How can I dynamically change the color of a cube?

[lovecode1]

How can I dynamically change the color of a cube?

I did the following:

cube.surface.color = (1.0, 0.0, 1.0, 1.0)

But in the scene I do not see any change. How can I force the change in the color.
When I do print(cube.surface.color) I do see that the color changed in the data structure but I do not see any change in the simulator.

What am I missing?

[duburcqa]

I'm afraid changing the color dynamically is not supported.

[lovecode1]

Thanks @duburcqa!
What about size or material changes—are any of them supported?

[duburcqa]

Material changes definitely not. Size it depends what you mean, it may be doable.

[lovecode1]

If I have a cube, can I scale it by 2 dynamically?
If so, which function allows scaling it at runtime?

My main goal is to create multiple episodes to train OpenVLA. To add more variance between episodes, I want to change materials and sizes to help train the model more efficiently.

Is there a better way to create these variant episodes? I’m assuming restarting Genesis for each episode is pretty time-consuming?

[jgillick]

One idea: create all the variants ahead of time, either in URDF/Mujoco or with gs.morphs.Box. Build them all in the scene, but only move the target objects into the primary workspace. At each reset swap the target objects in and out of the workspace. This isn't as easy as just modifying the size and material of the element at runtime, but I assume it should be easy enough.

[lovecode1]

@jgillick Thanks for the idea!
By moving in or out of the workspace, do you mean whether it is visible to the user—meaning, moving it out of the workspace and far away from the camera’s view? Am I correct?

[jgillick]

That's what I was thinking, yes. That said, I'm not familiar with your project, so I don't know if that would work for it.

[lovecode1]

Yes that should work.
Does the physics engine process items outside the camera view? I’m wondering if that adds load on the CPU/GPU, or if it only consumes memory without affecting processing time?

[jgillick]

(disclaimer: I'm not on the genesis team, so my answers might not be 100% accurate)
That's a good point, the physics engine will process anything in the scene, regardless of the camera view. As long as there aren't any complex collisions occurring, it might be okay.

Alternatively, I think you can combine objects during training to make them bigger or heavier. For example, combine two blocks to make a single object twice as big.

[MandiZhao]

@duburcqa following up on this -- I'm curious is there a way to dynamically rescale a loaded mesh at runtime? Assuming my env only has a few objects, I don't need to load too many different assets, but resizing the mesh during training time would be a super helpful DR option for RL