Graviti review for workplace video and multimodal training data

Graviti advertises catalog data and custom collection for physical AI and language models. One workplace specification describes head-mounted and fixed cameras used together, with time-aligned action labels and estimated pose. The fit is strongest for a buyer testing whether synchronized views reveal an action the first-person view misses. Inspect a representative sample before relying on listed volume. Graviti catalog and specifications

Verify what is delivered together

Ask whether listed hours are unique activity hours or accumulated camera hours. Compare timestamps, calibration notes, scene coverage and missing views in a short pilot. The site labels some products early access. Confirm the availability of the exact data, annotations, and formats you need. Request separate pricing for raw and annotated outputs and any minimum order.

Clarify consent, removal and downstream use

The company says contributors sign a revocable written agreement and that commercial licenses cover model training and deployment. Ask how contributor withdrawal interacts with the license for delivered files and already-trained models. Review role, site permission, access, retention, and redistributability against the signed agreement. A public statement about rights helps identify the right questions for the specific collection. Graviti rights overview

Graviti retail back-of-house capture

Graviti’s public skilled-work specification describes retail back-of-house as a capture environment and lists head-mounted RGB, inertial data and procedural annotations. Ask for a retail-specific sample, how hand–object contact is labeled and whether any accompanying inventory evidence comes from the same take. Request usable unique hours separately from headline collection targets and confirm the rights package for each store and worker. See the retail workflow guide. Graviti skilled work dataset

An original synchronized-view review

Consider a practice task where an operator puts a carton on a shelf. The head view is blocked for two seconds, while a fixed camera still shows the placement. Ask whether the shared task identifier and timestamps connect the two streams and whether the labels cite the view that actually supports the result.

Next, offset the supplied label interval in a copy used only for your own review. An evaluator should detect the mismatch rather than accepting an action label because its text sounds plausible. Do not alter a provider’s source and present the altered copy as its delivery. This is an independent test design, with no claimed test result.

Map availability and permission to each source layer

Keep a specimen manifest for the exact product and release. It should identify original views, processed files, motion signals, estimated poses and reviewed annotations. Ask whether missing views remain explicit and how an early-access feature affects delivery support. Availability of one workplace specification does not establish every possible scenario.

If a contributor agreement is described as revocable, ask what changes for future collection, distributed source files and existing model rights. Request the applicable written terms and evidence for the actual workers and sites. A general rights description cannot settle every item in a mixed workplace archive.