Replaces TensorDict with native dictionary

[Toni-SM]

Description

Remove TensorDict usage from Isaac Lab

Type of change

• Breaking change... for those ones using TensorDict specific API such as camera.data.output.to_dict() or camera.data.output.sorted_keys

Screenshots

Before

After

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

[pascal-roth]

Before a more detailed review, I was wondering what is the reason why we should prefer normal dicts instead of Tensordict (thought we would rather use more and not less tensordicts in the code)?

[Toni-SM]

Before a more detailed review, I was wondering what is the reason why we should prefer normal dicts instead of Tensordict (thought we would rather use more and not less tensordicts in the code)?

1. Env API handles state, observation, action, reward, terminated, truncated and info.

   • For state, observation and action, we support fundamental and composite spaces (at least in Direct workflow): Box (tensor), Discrete (tensor), MultiDiscrete (tensor), Dict (dict), Tuple (tuple). TensorDict does not support tuples, so there is no point in using it for spaces in single or multi-agent envs.
   • For single-agent envs, reward, terminated and truncated are just tensors not so in multi-agent (dict of tensors)... but since it cannot be used for spaces, it does not make sense that some components use it and others do not.

2. TensorDict is a PyTorch-scpecific thing which limits the implementation to that framework (making it difficult to scale Isaac Lab to other ML frameworks).

3. Nested TensorDicts is pythonic ....
   

[pascal-roth]

Okay makes sense to me, it was anyway basically the only place where we use them. @Mayankm96 if I remember correctly then you were in favor of using them, so tagging you here for viz

[vmoens]

Hey just a couple of comments based on the above discussion (not hoping that this PR will be reverted but I thought some of these comments were worth addressing)

Before a more detailed review, I was wondering what is the reason why we should prefer normal dicts instead of Tensordict (thought we would rather use more and not less tensordicts in the code)?

1. Env API handles state, observation, action, reward, terminated, truncated and info.

   • For state, observation and action, we support fundamental and composite spaces (at least in Direct workflow): Box (tensor), Discrete (tensor), MultiDiscrete (tensor), Dict (dict), Tuple (tuple). TensorDict does not support tuples, so there is no point in using it for spaces in single or multi-agent envs.

TensorDict does support tuples (it supports basically any python object) although most APIs won't work (like to(device) will just ignore these values). If you need that we could call tree_map over non-tensors by default, that could actually make sense!

2. TensorDict is a PyTorch-scpecific thing which limits the implementation to that framework (making it difficult to scale Isaac Lab to other ML frameworks).

On that I cannot disagree, it's definitely written for pytorch!

3. Nested TensorDicts is pythonic ....
   

hmmm ok?!

I'd just say that nested keys isn't really the main feature of the lib (and [k1][k2] or get(k1, {}).get(k2) will work perfectly too). The goal is to handle things like reshape ops / H2D transfers / arithmetic ops / functional calls / indexing etc as efficiently as possible across many tensors.