Fixes reset of sensor drift inside the RayCaster sensor

[zoctipus]

Description

The drift in raycaster didn't update when reset. This MR now updates it correctly:

before:

self.drift[env_ids].uniform_(*self.cfg.drift_range)

now:

self.drift[env_ids] = self.drift[env_ids].uniform_(*self.cfg.drift_range)

Fixes #1820

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist

• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[kellyguo11]

Is this required for uniform_? I think that might be an inplace operation already.

[zoctipus]

Thank you for being careful, I have checked again in my debugger

>self.drift[env_ids]
tensor([[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        ...,
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]], device='cuda:0')
>self.drift[env_ids].uniform_(*self.cfg.drift_range)
tensor([[-0.0657,  0.0741, -0.0464],
        [-0.0895,  0.0378,  0.0849],
        [ 0.0414, -0.0142,  0.0127],
        ...,
        [ 0.0647,  0.0327, -0.0896],
        [ 0.0268, -0.0399, -0.0197],
        [ 0.0034, -0.0036,  0.0639]], device='cuda:0')
>self.drift[env_ids]
tensor([[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.],
        ...,
        [0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]], device='cuda:0')
        

as you see the self.drift is not updated after the uniform_ operation, though it returns a uniformly sampled values

[zoctipus]

I do remember styles in pytorch where they use _ to indicate in place operation, it is indeed strange how it didn't work here. My torch version is 2.5.1+cu118

[Mayankm96]

The problem isn't with the in-place operation. Rther if you are indexing a torch tensor with explicit indices, it makes a "copy" of that tensor. So the original tensor isn't getting assigned but rather the copy of it.

Nice catch!

[zoctipus]

@kellyguo11
Upon further research, I found out this indexing is causing the issue

Comparing below two cases

Case 1:

>self.drift
tensor([[0., 0., 0.],
        ...,
        [0., 0., 0.]], device='cuda:0')
>self.drift.uniform_(*self.cfg.drift_range)
tensor([[-0.0657,  0.0741, -0.0464],
        ...,
        [ 0.0034, -0.0036,  0.0639]], device='cuda:0')
>self.drift
tensor([[-0.0657,  0.0741, -0.0464],
        ...,
        [ 0.0034, -0.0036,  0.0639]], device='cuda:0')

Case 2:

>self.drift[env_ids]
tensor([[0., 0., 0.],
        ...,
        [0., 0., 0.]], device='cuda:0')
>self.drift[env_ids].uniform_(*self.cfg.drift_range)
tensor([[-0.0657,  0.0741, -0.0464],
        ...,
        [ 0.0034, -0.0036,  0.0639]], device='cuda:0')
>self.drift[env_ids].     <------------------   change did not take effect
tensor([[0., 0., 0.],
        ...,
        [0., 0., 0.]], device='cuda:0')

it seems like the advance indexing returns coped data instead of of original data's view, and uniform_ indeed changed the data in place, but the change is on copied data!