Answers on how Curatrix datasets are captured, cleared, delivered and priced — and how tradespeople can contribute.
Each release contains egocentric video of skilled tradespeople performing a specified manipulation task, cut into episodes with a defined start and end state. Every episode carries annotations (subtasks, phases, tools, objects, outcome states), a manifest entry linking video, sensor streams and metadata, and provenance records covering contributor consent, site authorisation and licence scope.
Qualified tradespeople wear head-, helmet- or chest-mounted cameras while performing the task at approved workshops, construction sites, residential or industrial locations. Capture follows a written protocol agreed with the customer; nothing is recorded opportunistically. Each release lists its sensor rig so you can filter by viewpoint.
Real job sites contain the variation that breaks policies trained on narrow data: changing light, clutter, worn tools, partial occlusion and long task chains. Field data also captures genuine skill — how an experienced worker grips, applies force and recovers from small errors. Simulation and lab data are useful complements, but they do not replace it.
Beyond hours, each datasheet reports contributor count, number of sites and environment types, the ratio of recorded to accepted hours, annotation density per episode and rights coverage. Episodes are accepted only if they meet the written acceptance criteria for the task.
Every capture session is linked to a signed contributor release that names the purpose: training and evaluation of machine-learning models, including commercial use. The release is stored against the session, so any episode can be traced back to the agreement covering it. Contributors are paid for accepted hours.
Recordings are uploaded over an encrypted channel into access-restricted EU-region storage. Automated detection flags faces, readable text and screens; a human reviewer then checks every clip before it enters a release. Site owners sign a location release before any capture takes place.
Releases ship under evaluation, research, commercial or enterprise terms. The scope is recorded in the dataset manifest alongside provenance data. If you need broader rights later, the upgrade path is defined up front rather than renegotiated from scratch.
Video as MP4, tabular annotations and metadata as Parquet, the manifest as JSON, and synchronised multi-stream recordings as MCAP. Loaders for LeRobot and ROS are planned. The catalogue shows which formats each release includes today.
Yes. Every release is immutable and carries a version. Later releases add episodes or corrections without rewriting earlier ones, so an experiment run against a given version can be reproduced exactly.
Yes. Request access to a release from the catalogue and we will provide an evaluation sample with the full manifest and annotation schema, so you can test your loaders and evaluate the data against your own benchmark first.
Yes. You describe the manipulation task, target environment and evaluation setup; we turn that into a written data specification, recruit qualified tradespeople, record under protocol, process and annotate the footage and deliver a versioned release. Gaps found in your evaluation feed the next collection round.
Catalogue releases are priced per release and licence scope. Custom collections are quoted on the data specification: number of episodes, contributors, sites, sensor rig and annotation depth. Book a technical call and we will scope it with you.
Qualified tradespeople record agreed tasks with a camera we provide, at sites that have authorised capture. You are paid per accepted hour of footage. Details on eligibility, equipment and payment are on the tradespeople page.
The blog goes deeper on capture, rights and evaluation. For anything specific to your task, book a technical call.