DatasetFeatured
Open X-Embodiment
Large real manipulation corpus across 22 embodiments (1M+ trajectories) — key VLA pretrain base.
Dataset
OXE Magic Soup (subset recipes)
Common Open X-Embodiment mixture recipes used to reproduce VLA pretraining.
Overview
OXE Magic Soup refers to curated Open X-Embodiment mixture recipes used in open VLA pretraining papers. The mixture—not just “all of OXE”—often determines stability, transfer, and compute cost.
This entry exists so practitioners treat data recipes as first-class artifacts: document subset weights, filtering rules, and normalization, just like model hyperparameters.
Who it is for
Engineers reproducing open VLA pretraining; dataset mixture researchers.
Key highlights
- Actionable OXE subset recipes
- Critical for VLA reproducibility
- Community knowledge compressed
When to use
- Reproducing OpenVLA-style pretrain
- Designing new OXE mixtures
When not to use
- Single-dataset finetune only
Getting started
- 1Collect mixture tables from OpenVLA/Octo papers and code
- 2Implement subset filters and action transforms
- 3Log per-subset sample counts during training
- 4Ablate one subset at a time when debugging collapse
Papers & reading
Caveats & pitfalls
- Recipes drift across code versions.
- Licenses remain per-component.
Content reviewed 2026-07-23