[ZJLi2013]

Problem and outcome

scripts/README.md lists scripts/assets/import_objaverse.py and scripts/train_smolvla.py, but neither file was in the tree. The eight Objaverse assets in assets/objects/ had no reproducible source, and there was no training entry for the recorded LeRobot datasets.

Both scripts are now in the tree, unchanged from the versions that produced the assets and the post-training runs. pyproject.toml gains a lerobot extra; robotsmith/sim/recorder.py already imported lerobot lazily without declaring it.

• import_objaverse.py needs objaverse and RocRecon's mesh_to_urdf (it raises with install instructions if RocRecon is missing).
• train_smolvla.py needs torch and lerobot (pip install -e ".[sim,lerobot]").

Layer

• Scripts
• Packaging (new optional extra)

Evidence

python -m compileall -q scripts/assets/import_objaverse.py scripts/train_smolvla.py   # exit 0
python scripts/assets/import_objaverse.py --help                                   # prints usage
git diff --check                                                                   # exit 0

Neither script was run end to end for this PR: the import needs network access to Objaverse plus RocRecon, and training needs a GPU.

AI use

• Tool/model: Cursor (Claude Opus 5.5)
• Extent: located the two scripts in the maintainer's earlier history, checked them for private paths and API drift against main, added the extra and changelog lines, and wrote this description.
• Human verification: not yet; check the unpinned lerobot extra and whether either script needs a smoke run before merge.