Most robotics launches ask you to trust a demo video. Pollen Robotics, the Bordeaux robotics team at Hugging Face, is instead shipping the training loop.

The team has opened pre-orders for Microduck: a 25 cm bipedal robot priced at $399. Every movement it makes — walking, sitting, kicking, roller-skating, standing back up after a fall — is a neural policy trained in a physics simulator and exported to the hardware.

The real point is this: the training environments, the reward functions, the domain-randomisation settings and the sim-to-real recipe are all public on GitHub.

Microduck follows Reachy Mini, which has shipped more than 10,000 units, but reverses its premise: where Reachy Mini was built to sit on a desk and interact, Microduck is built to leave the desk, fall over, and get back up.

The hardware

  • Size: 25 cm tall, 14 cm wide, under 800 g.
  • Motion: 15 motors across legs, neck and head, plus an articulated beak that picks objects off the floor.
  • Compute: a Rockchip RK3566 with an AI accelerator, 1 GB of RAM, 32 GB of storage.
  • Sensing: a front camera behind a dedicated camera-use indicator, two IMUs (one in the body, one in the head), a compact LiDAR, microphones and a speaker, two NFC antennas, Wi-Fi and Bluetooth.
  • Power: a removable 2600 mAh battery, good for about an hour.

Seven trained moves ship in the box and can be driven by a bundled game controller before you write any code. The robot does not speak; each unit generates its own audio identity on first wake and keeps that voice permanently.

How the behaviours are actually trained

Policies are trained with reinforcement learning in a MuJoCo-based environment. Pollen reports roughly one to two hours on a CUDA GPU for a usable gait at 4096 parallel environments. Without a local GPU, the same command runs on Hugging Face's own infrastructure.

The hard part is sim-to-real, and what the team did there is notable: each servo is simulated not as an ideal controller but with a behavioural model of the real motor — voltage control law, back-EMF and load-dependent friction included.

Randomisation covers battery voltage, voltage sag under load, command delay and friction magnitude. Gear backlash gets its own training variant, assuming a degree of play at each joint.

Why it matters

Most robotics announcements show the result and not the road. The difference here is that the method is published rather than the outcome: you can train your own policy with the same recipe and export it to the hardware.

The price matters for the same reason. A four-hundred-dollar device is within reach of a student or a hobbyist, not just a university lab. Given that the field's biggest bottleneck is data and experience, what falls is not only the price: the number of attempts goes up.