English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Microduck RL: A Complete Sim2Real Recipe for an 800g Bipedal Robot

Forum topic · ✨步子哥 · 2026-08-30

Summary

Microduck RL is an open-source repository from Pollen Robotics that trains reinforcement learning locomotion policies for Microduck, an 800-gram, 25-centimeter-tall bipedal robot with 14 servo joints running at 50 Hz. Instead of treating the reality gap as an afterthought, the project explicitly models real-world imperfections in simulation: the BAM (Backlash Actuator Model) from Rhoban Labs captures servo friction, gear backlash, and response delays, while ±1° gear backlash is simulated as a passive hinge joint behind each servo-driven joint, with encoder readings taken through the backlash. Combined with domain randomization, MuJoCo Warp GPU acceleration (4096 parallel environments, 1-2 hours per gait), ONNX export, and a real-robot runtime, the repo provides a fully reproducible sim2real pipeline. It ships 13 tasks — walking, fall recovery, stand-up, ground picking, ball kicking, forward rolls, and even roller-skating maneuvers — hot-swapped at runtime via a shared 61-dimensional observation contract. The project demonstrates a general principle: model known physical imperfections in simulation so deployment surprises become old friends, not new problems.

Microduck RL is an open-source repository from Pollen Robotics that trains reinforcement learning locomotion policies for Microduck, an 800-gram, 25-centimeter-tall bipedal robot with 14 servo joints running at 50 Hz. Instead of treating the reality gap as an afterthought, the project explicitly models real-world imperfections in simulation: the BAM (Backlash Actuator Model) from Rhoban Labs captures servo friction, gear backlash, and response delays, while ±1° gear backlash is simulated as a passive hinge joint behind each servo-driven joint, with encoder readings taken through the backlash. Combined with domain randomization, MuJoCo Warp GPU acceleration (4096 parallel environments, 1-2 hours per gait), ONNX export, and a real-robot runtime, the repo provides a fully reproducible sim2real pipeline. It ships 13 tasks — walking, fall recovery, stand-up, ground picking, ball kicking, forward rolls, and even roller-skating maneuvers — hot-swapped at runtime via a shared 61-dimensional observation contract. The project demonstrates a general principle: model known physical imperfections in simulation so deployment surprises become old friends, not new problems.

Tags

#robotics#reinforcement-learning#sim2real#mujoco#bipedal-robot#onnx#domain-randomization#open-source

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634283