Microduck RL
Reinforcement learning environments for a small bipedal robot
What it is
A set of MuJoCo Warp training environments, built on mjlab with PPO, for teaching the 25 cm Microduck biped to walk, stand up, sit, kick a ball, roll, and skate on passive wheels. It models the Dynamixel XL330 servos down to their voltage control law, adds domain randomization and gear backlash, and exports policies to ONNX for the real robot. Robotics researchers and hobbyists working on sim-to-real locomotion can train on a CUDA GPU or Hugging Face Jobs and test policies with keyboard control in simulation.