cua-lite/lite.cuaworld-assets
lite.cuaworld materials Maintained environment materials for cua-lite's lite.cuaworld.* software environments — forked from cmu-l3/gym-anything (CUA-World, MIT) and curated/edited/expanded by us. Published as the Hugging Face dataset cua-lite/lite.cuaworld-assets. This repo holds content only (no engine code): per-environment env.json, install/setup scripts/, per-task assets (tasks/<task>/…), data/, config/, assets/, optional post_build.sh, and a curated registered.json. The… See the full description on the dataset page: https://huggingface.co/datasets/cua-lite/lite.cuaworld-assets.
lite.cuaworld materials
Maintained environment materials for cua-lite's lite.cuaworld.* software environments — forked from cmu-l3/gym-anything (CUA-World, MIT) and curated/edited/expanded by us. Published as the Hugging Face dataset `cua-lite/lite.cuaworld-assets`.
This repo holds content only (no engine code): per-environment env.json, install/setup scripts/, per-task assets (tasks/<task>/…), data/, config/, assets/, optional post_build.sh, and a curated registered.json. The build/run engine lives in the cua-lite repo (lite/gym/envs/lite/cuaworld/).
How it's used
- End users never clone gym-anything. cua-lite's
scripts/install.sh <sw>resolvesdata/assets.lock.yaml(HF repo + pinned revision), fetches the needed<env>/into a gitignored.cache/, and bakes it ontocua-lite/lite.cuaworld.base(session userga). - Maintainers onboard a software by importing the pristine upstream subtree as a baseline commit, then editing on top. Each env carries an
AGENTS.mdwith maintainer notes; upstream task-asset defects are recorded per-env inUPSTREAM_ISSUES.md.
Layout
<upstream_env>/ # one dir per onboarded env, e.g. pymol_env/
├── SOURCE.json # provenance: upstream repo + commit imported from
├── AGENTS.md # maintainer notes (install/setup/launch/gotchas/adaptations)
├── UPSTREAM_ISSUES.md # (if any) upstream task-asset defects — record-only
├── env.json # hooks (pre_start install / post_start setup)
├── scripts/ # install_<sw>.sh / setup_<sw>.sh / task_utils.sh (edited in place)
├── tasks/<task>/ # task.json, setup_task.sh, export_result.sh, verifier.py, …
├── data/ # seed inputs (staged to /workspace/data)
├── config/ # curated app prefs (staged to /workspace/config)
├── assets/ # task input files (staged to /workspace/assets)
├── post_build.sh # optional extra build step (e.g. boot-time app launcher)
└── registered.json # curated splits: {"eval":[...], "train":[...], "long_horizon":[...]}Cross-cutting upstream issues (e.g. the sample_trajectory_frames(n=) packaging bug) are in the repo-root UPSTREAM_ISSUES.md. We only carry environments we actually onboard — not all upstream envs.
Provenance & license
Each env's first commit is the pristine upstream import (SOURCE.json records the exact cmu-l3 commit). All our edits live in later commits, so git diff <import-sha>..HEAD -- <env>/ is exactly what we changed vs upstream. MIT-licensed; the cmu-l3 copyright notice is retained in LICENSE.
