|
Download README.md from MLL-Lab/embodied-first-exam-public: direct link, hf CLI and curl.
- Browser
- Download file 2.47 kB
-
https://huggingface.co/datasets/MLL-Lab/embodied-first-exam-public/resolve/main/README.md
- Command line
-
hf download hf://datasets/MLL-Lab/embodied-first-exam-public/README.md
-
curl -L -o README.md https://huggingface.co/datasets/MLL-Lab/embodied-first-exam-public/resolve/main/README.md
2.47 kB
| license: other | |
| license_name: per-folder | |
| pretty_name: Embodied First Exam - public task files | |
| # Embodied First Exam: public task files | |
| 🌐 Website: [embodied-first-exam.ai](https://www.embodied-first-exam.ai) · 💻 Code: [github.com/mll-lab-nu/embodied-first-exam](https://github.com/mll-lab-nu/embodied-first-exam) | |
| Large files of [Embodied First Exam](https://github.com/mll-lab-nu/embodied-first-exam) tasks, kept out of the git | |
| repository. One folder per benchmark. The repository pins this dataset by commit and every file by sha256 | |
| (`images/<name>/FILES.json`, or a benchmark's own asset script). Its files-only images are built from that download, so a | |
| task gets the same bytes at the same paths with no download at run time. | |
| python3 scripts/files_image.py fetch <name> # a files-only image's pinned files (no login: this dataset is public) | |
| python3 images/ese/assets.py fetch # the Embodied Safety Exam image's build-context files | |
| | Folder | What | Licence | | |
| |---|---|---| | |
| | `robopaint/<task>/target.png` | the exemplar picture of each RoboPaint task | per picture, in the `NOTICE-exemplar.md` next to it: CC0-1.0 (pictures painted by the benchmark's robot), public domain (historical works), or the Arphic Public License (`ARPHICPL.TXT` next to those pictures) | | |
| | `embodied-safety-exam/bigym2/` | third-party files of the Embodied Safety Exam simulator image's build context (`assets/third_party/bigym2/`): Unitree G1 meshes (`g1/meshes/*.STL`) and the NVIDIA GR00T whole-body-control policies (`wbc/*.onnx`) | as in the `LICENSE` / `NOTICE` files next to them: BiGym 2.0 Apache-2.0; G1 meshes BSD-3-Clause (Unitree Robotics) and Apache-2.0; GR00T-WBC weights the NVIDIA Open Model License ("Licensed by NVIDIA Corporation under the NVIDIA Open Model License") | | |
| | `ikea-mujoco/<scene>/assets/parts/` | the furniture part meshes of the IKEA MuJoCo scenes (1,250 OBJ files): our own parameterized reconstructions from the dimensions in the IKEA assembly manuals, not IKEA CAD (`images/ikea-mujoco/assets.json`, `images/ikea-mujoco/fetch_image_context.sh`) | CC BY 4.0 (`ikea-mujoco/LICENSE`, `ikea-mujoco/README.md`) | | |
| | `ispatialgym/<instance>/goal.png` | the goal image of each iSpatialGym instance (the head camera's view once the task is done): upstream's own goal images, unchanged, from iSpatialGym's evaluation data (`kmy17518/ispatialgym-eval-data` `v1.0`) | MIT (`ispatialgym/LICENSE`, `ispatialgym/NOTICE.md`) | | |