feat: add LightWheel embodiment

[ElmoPA]

Adds LIGHTWHEEL_INTRINSICS (fx=fy=786.6216072, cx=960.0, cy=728.0) and a
LightWheel class that delegates transforms and keymap to Aria.

[ElmoPA]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• feat: add LightWheel embodiment #381 👈 (View in Graphite)
• Contribution Update #382
• refactor: move intrinsics/extrinsics into embodiment classes #380
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[github-actions]

Claude Code Review

Summary

Adds a LightWheel embodiment class that delegates get_transform_list and _get_keymap to Aria, plus a new LIGHTWHEEL_INTRINSICS matrix.

Key concerns

1. Silent None return for unsupported modes. get_transform_list only handles "cartesian" and implicitly returns None for all four keypoints modes. Downstream code expects a list[Transform] and will crash with a confusing NoneType error. Either raise NotImplementedError explicitly or delegate all modes to Aria if they're supported.

2. EXTRINSICS = None. Aria has a defined ARIA_T_RGB_CPF. If LightWheel reuses Aria's transforms (and presumably Aria's coordinate frame conventions), it likely needs an extrinsic too — otherwise any pose re-expression to head/CPF frame will be wrong or fail. Please confirm whether LightWheel data is already in CPF frame, and if not, define the appropriate extrinsic.

3. Intrinsics dtype unspecified. Unlike MECKA_INTRINSICS above it, this array omits dtype=np.float64. Minor, but inconsistent and could trip downstream type assertions.

4. ACTION_STRIDE = 3 unjustified. Aria's stride should be checked — if LightWheel mirrors Aria semantics, the stride should match Aria's, not be hardcoded independently. Please verify against an actual LightWheel dataset's frame rate.

5. No registration / enum check. Is lightwheel (or lightwheel_bimanual?) registered in the embodiment enum / dispatch used by rldb loaders and configs? Without that, this class is unreachable.

6. No tests / no example data path. A new embodiment normally needs at least a smoke test loading a small zarr and running the transform list end-to-end.

Suggestions

• Replace the silent fall-through with:

  if mode == "cartesian":
      return Aria.get_transform_list(mode="cartesian")
  raise NotImplementedError(f"LightWheel does not support mode={mode}")

  …or, if Aria's keypoint modes are valid for LightWheel, just return Aria.get_transform_list(mode=mode) for all modes.
• Add dtype=np.float64 to LIGHTWHEEL_INTRINSICS.
• Document (in a comment) the camera/source for the intrinsics (sensor model, resolution 1920x1456?).
• Confirm extrinsics: if LightWheel images are already rectified into a CPF-equivalent frame, note that explicitly; otherwise add the correct T_RGB_CPF.
• Register LightWheel wherever Aria/Mecka/Scale are dispatched.
• Add at minimum a small unit test instantiating LightWheel.get_transform_list("cartesian") and confirming the keymap matches Aria's.

Verdict

Request Changes — the silent None return is a latent crash, and the missing extrinsics + enum registration need to be resolved before this is usable in training.

---

Reviewed by Claude · Review workflow