FULL STORY

Microduck: From Open-Source RL Stack to Handstands

Pollen Robotics open-sourced the RL training stack for its tiny bipedal robot Microduck. Days later a team trained it to perform handstands, with more skills on the way.

2026-08-28 ~ 2026-08-30 · 2 episodes · 8 posts

Episode 1 · Pollen Robotics open-sources Microduck bipedal robot RL training stack (2026-08-28, 5 posts)

Pollen Robotics open-sourced microduckrl, the reinforcement learning training stack for its miniature bipedal robot Microduck, and demonstrated a complete sim2real loop: policies trained in simulation are exported as ONNX and deployed directly onto the physical robot, where the robot can also learn new tricks. Built on the MuJoCo Warp (mjlab) simulator and PPO, with the hardware available for purchase, the project is a rare open-source package combining a training environment, a full sim2real recipe, and purchasable hardware.

Confirmed

  • The training environment is built on MuJoCo Warp (mjlab) with PPO, and the microduckrl library is fully open-sourced
  • A complete sim2real loop was demonstrated: train in simulation, run in reality, and teach the robot new skills
  • Policies are exported as ONNX and deployed to the physical robot
  • The simulation explicitly models actuator behavior, battery voltage, friction, and even gear backlash to close the sim-to-real gap; the release also details techniques such as head tracking and backlash simulation
  • The project includes training for multiple skills such as walking
  • Microduck hardware is available for purchase

Why it matters

  • It is a rare three-in-one open-source offering — training environment, full sim2real recipe, and purchasable hardware — lowering the barrier for researchers to reproduce and iterate RL policies on a real bipedal robot
  • The explicit modeling of physical details like actuators, battery voltage, friction, and gear backlash provides a referenceable engineering template for sim2real practice

Episode 2 · Microduck Robot Masters Handstand, Eyes Head-Spin Next (2026-08-29, 3 posts)

The Microduck robot has learned to perform a handstand, with its team now aiming to teach it head-walking, hopping like the Pixar lamp, and even head-spin dance moves, inviting community strategies for real-world demos.