[RyanPCo]

Branch 3 of 8 in the split of #640, which is the tip of the stack and carries the map of all eight.
Unit suite on this branch: 460 passed, 0 failed.

compute_hand_pose_xyzquat built the palm normal as cross(thumb_dir, pinky_dir).
Thumb and pinky are anatomically swapped between hands, so that cross product
already mirrors chirality for free; applying rot_left on top of it flipped x/y
a second time and landed the LEFT hand on the RIGHT hand's spatial convention.
Both hands now use the same rot_right relabel.

Reconverting every mecka zarr would be the obvious fix, so this avoids it:
rot_left == rot_right @ diag(-1, -1, 1), which makes right-multiplying the
stored left pose by Rz(180 deg) exactly equivalent to reconverting with
rot_right. RotateLocalFrame does that right-multiply (R_new = R_old @ R_fix),
relabelling the pose's own axes without moving its origin, on (7,) poses and
(T, 7) chunks alike, and leaves zero-norm quaternions alone because those mark
padded frames.

Human.get_transform_list(fix_mecka_left_wrist=True) prepends it to the raw LEFT
pose keys before any frame math: ee_pose in the cartesian modes, wrist_pose in
the keypoints modes, since the converter wrote the same rotation into both.
data/mecka.yaml opts in; aria and scale must not, because they run different
converters, and neither must mecka data reconverted after this commit, which
would double-flip.

test_wrist6d_roundtrip runs the real transform lists on synthetic world-frame
poses, parametrized over the flag, and compares every stage against plain
numpy/scipy SE(3) math that shares no code with the pipeline.

Co-Authored-By: Claude Opus 5 (1M context) [email protected]

🤖 Generated with Claude Code

[RyanPCo]

• hpt: never re-initialise pretrained stems; hash-check them; mask text padding #645 : 2 dependent PRs (#646 , #664 )
• obs: proprio history (K proprio steps, one token each, learned time embedding) #650
• loss: mask the repeat-last action padding out of the HPT loss #649
• resnet: keep a frozen backbone in eval mode; give backbones their own LR group #652
• augs: per-sample image augmentation for HPT train-mode augs #651
• eval: deterministic validation (seeded prompts + flow sampling); log-only grad-norm MAD #644
• data: mecka_fold_flagship_top3_hpt_6d, the 3-operator flagship train split #643
• val: subsampled metric loaders + pinned contiguous video loaders #642
• infra: mecka 6D / keypoint recipes, 300M models, defaults, and the rollout revert #640
• eval: layout-aware frame-invariant metrics, and a train-split viz head #659
• data: stop letting one bad episode or a rotation channel take a run down #658
• norm: guard degenerate channels, horizon pooling, and no rot6d normalization #657
• keypoints: the 144-D wrist-first hand-keypoint action becomes the human default #656
• mecka: fix the double-mirrored LEFT wrist frame, and correct existing zarrs #655 👈 (View in Graphite)
• 6d: PI goes 6D-only; Eva and Human gain the continuous-rotation modes #654
• 6d: continuous rotation representation, bimanual layouts, shared norm helpers #653
• train: give each Slurm slurm job gets its own cache dir to avoid conflict #638 : 1 other dependent PR (#639 )
• write git version into the run dir for reproducibility #637
• train: don't load PI base weights on resume and verify that checkpoint path exists #636
• configs: declare proprio and actions dims once per HPT model config #634 : 1 other dependent PR (#635 )
• cache norm stats for data configs to minimize recomputation #633
• import cloud packages lazily to support local training #631 : 1 other dependent PR (#632 )
• cache dataset resolver for faster training initialization #630
• validate episode hashes in data config before training #628
• modify pi camera map to match data #629
• add tests to ensure training works between every model and data vendor #626
• run CI tests on every PR #625
• add env variables for user-specific paths #624
• build: set pyproject and uv as the dependency managers #623
• fix(configs): fix paths for data hydra yaml configs #622
• refactor(algo): delete ACT #627 : 1 other dependent PR (#621 )
• tests: create dedicated test folder with unit and integration tests #620
• add pyright type checker #619
• main

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

[RyanPCo]

@greptileai

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Fixes hand pose convention in motion capture data processing.

The PR should not merge until mixed-convention datasets cannot silently receive the wrong left-wrist correction.

Findings

1.  Later episodes escape validation ▶
2.  Convention test shares implementation ▶

Summary

The PR corrects the left-hand pose relabel in the Mecka converter, adds a load-time rotation for existing Mecka and Scale zarrs, and adds convention checks and pipeline tests.

• Mixed old and newly converted episodes can evade the three-episode convention check.
• The new convention tests lack an independent assertion of the converter’s hand-frame orientation.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Stored episode poses] --> B{Dataset correction enabled?}
  B -->|Yes| C[Rotate left pose local axes]
  B -->|No| D[Keep stored pose]
  C --> E[Human frame transforms]
  D --> E
  E --> F[Training actions and proprioception]
  A --> G[Check first three episodes]
  G --> H[Warn or reject detected mismatch]

Loading

_{Reviews (1) · Last reviewed commit: "mecka: fix the double-mirrored LEFT wris..."}

[greptile-apps]

Later episodes escape validation If a dataset contains both older and reconverted episodes, this check examines only the first three in sorted path order. When those three are older, a later reconverted episode still receives the left-wrist rotation but is never checked. Its wrist frame is double-flipped, so training mixes incompatible frames.

[greptile-apps]

Convention test shares implementation The synthetic stored rotations come from _unrelabeled_hand_frame, which the converter and convention checker also use. A mistake in that helper could therefore leave these tests passing while both production paths produce the wrong orientation. An independent expected-frame check for both hands and the historical left-hand relabel would catch that regression.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[RyanPCo]

Merge activity

• Oct 19, 7:23 PM UTC: A user started a stack merge that includes this pull request via Graphite.
• Oct 19, 7:35 PM UTC: Graphite rebased this pull request as part of a merge.
• Oct 19, 7:35 PM UTC: @RyanPCo merged this pull request with Graphite.