[Rohithmatham12]

Summary

• clarify that inverse dynamics predicts action trajectories but does not synthesize full LeRobot observation.state records
• point users to the checked-in DROID LeRobot sample metadata for timestep-aligned state streams
• document that current action-layout and normalization references are the checked-in AV, DROID, and UMI assets/notebooks until additional embodiment assets are released

Why
Issue #192 asks how to obtain observation.state alongside predicted actions for LeRobot-format episodes. Issue #184 asks for action channel layouts and normalizers across embodiments. This PR documents the current supported boundary and points to the concrete checked-in sample without claiming unreleased embodiment-specific layouts.

Testing

• git diff --check

Addresses #192
Related to #184

[mli0603]

Thanks @Rohithmatham12. Will consolidate the your PR and the docs! We have a PR landing soon and we will take a look.

[MaciejBalaNV]

@mli0603 @Rohithmatham12
Is this PR still relevant? There were a few changes recently in the action normalization on cosmos_framework side.

[Rohithmatham12]

Thanks for checking. I rebased this on current main and refreshed the normalization wording.

I think the PR is still relevant for the observation.state / LeRobot part of #192: inverse dynamics returns predicted action trajectories, but it still does not synthesize a full LeRobot episode record, so users need to align predicted action rows with the source dataset's timestep-level observation.state, frame, timestamp, and camera streams.

For the normalization side, I updated the new section so it points at the current action-definition table on this README as the layout summary for AV, DROID, UMI, and human hand-pose examples, with the JSON assets/notebooks showing how those layouts are normalized and passed to each backend. It no longer implies the older checked-in assets are the only reference after the cosmos_framework changes.

If you prefer to fold this into a broader docs update, that is fine too, but I think the state-vs-action clarification is still a useful boundary to keep documented.

[yaoxu-crypto]

Thanks for this, @Rohithmatham12. The Cosmos cookbooks and evaluation suites have moved to NVIDIA/cosmos-framework under .agents/skills/ and were removed from this repo in #384, so this PR can no longer merge here.

Your clarification is still relevant: the action README in Cosmos Framework doesn't cover the observation.state alignment from #192. Could you re-submit it against .agents/skills/cookbooks/cosmos3/generator/action/README.md?

The relative paths are unchanged, just under .agents/skills/ (for example cookbooks/cosmos3/... is now .agents/skills/cookbooks/cosmos3/...). Please note that Cosmos Framework requires DCO sign-off on each commit (git commit -s, see CONTRIBUTING.md). I'm closing this PR because it can't merge here, but the work is welcome there. Please link this PR from the new one so the earlier discussion and reviews carry over. Thanks for your patience, and sorry it sat here while the move happened.

[Rohithmatham12]

Thanks for the guidance. I re-submitted the documentation against the moved README in NVIDIA/cosmos-framework#353 with a DCO-signed commit and linked this PR for the earlier context.