Bipedal Robots Are Officially Walking Into Everyday Life IoT
Microduck is an open-source tiny biped robot by French robotics company Pollen Robotics: about 25 cm and 800 g, its walking policies are trained with MuJoCo + PPO reinforcement learning and transferred to real hardware via sim2real, driven by a 50 Hz real-time loop on a Rockchip RK3566 running fifteen servos. With 900+ GitHub stars, it marks the moment bipedal robots officially walk into everyday life.
A Duck That Picks Itself Up Microduck is a duck-shaped biped: drive it with a gamepad, knock it over and it stands itself up, put wheels on and it rolls forward, lower its beak and it picks up small objects. It also sits, kicks a ball, and quacks in a voice of its own. At about 25 cm and 800 g, it is the first true desktop-scale biped robot.
Reinforcement Learning Makes Bipedal Walking Real Bipedal walking is one of robotics' hardest classic problems. Microduck's gait is not hand-written kinematics — it is trained with PPO in MuJoCo simulation under domain randomization (model mass, friction, latency), exported to ONNX, and deployed into the robot's 50 Hz control loop. That is sim2real: the transfer that takes walking robots out of the lab and onto ordinary desktops.
System Design Behind 25 cm and 800 g Small size did not mean simple internals: a Rockchip RK3566 runs a cluster of Rust daemons — robotd owns the 50 Hz control loop and motor bus; updaterd installs signed updates and rolls back on unhealthy boot; mediad streams the camera over WebRTC; tofd serves the depth sensor. Bluetooth, the gamepad and the phone app all speak one JSON-RPC contract over Unix sockets — a miniature industrial-grade real-time system.
Same Technology, Different Scale 50 Hz real-time control, signed OTA updates, sensor fusion, wireless links — the technologies that make Microduck work are exactly what smart street light controllers and concentrators (CCU) handle every day. Cassan's 4G/LoRa per-lamp controllers run the same kind of real-time loops and remote OTA upgrades, bringing robot-grade reliability to every street light.
What Open Source Means Here Apache-2.0, 642 commits, 900+ stars: anyone can read the control-loop source, reproduce the update machinery, and train their own policies in simulation before deploying to hardware. Open source turned the bipedal robot from an expensive research device into a learnable, improvable, replicable engineering sample — the biggest driver of robots entering everyday life.