Collections

Overview

A collection is a named, versioned container that groups related resources — datasets, files, or events — so you can retrieve, share, and build on a curated set as a single unit. For example, a file collection might gather front-camera images from ten separate flight datasets into a training set, without duplicating any data.

Reach for a collection when the grouping itself carries meaning: a benchmark set of missions for regression testing, a labeled set of events for model training, or a hand-picked set of files to share with a teammate. Because a collection references its members rather than copying them, the same dataset, file, or event can belong to any number of collections.

Every collection gets a unique identifier beginning with the cl_ prefix and, like datasets and files, supports tags for search and organization.

Resource types

Each collection holds exactly one resource type, chosen at creation time:

• Dataset collections reference whole datasets. Use one when the unit of work is an entire session or mission — for example, every dataset captured on a particular firmware release.

• File collections reference individual files. Use one when the grouping cuts across datasets — for example, all front-camera images regardless of which dataset they live in.

• Event collections reference events. Use one to curate labeled moments across many logs, such as all “hard braking” events found in a month of autonomous vehicle data.

The resource type is permanent; to group both files and whole datasets, create two collections.

Versioning

Every change to a collection — adding or removing resources, or updating the name, description, or tags — increments its version number and is recorded in its change log.

Each version records the collection’s exact membership, so you can retrieve the collection as it existed at any past version. This makes collections a reliable handoff unit: record the version number when you hand a collection to a training pipeline or a teammate, and anyone can recover the same inputs later.

Versioning tracks membership, not the members themselves. If a referenced resource is deleted, the change log still shows that it once belonged, but its content is no longer accessible.

Sharing and deletion

You can share a collection with specific users by granting them viewer or editor access on the collection itself, without changing any permissions on the underlying resources. Sharing extends to the contents: anyone who can view a collection can also view the datasets and files it references. Because sharing a collection effectively shares its members, adding a resource to a collection requires view permission on that resource.

Deleting a collection removes only the container; the datasets, files, and events it referenced remain untouched.

Working with collections

Create and curate collections from the Collections page in the web app, or programmatically with the CLI and SDK. The Working with Collections user guide walks through the full workflow — creating a collection, adding and removing resources, pinning a version, and retrieving it from a pipeline — in the web app, CLI, and SDK. For the complete command set, see the roboto collections CLI reference.