DatasetFeatured
DROID
Highly diverse real Franka manipulation data (~76K trajectories / 350h) for real-world finetune.
DatasetFeatured
Open X-Embodiment
Large real manipulation corpus across 22 embodiments (1M+ trajectories) — key VLA pretrain base.
Why: Currently the largest cross-embodiment real robot manipulation dataset
Overview
Open X-Embodiment (OXE) is the multi-lab, multi-robot dataset collection that powered the open VLA wave. It aggregates trajectories across many embodiments and institutions, enabling cross-embodiment pretraining that single-lab datasets cannot support.
OXE is not one homogeneous corpus: licenses, sensors, and action spaces differ by constituent dataset. Successful users define explicit mixtures (sometimes called “magic soup” recipes), normalize actions carefully, and track which subsets dominate training.
If you train or finetune generalist policies, understanding OXE composition is as important as understanding model architecture.
Who it is for
Anyone pretraining or studying generalist VLAs; dataset mixture researchers.
Key highlights
- Cross-embodiment scale unmatched by single labs
- Foundation for OpenVLA / Octo-style pretraining
- Public mixture recipes in follow-on work
- Forces principled action normalization thinking
When to use
- Pretraining generalist policies
- Studying transfer across robots
- Building open VLA baselines
When not to use
- You only need a single clean real-robot finetune set (prefer DROID / Bridge)
- Commercial use without checking per-subset licenses
Getting started
- 1Read the OXE paper and dataset cards for subsets you will touch.
- 2Start from a published mixture recipe rather than inventing one.
- 3Normalize actions and timestamps per robot embodiment.
- 4Track subset weights — silent dominance is common.
- 5Evaluate transfer on held-out embodiments.
Papers & reading
How it compares
Meta-dataset mixture, not a single robot corpus — check per-subset licenses.
DatasetFeatured
AgiBot World
Large-scale real bimanual dataset from AgiBot across diverse scenes and tasks.
1M+ trajectories (reported)Research / check officialBest for pretrainBest for finetuneBimanualSingle-armUpdated 2026-02-01
Caveats & pitfalls
- Per-component licenses differ — commercial use is non-trivial.
- Heterogeneous quality: more data ≠ better data.
- Download and storage costs are significant.
Content reviewed 2026-07-23