| # BEHAVIOR-1K offline bundle | |
| ## Contents and provenance | |
| - `archives/behavior-source-overlay.tar` contains BEHAVIOR-1K pinned to commit `b1979916ec1549b10a4e65e630bc6504a9af1b00`, plus the validated overlay/configuration. | |
| - `archives/behavior-datasets.tar` contains the prepared BEHAVIOR-1K assets, OmniGibson robot assets and challenge task instances. The machine credential `omnigibson.key` is intentionally absent. | |
| - `common/containers/behavior-isaacsim/` is the amd64 OCI layout for `harbor.smoa.cc/vhaowenyan/behavior:3.9.2`, including the Isaac Sim runtime. | |
| Model checkpoints, run outputs, logs, caches and credentials are excluded. | |
| ## Archive layout | |
| The canonical dataset archive is `archives/behavior-datasets.tar`. Extract it with `tar -xf archives/behavior-datasets.tar -C <workdir>` and point `BEHAVIOR_ASSET_ROOT` at the extracted `datasets/` directory. The archive excludes the machine credential `omnigibson.key`. | |
| ## Offline setup | |
| 1. Download this folder and `common/`. | |
| 2. Extract both behavior archives into a work directory. | |
| 3. Import the OCI layout, for example `skopeo copy oci:/opt/starbench-assets/common/containers/behavior-isaacsim docker-daemon:behavior:3.9.2`; or use the equivalent local `ctr`/runtime import. | |
| 4. Set `BEHAVIOR_ASSET_ROOT` to the extracted `datasets/`, `BEHAVIOR_SOURCE_ROOT` to the extracted BEHAVIOR-1K source, and apply the extracted overlay/config paths from the launcher. | |
| 5. Start the adapter inside the Isaac Sim container. Accept NVIDIA's Omniverse license on the target host/container runtime as required by the installed Isaac Sim release. | |
| This bundle cannot replace the NVIDIA kernel driver/GPU runtime. Run `common/verify_bundle.py` before launching. | |