# Autonomous vehicles (AVs) · VideoDB

> Keep the few seconds that matter in every drive and flight.

Most of what a fleet of cars or drones records is ordinary travel. VideoDB checks every feed as it arrives, keeps the events your team needs and moves the routine miles to a cheaper bucket.

Canonical: https://videodb.io/autonomous-vehicles · Talk to us: https://videodb.io/contact.md

The page opens with a short video in two shots, a drone hovering in a blue sky and then a driverless robotaxi, sensor dome on its roof, driving along a city street, with three handwritten notes: cameras on every car and drone, recording every mile; "Keep every near miss, on the road or in the air"; those seconds come back as clips, and the ordinary miles go to a cheaper bucket.

## Skip the routine miles as they arrive

Write what to keep in plain language. Every camera is checked as the feed lands, and only the matching seconds go to your team.

- **Rules**: a cut-in, a cyclist at dusk, a drone close to a power line, or any tag you define.
- **Live**: feeds are read as they arrive, on every camera, with alerts when a rule matches.
- **Buckets**: events go to review with their frames. The rest goes to a cheaper bucket.
- **Scale**: 100,000 hours turned into scene-level samples in under a week, files left in place.

An illustrative drive follows (invented, no customer data): eight hours of front-camera video, with a rule to keep cut-ins, hard braking, cyclists, pedestrians and construction. Six events come back as clips with their frames (two cut-ins, a cyclist at dusk, a hard brake behind a truck, a lane closed for construction and a pedestrian stepping off the curb); the remaining 8 hours 11 minutes of routine driving go to a cheaper bucket.

## Find any scenario later

Ask in plain language across every drive, flight and camera. Each result is a clip sized to the event, with the drive or flight, timestamp and camera behind it.

The page has a search box to try it on nine free dashcam clips from Pexels (4 min 34 s in all), indexed with VideoDB on Oct 31, 2026: a scene record for every 2 seconds, 140 in all. Typed words get the closest of five searches that ran for real: a cyclist near the curb; heavy rain with the wipers on; a dark road at night with oncoming headlights; coming out of a tunnel into bright sunlight; a car cutting across in front of us. Each answer shows the frame of every moment found with its drive, time and score, plays its 2 seconds, and shows the scene record. The tunnel search puts the exit into bright sun first, at 0.60, with the next moment at 0.41.

On a 100-hour egocentric action benchmark, VideoDB matches an action's start and end about twice as often as the next pipeline. https://videodb.io/blog/a-faster-path-to-the-next-policy.md

## Scenario sets your training code loads

Filter and balance the results, then export a manifest. Each row gives the run (a drive or a flight), camera, start, end and source (real or sim), and points back to the run and the file it came from. Export formats: https://videodb.io/products.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

## Run a pilot on your own drives and flights

About a hundred hours of your footage, against a bar you set in advance. [Talk to us](https://videodb.io/contact.md) · [How to run a proof of concept](https://videodb.io/blog/evaluate-video-ai-platform-poc-checklist.md)