refactor: move intrinsics/extrinsics into embodiment classes

[ElmoPA]

Move ARIA/MECKA/SCALE intrinsics and EXTRINSICS dict out of egomimicUtils
and into the embodiment classes (Aria, Mecka, Scale, Eva) as class attributes.
Remove CameraTransforms class and VIZ_INTRINSICS_KEY indirection; viz functions
now take an intrinsics matrix directly via cls.INTRINSICS.

[ElmoPA]

• feat(zarr): persist camera intrinsics in zarr.json #386
• feat: add LightWheel embodiment #381
• Contribution Update #382
• refactor: move intrinsics/extrinsics into embodiment classes #380 👈 (View in Graphite)
• main

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

[SimarKareer]

Just type it directly into here instead of this ugly code above

[github-actions]

Claude Code Review

PR Review: Move intrinsics/extrinsics into embodiment classes

Summary

Refactor moves camera intrinsics/extrinsics constants from egomimicUtils.py into embodiment classes as class attributes, and removes the CameraTransforms / VIZ_INTRINSICS_KEY indirection. The direction is good, but the implementation has several correctness bugs that will break training/rollouts.

Key concerns

🔴 1. Bogus LW_BIMANUAL enum alias — will raise ValueError

LW_BIMANUAL = 5  # alias of ARIA_BIMANUAL — routes through aria handling for now

Python Enum does not allow duplicate values to be defined as separate members — LW_BIMANUAL will silently become an alias such that EMBODIMENT.LW_BIMANUAL is EMBODIMENT.ARIA_BIMANUAL, and EMBODIMENT_ID_TO_KEY[5] will resolve to whichever name member.name iteration picks (likely ARIA_BIMANUAL). Even if intentional, this collides with the embodiment-string convention (must match exact enum name). This also looks unrelated to the PR's stated purpose — please remove it or split into a separate PR with proper enum handling (@enum.unique would have caught this; consider IntEnum with a distinct value, or aria_bimanual as the canonical string).

🔴 2. Lost extrinsics — Eva.EXTRINSICS is only x5Dec13_2

The original EXTRINSICS dict contained many calibrations (ariaJul29, ariaFeb17, ariaMar4, ariaJun7, ariaOct18_arx, x5Nov15_2, x5Nov18_3, x5Dec10_2, x5Dec13_2, etc.). This PR drops all of them and keeps only x5Dec13_2 as Eva.EXTRINSICS. Any episode/checkpoint/visualization script keyed off an older calibration is now silently using the wrong extrinsics. Since poses are stored in SLAM world frame and re-expressed at training time, this can silently corrupt training targets for older datasets without raising.

Also: _build_eva_bimanual_eef_frame_transform_list and _build_eva_bimanual_transform_list previously took extrinsics_key as an argument. Removing the parameter eliminates the ability to switch calibrations per dataset.

🔴 3. rollout.py — extrinsics_key removed from PolicyRollout signature

extrinsics_key="x5Dec13_2" was a function arg threaded through main(). Removing it hardcodes one calibration into rollouts. If multiple robot calibrations are used across rollouts, this is a regression. At minimum, keep the parameter (or a calibration enum) even if it currently maps to a single Eva.EXTRINSICS.

🟡 4. Aria class body is malformed

return super().viz(image, viz_data, mode=mode, **kwargs)
    FINGER_EDGES = [

Missing blank line + indentation is suspect — visually looks like FINGER_EDGES was meant to be at class scope but is now positioned after a return inside the viz method's source range. Re-read: actually it's at class scope (no leading whitespace inside def), but Aria.FINGER_EDGES duplicates what's already on Human. Either remove or confirm Aria override is intended. Same for FINGER_COLORS, FINGER_EDGE_RANGES, DOT_COLOR, ACTION_STRIDE=3 (already inherited).

Also missing blank line between Aria and Scale classes.

🟡 5. Scale.EXTRINSICS = None and Mecka.EXTRINSICS = None

Previously these were identity dicts ({"left": np.eye(4), "right": np.eye(4)}). If anything downstream does embodiment.EXTRINSICS["left"] it will now TypeError. Prefer keeping the identity dict for type stability, or document that None means "no extrinsics applied."

🟡 6. Notebook references undefined LightWheel

vis = LightWheel.viz_transformed_batch(batch, mode="traj")

LightWheel is not imported or defined in this diff. Notebook will NameError.

🟡 7. Circular-import risk

viz_utils.py no longer imports INTRINSICS (good), but egomimic/rldb/embodiment/human.py now defines ARIA_INTRINSICS etc., and eva.py imports from egomimic.rldb.embodiment.human import ARIA_INTRINSICS. This is fine, but the constants conceptually belong in a cameras.py / calibrations.py module rather than buried in human.py — Eva is not a human embodiment and importing from human for camera intrinsics is confusing.

🟡 8. Test coverage

No new tests. At minimum:

• A test asserting each embodiment class exposes a valid INTRINSICS (3x4 numpy array) and EXTRINSICS (dict with left/right 4x4 matrices, or explicitly None).
• A regression test that Eva.EXTRINSICS["left"] matches the previous EXTRINSICS["x5Dec13_2"]["left"] bit-exactly.

Suggestions

1. Drop the LW_BIMANUAL change from this PR. If needed, do it properly in a separate PR with a unique enum value.
2. Preserve all extrinsics calibrations. Move them into a CALIBRATIONS: dict[str, dict] class attribute on Eva (and a

---

Reviewed by Claude · Review workflow