How to add my own robot and train it through this pipeline?

[vegfood]

I'm trying to add my own robot to the existing training pipeline but encountering some configuration issues. My robot uses motor-type actuators, while the current demo appears to use a PD-based position controller. I need to understand:

1. How to modify the robot configuration to adapt to motor drive?

2. Which parameters or controller settings need adjustment?

3. Which parts of the training pipeline require corresponding modifications?

4. Are there any example configurations for motor-driven systems that I can reference?

[louislelay]

Hi @vegfood, thank you for posting that!

In mjlab you can skip actuator injection and rely on the actuators defined in your XML by leaving the EntityArticulationInfoCfg’s actuators field empty in the asset config so controls are forwarded to your MJCF actuators (e.g., motor-type).

For motor-type actuators, as it is and depending on what you want your policy to output, it would require some changes:

• If you want your policy to output effort, feed those to the motor actuators.
• If you want your policy to output positions, add a light impedance layer (τ = Kp (q*−q) + Kd (−qd)) to compute torques and feed those to the motor actuators.

For now the public examples currently targets Unitree G1 and Go1, with impedance via position actuators We plan to expand actuator configurations, once that lands, the workflow above will be much more plug and play.