Simulation Stability and Tuning Model Controllers

[peterdavidfagan]

This is not explicitly a bug but rather a question that likely belongs in a discussion section.

Which model is the issue affecting? Franka Emika Panda, UFactory Lite6

What is the issue? I am attempting basic pick and place behaviours in an object manipulation benchmark that leverages MuJoCo. I am struggling to run reliable control on existing models. Thus far I have attempted to configure physics time stepping (here) & control time stepping (here) as well as the actuator settings for the above robots. I am directly controlling the robots by passing joint targets to robot actuator models, I have mostly tested with direct position controllers but plan to attempt using PD controllers.

I wished to ask what is the best practice in this situation. Should I fix my sim step and control step and then tune controllers to this? If I am tuning controllers is it expected to perform this tuning manually or does there exist and open-source package for MuJoCo that provides automated tuning.

My current set of code can be found here. To reproduce what I am observing the following is required:

• System installation of poetry
• Clone repo and submodules (Note: the following submodules can be removed vcp, ros2_robotics_research_toolkit, transporter_networks)
• Install poetry env with poetry install
• Run the following script python transporter_network_data_generation.py

Is there any additional context you can provide (e.g., a spec sheet or a URDF to show a value mismatch)? I am happy to create a simpler reproducible example. I am mostly looking for advice that will help fast track my resolution of this issue. Thanks for open-sourcing this software.

[peterdavidfagan]

Moving to MuJoCo issues/ask for help as it is likely a more suitable location.

[peterdavidfagan]

Issue moved to here.