agnesluhtaru/oxe-magic-soup-plus-nl-commands

OXE Magic Soup++ Natural Language Commands

Per-step natural-language instructions extracted from the datasets in the OpenVLA oxe_magic_soup_plus mixture of Open X-Embodiment (OXE). Source datasets were downloaded and preprocessed with rlds_dataset_mod, following OpenVLA, then processed with oxe-stats.

Files

One CSV per source dataset:

• —oxe-<dataset>_nl_only.csv — train split
• —oxe-<dataset>_validation_nl_only.csv — all non-train splits, merged

Columns

• —dataset_name — source OXE dataset
• —split — train or validation
• —source_split (validation files only) — original TFDS split name before merging
• —episode_id, step_id — position within the source episode
• —language_command — the per-step natural-language instruction

Source datasets

• —austinbudsdatasetconvertedexternallytorlds
• —austinsailordatasetconvertedexternallytorlds
• —austinsiriusdatasetconvertedexternallytorlds
• —bc_z
• —berkeleyautolabur5
• —berkeleycablerouting
• —berkeleyfanucmanipulation
• —bridge_orig
• —cmu_stretch
• —dlredansharedcontrolconvertedexternallyto_rlds
• —dobbe
• —droid
• —fmb_dataset
• —fractal20220817_data
• —furniturebenchdatasetconvertedexternallytorlds
• —iamlabcmupickupinsertconvertedexternallyto_rlds
• —jaco_play
• —kuka
• —language_table
• —nyufrankaplaydatasetconvertedexternallyto_rlds
• —roboturk
• —stanfordhydradatasetconvertedexternallytorlds
• —taco_play
• —toto
• —ucsdkitchendatasetconvertedexternallytorlds
• —utaustin_mutex
• —viola

License

Each source dataset carries its own license from Open X-Embodiment; check the individual dataset before reuse. No single license applies to the merged mixture.