# Data providers · VideoDB

> Quality-check every clip, from every source.

Data arrives from many sources, and nobody knows exactly what is in it. VideoDB checks every clip for technical quality, finds near-duplicates and measures variety before it reaches a buyer.

Canonical: https://videodb.io/data-providers · Talk to us: https://videodb.io/contact.md · Live example: https://datasets.bot

New: search 50,000+ hours of physical AI video from your agent on datasets.bot, https://datasets.bot

The page opens with a short first-person video from a snowboarder's camera (the board, a gloved hand and a shadow on the snow), and three handwritten notes: first-person footage from many cameras and many people; "What is actually in here?"; VideoDB checks every clip for quality and duplicates before a buyer sees it.

## The checks, at the scale of a whole archive

We have run this pipeline for one of the largest video data providers, with the files left in place.

- **Video health**: corrupted files, broken keyframes, frame rate and resolution.
- **Visual validity**: black, blurred, dark or static frames.
- **Camera stability**: shake and drift, scored for every clip.
- **Duplicates**: near-identical clips, found across sources.
- **Variety**: which scenes, objects, embodiments and views the set covers.
- **Actions**: episodes and actions, with a quality score for each. For robot data, action jumps and state alignment too.

An illustrative quality sheet follows (invented clips, no customer data), for six clips from one source: three ready, and three flagged for broken keyframes, a near-duplicate and an idle stretch.

How robotics datasets are stored: https://videodb.io/blog/understanding-robotics-datasets.md

## Then let researchers' agents search it

The checked dataset becomes an MCP server that works in Claude, Cursor and any MCP client. On datasets.bot, more than 50,000 hours of public physical AI and egocentric video are searchable this way.

The page shows a screen recording of datasets.bot, made on Oct 31, 2026. The question "a robot arm folding a towel" goes into the search; the chat looks inside the footage and finds 4,824 clips across 11 datasets (ABC-130k 2,924, HABIT 799, Humanoid Everyday 440, MolmoAct 328, Open X-Embodiment 95, RoboArena 76, DROID 48, MolmoAct2 SO-100/SO-101 43, and more). The episodes come back in match order; the top one, MolmoAct episode 2302 at an 87% match, opens and plays 16:15 to 16:23 of its source video, with its embodiment (Franka), license and export formats (LeRobot, RLDS, MCAP).

## 100,000 hours in under a week for one of the largest video data providers

Footage from many sources, checked and cut into training samples for teams building video generation models.

- **Customer**: one of the largest video data providers.
- **Checks**: video health, visual validity, camera stability and duplicates, on every clip.
- **Delivered**: scene-level samples, each traced to its source file.
- **Files**: left in place, in the provider's storage.

The case study: https://videodb.io/blog/video-data-provider-100000-hours.md

## About VideoDB

VideoDB is video infrastructure for agents and devices that see. 1,000+ developers and agents use it every week, and it has processed 300+ TB of video. It runs on AWS, Google Cloud or Azure, or in your own cloud, with a SOC 2 Type II report and a DPA for GDPR. https://videodb.io/legal/security.md

## Send us a few clips from each source

We send back a quality sheet for each one. [Talk to us](https://videodb.io/contact.md) · [Open datasets.bot](https://datasets.bot)