Background: what Hugging Face acquiring Pollen Robotics means

Understanding Microduck’s product decisions is hard without one fact: Pollen Robotics was acquired by Hugging Face in 2025.

What that explains

Policies are .onnx files, not motion tables compiled into firmware. A trained behaviour can therefore be downloaded, swapped and shared the way model weights are — which is exactly Hugging Face Hub’s distribution model. The official framing is making robot behaviours easier for developers to share and reproduce.

Training runs on Hugging Face Jobs. Add --hf-jobs to any microduck_rl training command and it goes to remote compute, no local CUDA GPU required. For anyone without a graphics card that is a real lowering of the barrier — a usable walking policy takes roughly 1–2 hours of GPU time.

The online simulator is hosted on a Hugging Face Space, so people without hardware can try it first.

What it means in practice

If this direction continues, the expectation is that sharing a policy becomes as ordinary as sharing a model. Someone trains a better walking policy; others download the .onnx and swap it in — possible because every policy shares the same 61-dimensional observation contract and they are interchangeable.

Where the evidence ends

The acquisition was in 2025 and Microduck launched in August 2026. Public information does not describe how far the two organisations are integrated at a product level. The points above are inferred from published product characteristics, not from stated strategy.