise or success on your floor. Even convincing virtual motion depends on assumptions about contact and actuators. This is why the next question is about transfer: which assumptions remain accurate on the physical system?
When to move beyond the browser
Move to the training documentation when you want to change how a behavior is learned rather than simply observe it. Check prerequisites before installing a large toolchain. Keep your first goal small and reproducible. You can learn a great deal by comparing one baseline with one controlled change, especially if you record both a successful run and a failure.
Go to the source
Reviewed September 7, 2026. Prices, availability and software can change.