[RyanPCo]

Stops normalizing the rot6d rotation channels, guards zero-range channels against dividing by epsilon, and adds opt-in openpi-style horizon pooling for norm stats.

Behaviour changes

• rot6d channels are no longer normalized. This changes normalization for 6D recipes, so a checkpoint and its norm stats from before this PR are not interchangeable with a run after, and runs before and after are not comparable. This covers the cartesian 18/20-D layouts and the 144-D keypoint layout, action and proprio alike. The stat files are computed and stored as before; what changes is how they are applied. The loss now weights translation (unit scale) and rotation (raw) differently.
• ypr layouts keep the old behaviour.
• Channels with a stat range below NORM_MIN_RANGE (1e-4) are treated as constant. They normalize to 0 and unnormalize to their centre instead of going through the old (q99 - q1) + 1e-6 denominator, which scaled any off-convention value by 1e6.
• norm_stats.pool_horizon is off by default, so it changes nothing unless set.

rot6d passthrough

• A 6D rotation is two columns of a rotation matrix. It is already bounded in [-1, 1], and per-channel centre-and-scale pulls the pair off the manifold without adding resolution.
• TRI's LBM and UMI, the two projects that anchor an action chunk at near-identity the way this repo does, both exclude rotation from normalization. Per-timestep and normalized was the one combination with no precedent.
• It also removes the t=0 zero-range rotation cells at the source, leaving the degenerate-channel guard for the structural kp0 / kp9 position zeros.
• rot6d_channels() in pose_utils.py reads the indices off the existing layout tables. _apply_norm_one / _apply_unnorm_one pass those channels through, so the dataset path, the evaluator revert and the rollout revert all follow from one place.

Degenerate-channel guard

• The t=0 cell of a wrist-frame chunk is exactly the identity pose, so its quantile range is 0.
• The rule matches Diffusion Policy / robomimic (ignore ranges under 1e-4); GR00T maps min == max to 0.
• It lives in the shared _apply_norm_one / _apply_unnorm_one in action_utils.py, so the dataset and the tests agree.

Horizon pooling

• norm_stats.pool_horizon: true computes one set of stats per channel over every timestep of the chunk, as openpi does, instead of one per (timestep, channel) cell.
• The pooled stats are tiled back to (T, D), so normalize, the bounds check and checkpoints see the same layout either way.
• pool_horizon is part of the norm-stat cache key, so flipping it can't silently reuse stats computed the other way.
• Which is better is unsettled. The literature is split (LBM per-timestep with a stated rationale; openpi, RDT, GR00T, Diffusion Policy, UMI, lerobot and Octo pooled), and nothing here has compared them on a real recipe, so the default stays per-timestep.

Precompute script

egomimic/scripts/precompute_norm_stats.py computes and caches a recipe's stats on a CPU node outside a training run. It writes an explicit norm_stats.json and fills the content-keyed cache under the same key trainHydra uses, so a plain training run on the same episodes and recipe hits the cache.

Tests

• New tests/unit/test_eva_wristframe_norm_pool.py: pooled stats tiling, pool_horizon in the cache key, the constant-channel guard, rot6d passthrough per key and width, and no passthrough outside the 6D layouts.
• tests/unit/test_wrist6d_roundtrip.py passes the rot6d channels through its normalize / unnormalize round trips.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HjCQBqQCr2igFEMR7yUovw

[RyanPCo]

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

• 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 👈 (View in Graphite)
• 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
• 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]

Automated adversarial review (read-only; findings were checked by a second skeptical pass before posting): 1 medium, 1 low, 3 nit.

[RyanPCo]

medium (backward-compat): This fallback only applies to checkpoints loaded through ModelWrapper.from_state. trainHydra always builds fresh norm stats (rot6d_passthrough=True) and then loads weights through trainer.fit(ckpt_path=...) or load_checkpoint_weights. Nothing restores the checkpoint's saved norm state. So resuming, requeuing or mode=eval on a pre-#657 checkpoint (e.g. Eva cartesian_6d) silently feeds raw rot6d to a model trained on normalized rot6d, with no error. When ckpt_path is set, read the checkpoint's norm_stats_state and carry its rot6d_passthrough (False when absent), or fail on a mismatch. Also copy the flag in set_norm_stats_from.

[RyanPCo]

low (duplication): This replicates trainHydra's norm loop by copying it (keymap norm_mode rewrite, episode fingerprints, cache_inputs/norm_cache_key/write_cached, private _norm_run_metadata), and it has already diverged: it uses hydra.utils.instantiate rather than _instantiate_dataset, and it skips the norm_stats.norm_stats.get(emb) guard before the cache write. Since the whole point is producing the same cache key trainHydra would, consider extracting a shared compute_or_load_norm_stats(cfg, train_datasets, *, max_samples=None, write_cache=True) used by both.

[RyanPCo]

nit (stale-content): --data mecka_all_6d --model pi0.5_bc_mecka_6d don't exist under hydra_configs at this head. Use shipped ones (e.g. --data mecka --model pi0.5_bc_mecka) and drop the 'emimic venv' reference.

[RyanPCo]

nit (comments): This comment carries measurements ('28 action cells and 10 proprio channels on a real 144-D episode') and a comparison with Diffusion Policy/robomimic/GR00T that belong in the PR description. Two lines are enough: the threshold exists for structural-zero channels (wrist-frame t=0, kp0/kp9), and such channels normalize to 0 and unnormalize to their center.

[RyanPCo]

nit (comments): Keep one line on what pool_horizon toggles (stats per channel over the whole chunk vs. per (t, channel) cell). 'The two have not been compared on a real run here' and the project list are experiment notes for the PR description.

[RyanPCo]

Review findings: status after the fixup pass. The fixes are squashed into this PR's commits, and the whole stack was force-pushed.

Fixed

• Usage example names configs that do not exist and a personal venv: Example now uses --data mecka --model pi0.5_bc_mecka; venv mention removed.
• NORM_MIN_RANGE comment carries measurements and literature survey: Cut to the threshold's purpose and the structural zeros it covers.
• Config comment records experiment status: pool_horizon comment now only describes what it toggles.
• rot6d_passthrough compat only covers the from_state deploy path: set_norm_stats_from copies rot6d_passthrough. trainHydra settles ckpt_path and reads the checkpoint's saved norm_stats_state (mmap), then carries its flag into the fresh stats before wiring. Unit tests added.

Partial

• Script re-implements trainHydra's norm loop instead of sharing it: Fixed the concrete drift: uses trainHydra._instantiate_dataset, adds the norm_stats.get(emb) guard before writing the cache, and drops the seed=42 override. Did not extract a shared function from trainHydra.train.