News3 min read
Introducing OpenBot: the data engineering layer for Physical AI
OpenBot helps robot learning teams find the right datasets and models and validate robot data before it reaches training. Here is what is live today.
By OpenBot Team
Every robot behavior starts with a data decision. Which dataset matches your embodiment? Can you use it under its license? Will your own recordings load, and are they ready for training? Today those answers are scattered across papers, READMEs, Hugging Face cards, and whatever notes someone kept last time.
OpenBot is the data engineering layer for Physical AI. It helps you find the right datasets and models and validate robot data before it reaches training. It also says plainly what its evidence does not show.
What OpenBot is
OpenBot has three parts. Each one is labelled with its current status.
| Part | Status | What it does |
|---|---|---|
| Catalog | Live | Records for robot datasets and models, each linked to its upstream source |
| openbot-data | Open source · Local only | Python toolkit that inspects your own robot data on your own machine |
| Platform API | Candidate | API keys, a public API reference, and a Playground for developers |
The Catalog: find and compare
The dataset and model catalogs list embodied AI records with access, license, modality, embodiment, formats, and signals. Every record links back to its dataset page, paper, or repository, so you can check the original before relying on it.
Readiness scores in the Catalog help you narrow the list. They are not benchmark results or legal conclusions. When evidence is missing, the record says so instead of treating it as a pass. The Ego collection is a human-reviewed set of first-person and ego-to-robot datasets, compared task by task.
openbot-data: inspect your own data locally
openbot-data is an open-source Python toolkit for robot video and LeRobot datasets. It runs entirely on your machine. You don't need an account or an upload.
pip install openbot-data
openbot-data audit ./robot_videos --out ./audit.json --fail-on error
Use it to validate data before a training run. It can scan video directories and read their metadata. It can audit a dataset
with stable error and warning codes, snapshot and diff dataset versions, and
check whether a dataset is ready for ACT or SmolVLA training. It can also plan
conservative metadata repairs, which it writes to a new copy. It reads both
LeRobot v2.1 and v3 layouts. The current release on PyPI is 0.0.3.
Platform API: build on OpenBot
The Platform is open for registration. Developers can create API keys, read the API reference, and try requests in the Playground. The Platform is a candidate release: paid usage is not enabled yet, and we will announce each new API here once it ships.
How we work
- Sources first. A record is only as good as the source it points to, so we always link to the source.
- Honest status labels. Live, Local, Mock and Planned mean exactly what they say. Planned is never presented as available.
- Keep full datasets local. openbot-data runs without an account or a dataset upload.
Where to start
- Explore the Catalog to shortlist datasets and models for your task.
- Run openbot-data on your own recordings before the next training run.
- Follow the Guides for step-by-step procedures such as LeRobot readiness checks and license review.
This blog is where we will post product updates and company news. Subscribe to the RSS feed, or follow us on X and GitHub.