pollen-robotics/microduck

A Tiny biped duck robot 🦆

What it solves

Microduck provides the control software ("the brain") for a small biped robot, enabling it to perform complex physical movements like walking, rolling, standing up after being knocked over, and picking up objects using neural policies.

How it works

The system runs on a Rockchip RK3566 and consists of several Rust-based daemons that communicate via JSON-RPC over Unix sockets. A 50 Hz control loop drives fifteen servos based on neural policies exported as ONNX files. These policies are trained using reinforcement learning (PPO) in MuJoCo simulations with domain randomization to ensure they transfer successfully from simulation to reality (sim2real).

Who it’s for

It is designed for users who own a Microduck robot, as well as developers looking to build upon the platform or experiment with reinforcement learning for bipedal locomotion.

Highlights

  • RL-Driven Movement: Uses neural policies trained via PPO and MuJoCo for natural bipedal movement.
  • Robust Update System: Features a verified, health-gated, and reversible update mechanism to prevent bricking.
  • Multi-modal Interaction: Supports control via gamepad, Bluetooth, and laptop scripts.
  • Integrated Hardware Support: Manages a camera stream via WebRTC and a depth sensor.

Related

  • Project
    TheRobotStudio/SO-ARM100

    An open-source design for the SO-100 and SO-101 low-cost robot arms, designed for seamless integration with the LeRobot library to enable accessible AI robotics.

  • Project
    NVIDIA/warp

    A Python framework for GPU-accelerated, differentiable simulations used in robotics, physics, and machine learning.

  • Project
    NVIDIA/Isaac-GR00T

    NVIDIA Isaac GR00T N1.7 is a vision‑language‑action model that turns video + text instructions into continuous robot actions. It uses a Qwen3‑VL visual backbone, a diffusion transformer head, and a relative end‑effector action space to work across many robot embodiments. The repo provides installation scripts, data‑format specs, zero‑shot inference, fine‑tuning tools, benchmark evaluations, and deployment pipelines (ONNX/TensorRT) for platforms ranging from desktop GPUs to Jetson devices.

  • Project
    isaac-sim/IsaacLab

    A GPU-accelerated open-source framework built on NVIDIA Isaac Sim for simplifying robotics research in reinforcement learning, imitation learning, and motion planning.