What's this about?
Hugging Face's robotics arm, Pollen Robotics, unveiled Microduck: a 25cm bipedal robot under 800g, with 15 actuators, a camera, a small ToF LiDAR, two IMUs, microphones, a speaker and an articulated beak it uses to pick things up. It ships with seven pre-trained behaviours — walking, sitting, crouching, roller-skating, grabbing, getting back up after a fall — and the entire stack (SDK, simulator, RL pipeline) is open source under Apache-2.0. It costs $399 and ships before Christmas. Within about six hours it had passed $1,000,000 in sales, selling roughly one every five seconds.
What does "you train it yourself with reinforcement learning" actually mean?
Reinforcement learning is training by trial and error against a score. Nobody writes the walking code; instead you simulate thousands of ducks falling over, reward the ones that stay upright and move forward, and repeat until a walking policy emerges. The hard part is that a policy trained in a simulator usually fails on real hardware, because the simulator's physics are too clean — this is the "sim-to-real gap". Microduck's pipeline deliberately models the messy parts (actuator lag, dropping battery voltage, command delay, friction) so the policy that works in simulation also works on a real, cheap, slightly wobbly robot.
What's the community saying?
- The most-asked practical question is battery life — one hour per the spec sheet — followed immediately by whether it can find and dock with a charger on its own.
- Nostalgia for the Sony AIBO runs through the thread, with several people noting the original shipped in 1999 and nothing affordable filled the gap since.
- The recurring technical scepticism: "you can train deep learning to do anything" is easy to say, but will it do the thing reliably, and how often does it fall over?
- Several commenters treat the open-source stack, not the duck, as the real product — a programmable platform with a camera and LiDAR at hobbyist prices.
- A separate, much quieter post raises security concerns about the GitHub docs (see the AI Security section).