ci(gpu): point LIBERO at bundled package assets in nightly test

[shuheng-liu]

What this does

Fixes #319.

The nightly GPU test test_control_freq_reaches_real_robosuite_sim (added in #312) builds a real OffScreenRenderEnv to verify the configured control_freq actually reaches robosuite — the mocked sibling tests can't catch robosuite silently dropping the kwarg. Building a real env needs LIBERO's bddl_files/init_files/assets, but gpu_test.yml set LIBERO_CONFIG_PATH to an empty /tmp tree, so the test failed with:

FileNotFoundError: .../init_files/libero_10/LIVING_ROOM_SCENE2_..._in_the_basket.pruned_init

The empty-tree config is correct for the mocked CPU tests (they patch get_libero_path), but the real-sim GPU test needs actual assets. Note that just setting init_states=False in the test is not sufficient on its own: the env build in src/opentau/envs/libero.py reads the bddl file via get_libero_path("bddl_files"), so it would then fail on the missing bddl file instead. The workflow fix is the necessary part.

Changes:

• .github/workflows/gpu_test.yml: point LIBERO_CONFIG_PATH at the assets bundled inside the installed LIBERO package (the fork's MANIFEST.in does graft libero, shipping bddl_files/init_files/assets) via LIBERO's own set_libero_default_path(), instead of the empty /tmp tree. As a bonus this also drops a pre-existing SC2155 shellcheck warning on the old export VAR="$(pwd)/..." line.
• tests/envs/test_libero_control_freq.py: set init_states=False to focus the test on its control_freq assertion (matching the two mocked sibling tests); the env build still exercises the real bddl/asset path.

How it was tested

• actionlint on the edited workflow: the Run Tests step is shellcheck-clean, and the change removes a pre-existing SC2155 warning.
• Verified the LIBERO fork bundles the exact failing init_files/libero_10/*.pruned_init (~23 KB real file, not an LFS pointer) and the bddl_files/libero_10/*.bddl for the task under test, and that MANIFEST.in grafts them into the installed package.
• Could not run the GPU test locally (no NVIDIA GPU; LIBERO not installed). Validation is the nightly job — gpu_test.yml was manually dispatched on this branch via workflow_dispatch; see the linked run in the checks.

How to checkout & try? (for the reviewer)

# Dispatch the nightly GPU workflow on this branch:
gh workflow run gpu_test.yml --ref claude/magical-banach-eb58a8

# On a CUDA machine with the libero extra installed, run just the fixed test:
pytest -sx tests/envs/test_libero_control_freq.py::test_control_freq_reaches_real_robosuite_sim

Checklist

• I have added Google-style docstrings to important functions and ensured function parameters are typed.
• My PR includes policy-related changes.
  • If the above is checked: I have run the GPU pytests (pytest -m "gpu") and regression tests.

[shuheng-liu]

Review: the LIBERO-assets CI fix is correct and empirically validated.

I traced the workflow incantation end-to-end against the pinned LIBERO fork (45c5a01) and the mechanics hold:

• touch "$LIBERO_CONFIG_PATH/config.yaml" — libero/libero/__init__.py only prompts on if not os.path.exists(config_file). Pre-creating the file skips the interactive first-run input(), which would otherwise hit EOF and crash the import under CI. The empty file is harmless because nothing reads the config at import time (get_libero_path is lazy).
• set_libero_default_path() — reads LIBERO_CONFIG_PATH from the env and writes package-relative paths (benchmark_root = the installed libero/libero/ dir) into that same config.yaml, overwriting the empty file. The subsequent pytest inherits the exported LIBERO_CONFIG_PATH, so get_libero_path("bddl_files") resolves into the installed package.
• The assets actually land in the install: MANIFEST.in does graft libero, and — the load-bearing part — setup.py sets include_package_data=True, so the grafted bddl_files/init_files/assets end up in site-packages, not just the sdist.

Blast radius: the env-var change applies to the whole pytest -m "gpu" run, but test_control_freq_reaches_real_robosuite_sim is the only GPU-marked test that touches LIBERO config paths — every other LIBERO test is CPU/mocked and patches get_libero_path. No collateral on the other GPU policy tests.

init_states=False: doesn't weaken the test — the control_freq read-back is independent of init states, it matches the two mocked siblings, and the real bddl path is still exercised by _make_envs_task. Belt-and-suspenders with the workflow fix, which is fine.

SC2155: accurate — switching export VAR="$(pwd)/..." to a literal removes the command-substitution-masks-return-value warning, and the new quoted mkdir/touch lines don't introduce any.

Validation: the manually-dispatched GPU run (26306227758) shows Run Pytest on GPU = success; CPU / pre-commit / check-checklist are all green.

Two optional, non-blocking notes:

1. The workflows now use two different LIBERO config strategies — gpu points at the installed package via set_libero_default_path(), while cpu/regression still point at the empty .github/assets/libero tree. Intentional and explained in the inline comment, but a future reader editing one might not notice the other differs; a one-line cross-pointer would help.
2. set_libero_default_path() trusts that the install contains the assets. If a future LIBERO bump dropped a file (or moved assets to LFS without materializing them), this would resurface as a FileNotFoundError at env-build time rather than anything clearer. Not worth guarding now — the nightly is the right place to catch it.

LGTM.

---

Generated by Claude Code