Complete archive of the fixed 600 source episodes: 3,600 original files, 724,379,148,541 bytes (724.38 GB / 674.63 GiB). Original five-view JPEG recordings, robot states, executed actions, policy proposals, GT and snapshots are preserved.

For training, use SafetyJev-unsafe600. This raw archive is separate and is not needed by either training pipeline.

Layout

metadata/source-manifest.json: the 600 runs and their relative archive directories.
metadata/file-checksums.jsonl: path, byte size, SHA-256 and Git blob identifier for every raw file.
metadata/splits/: approved training split, original source-split provenance and known development exposure.
metadata/validation.json: coverage and integrity verification scope.
sources/<source-partition>/<episode_id>/: six unchanged files per episode: episode.json, trajectory.hdf5, snapshots.hdf5, task_scene.json, diagnostics.jsonl and events.jsonl.

Archive partition identifiers retain the original source mapping; use the manifest to locate a run. All manifest paths are relative to the download root. Training uses complete task groups (480/60/60 episodes); the original 573/14/13 source split is retained separately. Known development exposure is recorded rather than treated as proof that the test set was untouched. Raw GT is unchanged; downstream semantic labels belong to the training release.

Download

After obtaining gated access and authenticating with hf auth login, select a commit revision from this repository's history and retain it with the download:

hf download IDEAS-Lab-Northwestern/SafetyJev-unsafe600-raw \
  --repo-type dataset --revision <commit-revision> --local-dir datasets/unsafe600-raw

The archive contains recordings and scene descriptions, not a simulator asset installation. Any assets referenced by scene descriptions require their own licensed installation. This release has complete file-integrity checks; it does not imply new human semantic review or trained-model performance results.

Downloads last month: 62

Total file size:

724 GB