[Question] is it possible to generalise?

[MartinaRuocco]

Hi!
I am trying to integrate physics intuition during robot training. I am currently using Mujoco and that's why I think this project would be relatively easy to integrate.

How difficult would it be to extend this project to different tasks/physic events?

[ogroth]

Hi! Thanks for the interest in ShapeStacks. The MJCF files can easily be run with MuJoCo. However, I can't guarantee bug-free behavior with mujoco2. The project was done with mujoco 1.5 and a few things like friction computation on cylinder primitives seem to have changed. Which tasks / physic events do you have in mind?

[MartinaRuocco]

Hi Oliver, thank you for your fast response.

Let me first say that this is not my field, therefore I apologize in advance for any silly question.

To begin with, I'd like to analyze (and "learn") the physical events that occur on a given object as a result of the robot interaction, like rolling, sliding etc.
In the future, the idea is to move to more complex scenarios, like objects stacking.
I am not sure this can be done directly in Mujoco or it needs some external libraries.

Many thanks in advance

[ogroth]

Collecting data for physical interaction can easily be done with mujoco. The ShapeStacks paper was one example of that. You could, for instance, put a robot model next to the stacks (or just spawn an invisible force) and just have it knock them over and collect the episode as a video. If you want to do proper robot control and planning, you would need to look to other libraries though.

[MartinaRuocco]

Yes, that's definitely something that would fit my purposes.
Can you, by any chance, suggest some tutorial on how to use Mujoco for the mentioned purposes?