Add observation term history support to Observation Manager

[jtigue-bdai]

Description

This PR adds observation history by adding configuration parameters to ObservationTerms and having the ObservationManager handling the collection and storage of the histories via CircularBuffers.

Fixes #1208

Type of change

• New feature (non-breaking change which adds functionality)

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

[KyleM73]

From the discussion in #1208, this PR introduces observation history tracking per-term. It was discussed that per-group tracking should also be implemented. In theory, the same effect can be had by giving all obs terms the same history length with the config provided in this PR. However, I want to reiterate what I think are the two best arguments for implementing per-group obs histories as well, and would be happy to hear feedback:

1. When writing deployment code, it is often more straightforward and natural to save the entire observation to a 2D buffer of size (H,D) for H history and D observation dimension. In the current implementation, separate buffers would need to be kept for each observation term as defined by Isaac Lab. However, it may not be trivial to do so, whereas to evaluate RL policies on hardware it is guaranteed that the observations will be stacked into the full length-D array at some point, making saving them as a per-group history easier to replicate on hardware.
2. Providing an implementation at the group level is a cleaner interface for developers than having to change the history length for each obs term. Especially when subclassing observation group configs from parent classes, it would be much nicer to define the history in the observation group post init (the interface currently used for enable_corruption and concatenate_terms, etc.) as opposed to having to change it for every term. Having used only my own version of the code in this PR for the last ~6 months, I have inadvertently missed changing the history length for one or two obs terms many times, requiring restarting the run. This bug can be avoided by optionally choosing to only change the history length in a single location, namely in the obs group post init.

The PR as-is resolves per-term observation history tracking. If it is decided that Isaac Lab should only support per-term observation histories then I am happy to close my comment/proposal with this PR. Otherwise, I'm happy to help implement per-group history tracking as well.

Thanks for putting this together James!

[jtigue-bdai]

I agree @KyleM73,the current implementation only covers term level that we talked about (in a slightly different implementation). I think having the option for group level is still worthwhile. I also agree we should be able to apply group level history via an overall group history_length. I was thinking, if set, the group level history length would override the term history length within that group.

[aravindev]

@jtigue-bdai Isn't it better to initialize the buffer to the latest data on the first append or right after a reset?

Otherwise, we assume that zero is a valid value for the observation, which may not always be true.
For example, if we are gathering a history of, lets say, the gravitational force, which may be defined as a strictly negative value, having a buffer filled with zeros as previous observations may not be within the expected distribution.

I would suggest that, during reset or at first init, all the indices in the history is initialized to the most recent data, possibly the one being passed into append()

[aravindev]

I cannot find how to provide a suggestion to lines that are not part of the changelist, so here we go.

In the method def __str__(self) -> str:,

You may need to re-compute the observation dimensions depending on whether flattening is enabled or not.
I see two deficiencies which can cause confusions:

1. Currently, the printed summary does not handle history length while computing the observation dimension.
2. The self._group_obs_dim is wrong and does not correspond to the actual observation dimension if history is used. This is quite critical to be fixed IMO.
3. The shape of the observation group printed in the summary is wrong.

Hope this helps!

[aravindev]

Suggested change

	# create history buffers
	if term_cfg.history_length > 0:
	group_entry_history_buffer[term_name] = CircularBuffer(
	max_len=term_cfg.history_length, batch_size=self._env.num_envs, device=self._env.device
	)
	# call function the first time to fill up dimensions
	obs_dims = tuple(term_cfg.func(self._env, **term_cfg.params).shape)
	# call function the first time to fill up dimensions
	obs_dims = tuple(term_cfg.func(self._env, **term_cfg.params).shape)
	# create history buffers
	if term_cfg.history_length > 0:
	group_entry_history_buffer[term_name] = CircularBuffer(
	max_len=term_cfg.history_length,
	batch_size=self._env.num_envs,
	device=self._env.device,
	)
	obs_dims = (obs_dims[0], term_cfg.history_length * obs_dims[1], *obs_dims[2:])
	if term_cfg.flatten_history_dim:
	obs_dims = (obs_dims[0], np.prod(obs_dims[1:]),)

This populates the correct _group_obs_term_dim into the dictionary. Later, this is used to compute the _group_obs_dim however, I assume that its computation does not need any changes.