Starter path · step 07 · 6 to 10 weeks
Legs and sim-to-real
Why a policy that walks in simulation falls over on the floor, and how the Open Duck Mini closes the gap.
The duck in the Hardware Pool is a research problem you can hold. It walks on a policy trained in MuJoCo and transferred to real servos. This module is about that transfer.
Do this
- Underactuated Robotics (MIT, Russ Tedrake; free textbook). Chapters 1 to 4 for why legged robots are hard, then the chapters on walking. Read slowly; it's dense and it's worth it.
- Open Duck Mini (Antoine Pirrone and contributors). The robot's home: bill of materials, print guide, assembly, the onboard runtime, and the sim-to-real notes. Read the sim-to-real document before anything else.
- Open Duck Playground. The training setup, built on MuJoCo Playground. This is where you train your own walking policy.
- Isaac Lab documentation (NVIDIA). The other major simulator for legged learning. Read the overview so you know what's out there; you don't have to switch.
Build
Train a walking policy for the duck in the Playground. Add domain randomisation: mass, friction, motor delay. Compare how far the randomised and unrandomised policies get on the real robot, if you have it. Log the falls.
Skip
- Building your own duck before you've walked a borrowed one.
- Model-based control of the whole biped. Learn it later if you still want to.
Done when: you've trained a policy for the duck, you can say what domain randomisation is for, and you've watched a real robot fall for a reason you can name.