Hugging Face unveiled Microduck on 27 August, a $400 bipedal robot about 25 centimetres tall, due to ship before year end. It walks, it can be fitted with roller skates, and new movements are trained by reinforcement learning in a physics simulation on a laptop or on Hugging Face’s own compute, then transferred to the machine in one step. The kit, simulator and training stack are on GitHub.
The robot comes out of Pollen Robotics, the French company Hugging Face bought last year and now runs as its hardware arm, the group behind its Reachy desktop robots. Hugging Face has consistently framed its robotics work against the closed and expensive end of the field. Chief executive Clem Delangue argues that keeping robots open and cheap is what stops the category being run by a few players shipping black box systems.
The contrast is deliberate. Tesla’s Optimus is quoted above $20,000, 1X’s NEO the same, and most humanoid programmes run on private rounds in the billions. A $400 toy that learns to skate is not competing with any of that on capability, but it puts a working robotics learning loop in front of hobbyists and classrooms for the price of a mid range phone, which is how Hugging Face built its position in machine learning.
The timing is the part left hanging. Microduck arrived in the same week The Information reported that Nvidia had agreed to buy Hugging Face for $12.9 billion, a deal neither company has confirmed. Nvidia has said it wants to be the default platform for general purpose robotics. What a deliberately cheap open hardware line looks like inside the company it was positioned against is a question the announcement did not touch.
