make config changes

[AnikethCheluva]

make config changes

branch for pi configs

[AnikethCheluva]

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: paligemma base weights for pi0.5 training #609
• make config changes #176 👈 (View in Graphite)
• refactoring aciton transforms #604 : 1 other dependent PR (#605 )
• main

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

[github-actions]

Claude Code Review

Summary

Adds two new Hydra configs: a cotraining data config for eva+aria right-arm "object in container" task, and a Pi 0.5 model config targeting mecka_bimanual. Both are additive and don't modify existing configs.

Key concerns

1. Hardcoded absolute path (pi0.5_mecka.yaml):

   pytorch_weight_path: /storage/home/hcoda1/5/acheluva3/jcedar/EgoVerse/...
   

   This is a user-specific PACE path and will break for every other collaborator and CI. Should be parameterized via an env var or a shared checkpoint dir convention (e.g. ${oc.env:EGOVERSE_CHECKPOINTS}/pi05_base_pytorch).

2. Domain/embodiment inconsistency in obj_cont_cotrain.yaml: The datasets are eva_right_arm and aria_right_arm, but there's no matching model config wired to these single-arm domains here. Just flagging — make sure the model config used with this data config declares domains: [eva_right_arm, aria_right_arm] with proper action converters, otherwise training will silently drop one domain.

3. pi0.5_mecka.yaml only declares mecka_bimanual but PR description says "branch for pi configs" / cotrain intent — if this config is meant to be used with obj_cont_cotrain.yaml, the domains don't match at all. Please clarify pairing.

4. Missing trailing newlines in both files ("No newline at end of file"). Minor, but our repo convention is to include them.

5. Action converter key casing: MECKA_BIMANUAL (uppercase) vs domains: ["mecka_bimanual"] (lowercase). Confirm action_utils rule lookup is case-insensitive — other configs in the repo I've seen use lowercase keys in rules. If it's case-sensitive, the fallback converter will silently be used, which would be a hard-to-debug training bug.

Suggestions

• Replace absolute checkpoint path with ${oc.env:...} or a config-level paths.checkpoints interpolation.
• Verify action converter rule key casing against egomimic/utils/action_utils.py and align.
• Add a brief comment at top of each yaml describing intended use (which train script, which paired model/data config).
• Add trailing newlines.
• If these configs are actually used together, add an experiment-level config under hydra_configs/experiment/ that composes them, so the pairing is explicit and testable.
• Consider a smoke test: python -m egomimic.trainHydra --cfg job with these configs to catch resolution errors before merge.

Verdict: Request Changes

Primary blocker is the hardcoded home-directory path — that will break anyone else who tries to use this config. Casing on the action converter rule is worth a 30-second check before merging.

---

Reviewed by Claude · Review workflow