microduck_rl — the training environments

Every move Microduck makes is trained here and exported as .onnx for the robot to load.

Its real value is that the whole sim2real recipe is in the code rather than in somebody’s head:

• The BAM M6 actuator model — voltage control law, back-EMF, Coulomb/Stribeck/load-dependent friction, for the Dynamixel XL330
• Domain randomisation over physical parameters — battery voltage, voltage sag under load, command delay, friction magnitude — not observation noise
• Backlash modelling at ±1° per servo, with the encoder on the output side of the play, matching the real structure
• A 61-dimensional shared observation contract, which is what lets walk, recover and trick policies hand over at any moment

More than a dozen tasks, from the main Mjlab-Velocity-Flat-MicroDuck to a full family of roller tasks. No GPU? Add --hf-jobs to run on Hugging Face Jobs.

See Train and deploy and the task pages.