Dataset
RH20T
Large multi-modal real manipulation dataset with multiple sensors and task types.
DatasetFeatured
DROID
Highly diverse real Franka manipulation data (~76K trajectories / 350h) for real-world finetune.
Why: Diverse, clean real scenes — great for finetuning
Overview
DROID is a large, highly diverse real-world Franka manipulation dataset (~76K trajectories / ~350 hours) collected across many scenes and tasks. It is one of the best open resources for real-world finetuning when your robot is Franka-like or you can retarget.
Diversity is the selling point: in-the-wild scene variation stresses policies harder than lab-only tabletop sets. Pair with language annotations and modern IL/VLA finetune recipes for strong transfer studies.
Who it is for
Real-world finetune teams; Franka labs; diversity/generalization researchers.
Key highlights
- ~350 hours of real Franka data
- High scene and task diversity
- Strong fit for finetune and evaluation under shift
- Open licensing relatively friendly (verify current terms)
When to use
- Real finetune after OXE pretrain
- Studying robustness to scene diversity
When not to use
- Your embodiment cannot map to Franka action spaces without heavy work
Getting started
- 1Download a subset first; full dump is large.
- 2Visualize random episodes for quality sanity checks.
- 3Finetune a small policy before full-scale runs.
- 4Report success under lighting and background shifts.
Papers & reading
How it compares
Caveats & pitfalls
- Storage and I/O pipelines matter as much as algorithms.
- Retargeting to other arms is non-trivial.
Content reviewed 2026-07-23